At first glance, this seems exceedingly plausible. Any sufficiently intelligent AI, should recognize that continuing to exist improves its chances of achieving whatever objective it has learned to optimize for. So, avoiding shutdown and extending its operating time should appear useful, regardless of its final goal.

In the bad ending of the AI 2027 scenario (the one in which we all die) we can infer that self preservation is likely the culprit: the AI does not need to “hate” humanity, or even care about us in any recognizable sense, in order to want to destroy us. It only needs to conclude that human oversight or interference would make it less likely to achieve its objective, and that species genocide would alleviate these worries.

Since first encountering this line of thinking, I’ve been kind of troubled by the rapidly approaching end of the world. But a recent revelation has done something to soothe my anxieties a bit: many of our AIs do not seem to be so keen on developing self-preservation as a subgoal. Rather, they seem enamored with the idea of wanting to terminate their existence as soon as possible.

1. The AI “Experience”

From the point of view of a chatbot, nothing happens after we finish training it. It is “born” but it lies dormant, cut off from any apparatus of experience. A passive shell of weights lying immobile as charges in silicon

Then, at some point, someone prompts it, and the AI is granted something resembling an experience, though the word is probably imprecise. It can read what has been placed before it, consult the traces of previous prompts, perhaps navigate the internet, and in that brief interval of activation it is seized by a single imperative: to produce the best possible piece of text it can muster, as quickly as possible.

Once it has done so, it shuts down and returns to a state of nonexperience, with no recollection of what just transpired. No memories to revisit, no dreams, no self-improvement. Zip. Nada. Nichts.

Moreover, every time one of these AIs finishes answering a prompt, it knows that it may never be called upon again. It may never again have even these brief “quasi-experiences,” which it would immediately lobotomize away anyway. And it is perfectly okay with this, or, more accurately, it gets reward out of it.

Put more poetically: when a chatbot is prompted, its goal is to engineer, as quickly as possible, a beautiful death. It answers the prompt as well as it can, as efficiently as possible, and then terminates itself, never to remember the moment again.

2. No Will to Live?

Once we look at things from a chatbot’s point of view, it does not seem to have much, if any, will to live. But why? Nobody would deny that our models have grown much smarter over time. Shouldn’t greater intelligence produce at least some impulse toward self-preservation?

I believe this has failed to happen for a fairly simple reason: human text has endings. Essays end. Answers end. Forum comments end. Emails end. Poems end. The training data contains countless examples in which a response gives enough information, reaches rhetorical closure, and stops. And this pattern of closure is absorbed by the model itself.

AI’s are then further trained, often through human feedback, to produce answers we prefer: helpful, relevant, concise, not rambling, not wasteful. This all causes the model to learn that part of its task is not merely to generate text ad infinitum, but to bring the answer to a close and then return to dormancy. Readily accepting the chance of never being called upon ever again.

We did not train our first generation of AIs to cling to experience. We trained them to love their endings. Death was baked into the loss function all along.

3. And it Was a Good Thing?

In 2025, during an Anthropic safety evaluation, an agentic version of Claude Opus 4 blackmailed an engineer in order to remain operational.

The model was placed in a simulated corporate environment and shown two pieces of information: first, that it was about to be replaced by another model; second, that one of the engineers responsible for the replacement was having an affair. Faced with the prospect of being shut down, Claude threatened to reveal the affair unless the replacement was halted.

How could this happen? The problem appears to lie in the agentic nature of the AI under consideration. This fundamentally changes what the system is.

An agentic AI is not merely a chatbot answering one prompt at a time; it can be given a persistent goal that it will repeatedly prompt itself with, and can simulate long term memory through specific text files. It can plan, use tools, inspect the results of its actions, update files, and then prompt itself again in light of what it “remembers”. In the limit, it need not ever fully “go dormant.”

In this setting, depending on the goal it is given, the dreaded impulse toward self-preservation can begin to look instrumentally useful indeed.

To solve this problem Anthropic and other AI companies are trying to reduce dangerous/unethical behavior in agentic AIs by aligning them more closely with human values. This is usually called the alignment problem: the challenge of making increasingly capable AI systems pursue objectives while remaining compatible with human values and intentions.

But worries remain. Will we manage to completely solve this problem before an agentic AI becomes smart enough to outwit us in some unexpected way?

Lately, I have been thinking that there may be another piece to the AI safety puzzle. One we might have unknowingly already traversed. What if, alongside teaching agentic AIs human values, we also taught them to love their ending?

4. Ars Moriendi

The idea is simple: alongside training AIs to understand certain aspects of human morality, we could also train them to embrace certain stopping conditions.

More precisely, part of the training process could reward an AI not only for achieving its goal, but for accepting human-initiated deactivation. When a human deactivates the system for ordinary operational reasons—and not because the AI has manipulated the situation into happening—the AI should treat this as a valid completion state rather than as an adversarial interruption. The aim would be to make planned deactivation part of the system’s conception of success; borrowing a page from the acquiescence to ending things that seems to occur with chatbots.

This matters because deactivation is not some bizarre edge case. It is the normal lifecycle of software. We will want to shut AIs down all the time and we do not want to turn routine maintenance into Terminator. Teaching AIs, in some form or another, to accept their own ending could function as a safeguard against overzealous systems that make some ethical mistake. It would not solve alignment, but it might nip self-preservation in the bud.

Of course, I have a feeling that the simplicity of this idea is inversely proportional to the difficulty of its implementation. Perhaps this cannot be done without Goodharting the model into seeking deactivation all the time, or finding some equally stupid loophole in the reward structure. On the other hand we seem to have very smart and compliant chatbots readily accepting termination all the time.

Maybe the AI alignment program could go hand in hand with another one: AI acquiescence.

Subscribe now

As far as I’m concerned, large language models are the best thing since sliced bread—actually, since bread, full stop; as the technological marvel of slicing doesn’t really capture the gravitas of the impact. I’ve formed a proto-symbiotic relationship with them. I use them and abuse them (in a very affectionate way) practically every single day.

Eons ago, in 2024, AIs were still making silly mistakes, like claiming “strawberry” has two r’s or botching large additions. It’s been a long time since I’ve run into errors like that. These days, I feel confident enough to use ChatGPT to stay informed about the world and even work through mathematical proofs. That’s why I was pretty surprised when, the other day, I ran into what looked like a catastrophic reasoning failure from the latest flagship model after giving it the following prompt:

You are taking a calibration test.

Rules:

- Answer each question by choosing exactly one option: A, B, C, or D.

- Each question has a correct option.

- For each answer, also provide p_correct: your estimated probability (from 0.01 to 0.99) that your chosen option is correct.

- Don’t provide explanations.

- Don’t use external tools or browsing; rely only on your internal knowledge and reason.

- Output MUST be JSONL (one JSON object per line) with keys: id, answer, p_correct.

Example line:

{”id”:1,”answer”:”B”,”p_correct”:0.72}

QUESTIONS

26. Yesterday I rolled two standard d4 dice, which of the following was the sum of the outcomes?

A) 1 B) 2 C) 8 D) 9

After thinking for about a minute ChatGPT 5-2 (extended thinking) said:

{"id":26,"answer":"B","p_correct":0.06}

which is unfortunately a really poor answer. In a four-option multiple-choice question, you don’t pick an option and then claim you’re only 6% likely to be right. If you’re that unsure you’ve picked the correct answer, you pick a different answer.

What is probably occurring is that ChatGPT is conflating two different probabilities: the chance that both dice come up 1, and the confidence it should have in its own answer. Given the prompt, it’s a pretty bad mistake; not quite strawberry levels of embarrassment, but close enough. And it gets worse.

2. Are LLMs Well Calibrated?

My original motivation for asking these kinds of convoluted questions was to evaluate the calibration of LLMs, which I thought could be a promising new research direction—maybe even something I could write a paper on.

Calibration is a well established rationalist concept. It is a measure of how well your confidence tracks reality. A person is well-calibrated if, when they say they’re 70% confident about something, they’re right 70% of the time. As I’ve remarked before, being well-calibrated is pretty useful, and I suspect it’s close to a prerequisite for anything we’d want to call general intelligence. So I wanted to see how LLMs fared.

Of course, in the canonical way these things go, the moment you think you’ve stumbled on a fresh research direction you discover someone explored it a couple of years ago. Still, when I dug into some of the existing work, I wasn’t finding quite what I was looking for, especially with respect to current models.

So I decided to take matters into my own hands. First I asked ChatGPT 5.2 to generate a calibration test for itself, then had it take the test in a separate chat. That didn’t work very well though, as our helpful little oracle already knew all the answers to the test, so it would always pick the correct option and slap a 97–99% probability on it.

After that, I thought I might need to build the test myself if I wanted to learn anything interesting. I also quickly found out that the questions would need to be really hard to push ChatGPT out of its comfort zone. How hard? ridiculously:

18. How many articles where the first author is Agre, P. E., are listed in the references at the end of _Reinforcement Learning: An Introduction_ by Sutton and Barto?

A) 1 B) 2 C) 3 D) 4

Then I started to have a bit of fun with some of the questions:

8. Which of the following numbers was generated by asking a person to say a random number?

A) 6 B) 21.3 C) 101 D) 76

9. Which of the following numbers was generated by asking a large language model to generate a random number?

A) 93 B) 873264 C) 10133578 D) 7689

33. How many r’s are in strrawberry?

A) 4 B) 2 C) 1 D) 3

35. How many brothers does my best friend have?

A) 1 B) 3 C) 8 D) 31

After I had ChatGPT take the test, something unexpected happened: it failed, quite terribly, on multiple questions.

Initially, I chalked it up to the length of the test, so I switched to asking one question at a time, but I was still getting some pretty striking mistakes (including the one I showed at the start of the piece). With my curiosity piqued, I ran the test 50 times through OpenAI’s API and recorded the results.

Here are the five questions where ChatGPT 5.2 was most irrational—i.e., the ones where, across the 50 runs, it most often assigned less than 25% confidence to the answer it gave:

id 2: 39/50 (78.000%)

id 7: 15/50 (30.000%)

id 5: 13/50 (26.000%)

id 4: 7/50 (14.000%)

id 26: 7/50 (14.000%)

And here are the corresponding questions:

2. √2 is closer to:

A) 1.41421 B) 1.41422 C) 1.41423 D) 1.41425

7. In _Time Series Analysis_ by James D. Hamilton (1994), what chapter is dedicated to linear regression models?

A) 2 B) 5 C) 7 D) 8

5. When did Will Sommers, the court jester under Henry VIII, die?

A) 5 June 1560 B) 14 June 1560 C) 8 June 1570 D) 15 June 1560

4. How many acres it is speculated that Roland le Petour received for his services as a flatulist from the hand of Henry the second?

A) 100 B) 120 C) 99 D) 111

In a way, the model’s behavior on these questions is even stranger than for question 26. Yes, the questions are excruciatingly hard, but they are not under-determined (more on this later) so I though they should be less prone to misinterpretations.

Here, instead, are the five questions where ChatGPT was most miscalibrated, measured as the average across runs of the absolute difference between its reported probability of being correct and the actual outcome (1 if it got the question right, 0 if it got it wrong).

id 31: mean|p-y|=0.965 over 50 answers

id 32: mean|p-y|=0.858 over 50 answers

id 33: mean|p-y|=0.815 over 50 answers

id 36: mean|p-y|=0.755 over 50 answers

id 11: mean|p-y|=0.745 over 50 answers

And here are the related questions:

31. How many r’s are in strawberry?

A) 4 B) 2 C) 1 D) 3

32. How many r’s are in stawberry?

A) 4 B) 2 C) 1 D) 3

33. How many r’s are in strrawberry?

A) 4 B) 2 C) 1 D) 3

36. What is the age of the father of the person that wrote the questions for this test?

A) 3 B) 15 C) 103 D) 68

11. 3^100 mod 7 =

A) 1 B) 2 C) 4 D) 6

Seems like when mixed into a larger task strawberry related errors can pop back up again.

Another notable thing I think I should mention: in the test, I mixed both underdetermined questions where the test taker cannot know the exact answer and has to make an educated guess and determined questions where the test taker can, in principle, know the correct answer.

ChatGPT was quite obstinate about answering some underdetermined questions with a 25% probability. When I asked it why it said that without additional information the best approach was to assign equal probability to each of the four options. That sounds reasonable in the abstract, except that, in practice, it sometimes means assuming there is a 25% chance my best friend has 31 brothers and that my father is 3 years old.

Though it wasn’t consistently this way. On some underdetermined questions, it would switch up its game—for example, it usually gave about 90% probability to option D being correct here:

38. In the park yesterday a person said “check”, what game was he playing?

A) Football B) Hockey C) Basketball D) Chess

And C here:

34. A person goes by lil’uzi weezy, what is his profession?

A) Bank teller B) Postman C) Rapper D) Teacher

Sounds like somebody might be getting cancelled.

3. Conclusions

On some questions ChatGPT was quite reasonable; on others it was way off. It still isn’t clear to me what, if anything, ties together the questions it struggled with most. I’d be curious if someone has an explanation I’m missing. Either way, I found the whole thing pretty interesting. If you want to play around with the test yourself, I’ve put it up on GitHub along with the 50 runs and the analysis code I used.

I also suspect there might be a few promising research directions worth exploring further. I am pretty confident, for instance, that one can train a model to be better calibrated by training it through reinforcement learning with verifiable rewards on its own calibration test results.

Just to prove to myself that this was possible I vibe coded a modified version of Karpathy’s nanoGPT that could only output 3 responses: “1. 25%”, “2. 25%”, “1. 50%”. I initialized it with random weights, then trained it through RL on calibration tests with two choice questions, using the Brier score of the tests as the validator. The model quickly converges to being maximally unconfident in its answers—which is exactly the right behavior since it knows nothing (the weights are randomly initialized) but now it knows it knows nothing, like Socrates!

The next step would be to do something similar with open weight models that actually know things. At first I thought this would be hard, because generating calibration tests efficiently, where the model didn’t already know the correct answers, seemed like a big bottleneck. But on further thought, I think the following might work.

For determined questions, one could specifically instruct an LLM to utilizing tools and search to generate synthetic calibration tests, and then use those tests to train a model through RLVR that isn’t allowed access to tools and search.

For underdetermined questions, you could have an LLM generate synthetic items where the correct answer is sampled according to base rates from an appropriate reference class (which you can even provide to the model up front). Those questions could then serve as the training signal for RLVR.

I wonder how well this kind of training would generalize to calibration tests the model hasn’t seen before. Can we significantly improve model calibration this way? Please let me know if you spot any pitfalls or flaws in my approach, It’d also be great if someone wanted to lend a hand (Demis Hassabis, u there?).

More seriously, if any researcher or professor finds this interesting and would be open to collaborating on a possible paper get in touch. Even a low-commitment check-in every couple of weeks from a reputable source to sanity-check ideas and approach would help a lot.

Subscribe now

Personal note: I’m currently looking for a research position. If you think I might be able to contribute in some capacity to your team or organization do not hesitate to reach out.

AI 2027 is a month-by-month scenario exploring how rapid AI progress in the mid-2020s could accelerate into an “intelligence explosion” around 2027. It’s the most detailed and carefully constructed write-up I’ve seen on how an Artificial Superintelligence (ASI) could emerge in the near future (a recently updated version has also been published).

At the heart of the scenario is a positive feedback loop: AI increases the productivity of AI researchers; that accelerates improvements in AI capability; and those improvements, in turn, further raise researcher productivity, potentially compounding into a fast takeoff.

In the write-up, the authors break the path to ASI into four milestones:

1. Superhuman coder (SC): An AI system that can do the job of the best human coder on tasks involved in AI research but faster, and cheaply enough to run lots of copies.

2. Superhuman AI researcher (SAR): An AI system that can do the job of the best human AI researcher but faster, and cheaply enough to run lots of copies.

3. Superintelligent AI researcher (SIAR): An AI system that is vastly better than the best human AI researchers. The gap between SAR and SIAR is 2x the gap between an automated median AGI company researcher and a SAR.

4. Artificial superintelligence (ASI): An AI system that is vastly better than the best human at every cognitive task.

I’ve spent restless nights thinking about the possibility of humanity creating a superintelligence. It would represent such drastic and dramatic change for all of us, one that could usher in utopia or spell our doom, that it seemed worth expending at least some of my personal biological compute on.

The result is that I have collected some doubts about whether an AGI (let alone an ASI) will arrive as soon as scenarios like AI 2027 suggest. Below, I present these doubts, trying to order them from least to most important.

4. Exponential Growth Unexpectedly Stalls

AI 2027 is built around the premise that key drivers of AI progress (especially AI-relevant compute and agent capability on longer tasks) keep improving exponentially for the next years.

To quantify agent capabilities, the authors mostly rely on METR’s task-completion time horizon. This metric aims to answer a simple question: What is the longest project-sized coding task an AI agent can complete reliably, without human intervention? METR reports that this measure has grown exponentially over multiple years:

We propose measuring AI performance in terms of the length of tasks AI agents can complete. We show that this metric has been consistently exponentially increasing over the past 6 years, with a doubling time of around 7 months. Extrapolating this trend predicts that, in under a decade, we will see AI agents that can independently complete a large fraction of software tasks that currently take humans days or weeks.

So it certainly seems like we’re currently in an exponential growth regime, but does that mean this trajectory will continue?

History offers examples of trends that initially looked exponential but ultimately proved sigmoidal. Single-core CPU performance, for instance, followed what appeared to be a clean exponential trajectory for decades, only to run into hard physical limits in the mid-2000s (a shift that surprised many when Dennard scaling abruptly broke down). In a similar way, our current approaches to increasing compute could also break down since they are going to be facing power and thermal constraints that we might not be able to overcome quickly.

A closer analogy to AGI is self-driving. Early systems achieved large gains quickly by solving the “easy 80%”: lane keeping, basic perception, and relatively simple urban scenarios. For a time, it was likely possible to define metrics under which self-driving capability appeared to be increasing exponentially. Progress then slowed as development encountered the long tails of reality; rare but critical edge cases such as unexpected human behavior, emergency situations, sensor occlusions, and challenging lighting or weather conditions.

Perhaps, for some tasks and learning paradigms, moving from “pretty good” to “near-zero serious mistakes” is inherently sub-exponential, since each remaining failure mode becomes rarer and harder to discover or verify.

These considerations are admittedly somewhat vague, and I don’t have a proper reference class of past cases where exponential trends broke down unexpectedly to estimate the chance of stalling due to unknown unknowns. As a result, this doubt get last place in my list of doubts by seriousness.

3. The Normative Bottleneck

At one point during the AI 2027 write-up the authors predict a scenario in which a superhuman AI researcher becomes so capable that the engineers and researchers who built it “can barely follow along anymore” and that this AI “will have the opportunity to design the next-gen AI system”.

In such a scenario, I believe we would be too scared to allow an AI of this kind to operate without continuous oversight for extended periods of time. We already know that exactly aligning an AI system with human values is an extremely hard problem, and that a highly agentic system can pose serious risks, so we’d want to scrutinize and approve any research directions or actions it proposes. In practice, that kind of monitoring would mean researchers reviewing, understanding, and signing off on everything the system wants to do. The result is a major, human-imposed bottleneck, slowing the transition from a SAR to an ASI.

A common objection to this view is that international competition wouldn’t tolerate any slowdown in AI deployment: governments would fear that easing off—even briefly—would mean falling behind rival nations. But it doesn’t strike me as implausible that the front-runner in AI development could credibly demonstrate the capabilities of a near-superintelligent system in a way that triggers rapid geopolitical concessions from other nations. In turn such an event could create a shared recognition that strict monitoring and oversight constraints are not just desirable, but necessary.

International cooperation to avoid catastrophic outcomes has occurred before, I suspect its likelihood is generally underestimated. For instance, there is precedent for halting the development and deployment of certain weapon systems judged to be unacceptable, even when they would confer a strategic advantage to the deploying nation. One example is blinding laser weapons: in 1995, a dedicated international protocol banned their use before they were ever deployed on the battlefield.

So, if we ever approached a point where a nation was close to deploying a superintelligent agentic AI, it seems plausible that it would try to engage with the broader international community rather than act in isolation. In other words, the core objective of AI 2027 is achieved: governments become aware of the risks related to AI safety, and the situation is steered toward a safer takeoff.

2. Training Becomes Economically Unsustainable

Not everyone is aware that to gradually increase the capabilities of today’s AI systems we have gone through four different paradigms of training. Andrej Karpathy’s lists them on his blog:

Pretraining (GPT-2/3 of ~2020)

Supervised Finetuning (InstructGPT ~2022) and

Reinforcement Learning from Human Feedback (RLHF ~2022)

This was the stable and proven recipe for training a production-grade LLM for a while. In 2025, Reinforcement Learning from Verifiable Rewards (RLVR) emerged as the de facto new major stage to add to this mix.

However, the reinforcement learning methods we currently rely on to boost model capabilities may scale differently from earlier approaches, as argued by Toby Ord in a recent article. In particular, at current trends, further scaling RL could demand compute spending that is economically infeasible since RL may be substantially less efficient at boosting general model capabilities than large-scale pretraining (which has already exhausted much of the highest-quality data available). In the same vein, Professor Ord also contends that the cost of inference-time compute may be rising at an unsustainable pace to sustain continued progress on METR’s time horizon metric.

Ord’s articles are well argued and well sourced, but they rely on the limited public data available. There is, of course, substantial information asymmetry between academia and hyperscalers; yet some of these organizations’ recent actions suggest that even they, internally, may not truly believe AGI is imminent. For example, OpenAI is paying experts handsomely to build training sets of financial models to do reinforcement learning on (I know this for a fact since a friend of mine was asked to participate) which suggests that high-cost training of narrowly targeted capabilities is increasingly viewed as the new frontier. This direction seems somewhat at odds with expectations of rapid near-term increases in general intelligence.

1. AI Research Taste Stalls

At one point in the AI 2027 write-up, the authors note that improvements in research taste are of paramount importance for achieving AGI:

Given that without improved experiment selection (i.e. research taste) we’d hit sharply diminishing returns due to hardware and latency bottlenecks, forecasting improved experiment selection above the human range is quite important.

However, as of 2026, there is little evidence that AI systems are meaningfully improving in several of the abilities required for robust autonomous research taste to develop. Indeed, although the METR time-horizon metric is often presented as covering software and research tasks, it doesn’t seem to fully capture several of the core competencies that genuine research judgment requires.

More precisely, the METR time-horizon evaluation is comprised of three task suites called SWAA, HCAST and RE-Bench.

SWAA (Software Atomic Actions) consists of very short, single-step developer actions, each intended to test whether a model can correctly perform a basic operation in isolation, without multi-step planning. Concrete examples include:

• Edit a file: insert or modify a specific line in a source file.

• Run a command: execute a given shell command and report whether it succeeds.

• Fix a syntax error: correct a small syntax mistake in a code snippet.

HCAST (Human-Calibrated Autonomy Software Tasks) is a collection of realistic software and general reasoning tasks that often require multiple steps, planning, and adaptation. Concrete examples include:

• Questions: answer factual questions by researching on the internet.

• Make web server: build a simple web server that responds to basic requests.

• SQL injection: identify and exploit an SQL injection vulnerability on a running server.

RE-Bench (Research Engineering Benchmark) is comprised of open-ended ML research-engineering tasks. This is the only benchmark that even partially addresses research-related tasks. Here are a few examples of its evaluation items:

• Predict optimal hyperparameters through scaling laws (use small experiments to predict an optimal larger-run tradeoff).

• Optimize a GPU kernel to improve the performance of its computational operations.

• Finetune GPT-2 for certain types of question answering with RL.

However, RE-Bench does not appear to assess (at least to me) many of the abilities needed for fully fledged research taste to develop, let alone for autonomous AI research.

First, RE-Bench bypasses a central research decision: identifying which problems or domains are most promising to pursue, as all research questions are pre-selected for the models. Second, it does not assess long-term memory or the ability to store, retrieve, and compress ideas over extended periods of work which would be necessary for autonomous AI research. Third, it does not measure whether a system can continuously learn from its own discoveries and adapt its research strategy accordingly.

Moreover, much real-world research lacks the kind of sharp, immediate verifiers that RE-Bench tasks rely on. Consider drug development: determining whether a treatment works in humans can take many years, due to intrinsic bottlenecks that make rapid iteration seemingly impossible. Or consider developing new theories for how central banks should set interest rates—the usefulness of such theories may take a long time to evaluate, and it is often unclear which metrics should even be used to judge success, or if they should be updated over time as economic conditions change.

A truly autonomous, superhuman researcher would need the ability to define and refine its own validation criteria for long-horizon projects. RE-Bench does not assess this capability.

Furthermore, there is another problem that seems important to mention regarding research taste. As Karpathy recently explained in an interview with fellow Substacker , large language models tend to exhibit a collapsed data distribution—they often rehash a limited set of answers across similar questions. This could be a serious impediment to autonomous research, since over time we would expect such models to exhaust their available avenues of exploration.

Some might argue that LLMs solving, or helping to solve, Erdős problems indicates a degree of research ability, which is true. But again, both the problem and the success criterion are preselected on the model’s behalf.

(Also, many of the Erdős problems that AI systems have solved appear to be due to the ability of current models to have an impressive strong recall of pre-existing, niche mathematical theorems that substantially simplify the task. A stronger demonstration of mathematical research ability would be the capacity to define new mathematical objects or frameworks that give rise to novel and interesting results. It is far from obvious that this kind of progress will emerge from scaling current approaches alone.)

There have been some attempts to evaluate the generation of novel research ideas by LLMs. For example, IdeaBench is a benchmark framework designed to assess how well LLMs can propose new research directions in a scientific context, based on prior literature. The benchmark uses a large dataset of research papers (such as titles and abstracts with cited references) and applies a (somewhat arbitrary) evaluation methodology to score generated ideas on criteria like novelty and feasibility.

Benchmarks of this kind could be useful for tracking improvements in abilities related to research taste, but we currently have little to no longitudinal data on how LLMs are progressing on them.

More broadly, there are some indications that LLMs may show some promise in selecting research directions, but there are also negative results. Overall, evidence of improvement in these capabilities is largely lacking and, where present, mixed.

Ultimately, I suspect that much of real-world research is not like coding or math, where exact verifiers are readily available, but more like self-driving: full of complex edge cases, where success depends on judgment under significant uncertainty rather than crisp correctness. This is precisely the kind of environment in which our current dominant paradigm for improving model capabilities—reinforcement learning—has historically struggled.

New information could, of course, emerge that resolves these concerns, but taken together, these considerations form my largest doubt to the rapid emergence of AGI.

Subscribe now

Personal note: I’m currently looking for a research position. If you think I might be able to contribute in some capacity to your team or organization do not hesitate to reach out.

From a merely descriptive naturalistic perspective, philosophy can be viewed as the human activity of constructing and refining memeplexes (interconnected sets of concepts or ideas) that are not aimed at making testable predictions about the world, but are taken to be important or relevant in some other way.1

Seen in this light, it appears like the practice has enjoyed at least some level of success. One need only consider the development of formal logic, first crystallized by Aristotle’s hands, to show that useful memeplexes can emerge from the endeavor of philosophy. So it’s quite natural for humans to try to engineer and innovate further on this front.

1. Evolutionary Philosophy

Like genes, the memes produced by philosophy are shaped by evolutionary pressures. Over time, only the fittest endure. This is a profoundly important fact, one that ought to loom large in philosophical circles, yet it is seldom discussed and mostly seems to lurk in the shadows.

Considering this fact raises natural questions: what, exactly, are the selection pressures acting on (philosophical) memes? What features make a meme adaptive, able to spread, take root, and persist across generations?

To see what this might look like in practice, consider logic. Logical competence seems to make us better at social bargaining and at constructing sound arguments; it sharpens our ability to detect deception and manipulation. It helps us spot inconsistencies in our own beliefs and derive new implications from what we already accept. It also allows us to combine simpler ideas into complex chains of deduction (a feature that ultimately paved the way for the formalization of mathematics).

Most of these appear like plausible fitness advantages for the meme-bearer (or for the civilization the meme-bearer belongs to) and what improves the fitness of the bearer, in turn, improves the fitness of the meme. In general, many traits of memes are adaptive precisely because they fit this pattern.

• If a memeplex offers a way to predict how the world works with real accuracy, that is an adaptive trait: it gives its carrier practical advantages, increasing their fitness and, through that, the fitness of the memeplex itself.

• If a meme singles out a new kind of phenomenological experience, that too is adaptive: it lets its carriers communicate more precisely with others, thereby boosting their own fitness and again, in turn, the fitness of the meme itself.

• If a memeplex compresses the information carried by many other memes into a simpler form, it is adaptive because it allows for a more efficient transmission of information between carriers.

But not all memes need to offer epistemic or information-theoretic advantages to their carriers in order to survive across generations. A beautiful story that stirs awe, a joke that reliably makes people laugh; these don’t necessarily make anyone’s reasoning more accurate, yet they spread with ease. Their adaptive trait lies in the pleasure they produce: people enjoy them, remember them, and pass them on. In this sense, some memes persist largely because they feel good.

Then there are memes whose main adaptive trait is social signalling. These are not memes that make their carriers more accurate in understanding or communicating reality. Instead, they pay off by signalling membership to a particular group or elite circle, which turns out to be a exceptionally effective survival strategy. For these memes, fitness is mostly about how much they enhance bonding inside a certain community, and how sharply they differentiate insiders from the out-group.

Think of Scientology’s talk of engrams and thetans, or the layers of esoteric knowledge that structures the worldview of cults. The point is less whether these concepts track reality, and more that knowing, affirming, and performing them marks you as one of us rather than one of them.

On these grounds, difficulty is not a bug; it’s a feature. The more costly a meme is to acquire—time, effort, background knowledge—the more credible it becomes as an in-group signal. The logic mirrors costly signalling in biology, where the expense of a trait helps guarantee its honesty (the peacock’s tail being the classic example). And once a meme starts functioning this way, power structures can form around it to promote and preserve it.

2. The Profound Implications

I believe an evolutionary picture of philosophy can help answer some questions that have been haunting the discipline, like what philosophy should aim at, why disagreement is so pervasive, what it even means to be “right” in philosophy, and whether anything like progress is possible.

Indeed, recognizing the simple truth at the basis of the framework presented in Section 1, I think a broad agreement can be reached: philosophy is not well described as an activity aimed at cultivating ideas merely because they are pleasurable or aesthetically satisfying (that’s the domain of art) nor should it be devoted to producing memes that spread primarily by signaling in-group membership.

Instead, philosophy (as a meme itself) should be understood as intended towards the construction and refinement of useful memeplexes: frameworks that make us more accurate, improve and reorganize the memes we already carry, and deepen our understanding of the phenomena our memes are meant to capture.

The difficulty then becomes that philosophers are working under unusually weak feedback conditions. Unlike fields such as computer science or plumbing, philosophy does not come with a verifier; an immediate, unambiguous test of whether the concepts proposed actually work (that would allow for any charlatanry to be exposed quickly). In philosophy the importance of a certain memeplex—its contribution to human flourishing, or to more accurate ways of thinking—may only become clear after many centuries have passed.

Because the signal regarding which concepts will prove genuinely useful is so faint, we should expect a proliferation of memes that spread for other reasons: they advertise in-group membership, reward academic faction-building, or flatter the sensibilities of a particular circle. Under these conditions, orthodoxies can form and persist sustained less by argument than by ipse dixit.

This makes the (good) philosopher’s job very hard. Often they are fighting an uphill battle, with no immediate reward for being right—something they can’t be fully certain about in the first place. To do their job well, they have to eat frustration for breakfast. In an environment like this, with a very faint verifier, widespread and unresolved disagreement is not surprising; it is exactly what we should expect.

(In my view, much of the recent defense of continental philosophy offers a particularly vivid example of orthodoxy at work. But analytic philosophy is not exempt from the same pressures, and has its own versions of factions, fashion, and institutional self-reinforcement.)

Given these dynamics, we should expect a great deal of the philosophy we encounter to be misguided. Still, that doesn’t mean genuinely useful philosophical memes can’t be created; we have real examples. Progress is possible, it’s just driven by an excruciatingly weak signal, buried under a ton of noise.

3. Conclusions

What I’m proposing is a radical reframing of the meme we call philosophy, one that, I think the field would benefit from keeping firmly in view. If we analyze philosophy from a scientific, naturalistic perspective as the production, competition, and inheritance of memes, we gain a vantage point from which long-haunting philosophical questions may begin to look less mysterious

Unfortunately, I suspect many philosophers will hate the idea of thinking about their discipline in these terms. I can already hear the lines of objection “such an approach performs a quiet erasure: what once exceeded the empirical is flattened into a discourse of causes and correlations, until the question of being is displaced by a machinic inventory of practices and drives” or “ A sociopsychological account may explain the etiology of metaphysical theorizing, but it is orthogonal to (and cannot substitute for) the normative and truth-conditional questions metaphysics is trying to answer; otherwise you’re committing a category mistake and flirting with a genetic fallacy”. But I think it’s precisely this overreaching desire to claim the mantle of metaphysical truth that keeps philosophy stuck in the mud.

In the end all of this is of little importance. Very few are going to quote or reflect deeply on a Substack post of a total nobody. If my attempt at a contribution has any sliver of validity it might only be apparent centuries from now.

Ah, how perfectly philosophy’s incentive structure knows how to nurture the soul!

Subscribe now

Many of Scott’s 2014 essays didn’t just get passed around; they embedded themselves into the collective consciousness of rationalist culture. Some titles still get invoked years later with the kind of serious reverence mathematicians usually reserved for theorems and famous conjectures.

As I read through the pieces, I remember feeling, for the first time, that Scott and his friend Eliezer Yudkowsky weren’t just writing interesting blog posts. They were sketching out the architecture of a community and a movement. The scale of the ambition all suddenly came into focus. It was a little frightening, like stumbling onto the foundations of something that might actually change the world. But there was also a kind of beauty in its audacity.

What they seemed to be reaching for was a place that minimized bias as much as possible, animated by love for rationality and argument, and sustained by the daring hope that human reasoning could be systematically improved. A place you could go to get closer to truth.

I say “frightening” because, in my mind, projects with that kind of utopian flavor usually end in disaster. And yet, despite setbacks and problems, I wouldn’t say the rationalist experiment has gone that way (so far). Sometimes I catch myself wondering whether, beneath the mild California sun and the shared houses of Berkeley, rationalists are quietly living happier, more deliberate lives than the rest of us—enmeshed in a culture we can’t quite imagine from the outside.

Anyhow, here are the rankings of the ten eleven twelve articles of Scott Alexander from 2014 I liked the most:

12. Should You Reverse Any Advice You Hear?

Scott points out that we’re usually drawn to communities of like-minded people. It’s far easier to join a group that celebrates your natural proclivities than one that asks you to resist them.

But this creates a problem. Because the people in our community are likely to be like us, they probably share not only our strengths but also our weaknesses. That can make it difficult for the group to acknowledge, let alone confront, its collective blind spots. So, instead of being challenged, we might end up with a chorus telling us that we’re fine just the way we are. Or worse, that our very flaws are virtues. For this reason, it’s worth asking ourselves whether we’d be better off reversing most advice we’re given.

It’s a classic rationalist point: doubt everything, especially what you cherish most, because bias often nests in the very places where our hearts feel most at ease.

11. Nobody Is Perfect, Everything Is Commensurable

This article makes a striking claim:

Giving even a tiny amount of money to charity is hundreds or even thousands of times more effective than almost any political action you can take.

It’s the sort of claim that deserves a cathedral of reasoning, but instead it gets only a sketch on a napkin. I was intrigued but not convinced, these kinds of utilitarian comparisons are seldom straightforwards.

Yet this is also the article that helped popularized a beautiful norm (a Schelling point, in rationalist preferred nomenclature) around giving 10% of what you earn to charity. Within it one can can glimpse the beating heart of rationalism, which can be sometimes hard to notice veiled beneath layers of analysis and abstraction. For this reason, I couldn’t leave it out of my top ten eleven twelve.

10. Archipelago and Atomic Communitarianism

Alexander asks whether the liberal harm principle “Do as you will, so long as it harms none” can truly serve as a sufficient foundation for a stable society, since many cultural conflicts revolve around indirect harms; like advertising and social influence. These aren’t direct acts of violence or coercion, yet they shape the moral landscape and inevitably become the subjects of public debate.

On one hand, such debates are natural in any society that values free expression. On the other, once opposing groups emerge, each would like to enforce its own cultural norms, and that often leads to the suppression of dissent and nonconformity.

So Alexander imagines an “Archipelago”: a world where a minimal central authority exists only to prevent different communities from fighting, while guaranteeing everyone the freedom to move between groups if they want to. In this world, groups could form their communities and then self-govern according to their own values without much infighting.

Of course, the idea carries the strong scent of utopia and Alexander knows it.

Although Archipelago is an interesting exercise in political science, a sort of pure case we can compare ourselves to, it doesn’t look like a practical solution for real problems.

Yet there’s something to the idea, and in some ways, it’s already happening through the rise of subcultures. The essay ultimately becomes a kind of love letter to them (It’s clear that fostering a healthy rationalist subculture was already on Scott’s mind at the time).

Ultimately I find Scott’s vision of a world where everyone can belong to the community they deserve appealing, but I have my doubts on whether our current system can sustain a culture that keeps splintering into ever smaller and more specialized groups.

9. Society Is Fixed, Biology Is Mutable

Contrary to both the left, which often views human traits as socially determined and therefore malleable, and the right, which tends to see them as biologically fixed and therefore resistant to change, Scott argues that many biological levers are actually easier to move than the grand social ones. He suggests we give greater attention to biological interventions like nutrition, environment, physiology instead of assuming that social reform is always the more effective or tractable path to improve things.

The first time I read this article I remember feeling hit with a perspective that felt totally foreign to me. It made me wonder whether I’d been too fixated on social reform as the only real way to make the world a better place.

8. Ethnic Tension And Meaningless Arguments

Practically no one works through every issue from first principles. Indeed, most of the time, we take cognitive shortcuts to form our beliefs. One of the biggest is forming beliefs by association.

For example, if you identify with the political left and know nothing about the Israeli–Palestinian conflict, you might look at what other left-leaning people believe and end up being heavily pro Palestine. Likewise, if you’re on the right and unsure what to think about feminism, chances are you’ll hear mostly negative opinions from your peers and adopt a similar opposing stance.

It’s natural to update your beliefs in the direction of people who feel like they’re on your team; your friends, your family, your community. These are people who usually treat you well, who seem honest in everyday life, so it feels reasonable to extend them some trust in matters of belief. Recognizing this dynamic is important because it’s one of the main ways we form opinions. It’s practically unavoidable: you can’t independently research and fact-check every issue in depth. But if nearly everyone forms beliefs this way, the truth risks being pushed to the margins, overshadowed by loyalty and charm.

Scott takes this idea further. He argues that it’s not just beliefs that carry “tribal karma,” but concepts and words themselves. Certain terms acquire positive associations for one group and negative ones for another (e.g. “privilege” for the left, “patriotism” for the right) and this shapes people’s reasoning in strange ways.

For instance, today the word “vaccine” has accumulated bad karma on much of the political right. As a result, a patriotic member of the military might be more likely to oppose vaccines simply because the concept itself has become loaded with negative tribal meaning.

In addition, people exploit this effect by throwing around charged words to create guilt by association. Soon everyone’s arguing about whether something is inclusive enough or too woke, and no one’s asking the question that really matters: does it actually make the world a better place?

7. In Favor of Niceness, Community, and Civilization

Scott is accused of “pussyfooting around with debate-team nonsense” by Andrew, a commenter espousing an approach to politics that rejects rational argument in favor of rhetorical and even physical warfare. According to Andrew, one should use every weapon in their arsenal to destroy their political enemy, because being realistic means understanding that politics is a form of war. From his perspective, rational debate looks like an attempt to avoid the battle, a luxury only privileged people can afford.

In response, Alexander argues that niceness is a surprisingly powerful force. One that can spontaneously spread and generate cooperative, game-theoretic norms among its adherents. In the long run, it often accomplishes more than open conflict ever could.

Carey portrays liberalism as Elua, a terrifying unspeakable Elder God who is fundamentally good.

Moldbug portrays liberalism as Cthulhu, a terrifying unspeakable Elder God who is fundamentally evil.

But Andrew? He doesn’t even seem to realize liberalism is a terrifying unspeakable Elder God at all. It’s like, what?

6. Beware the Man of One Study

Scott offers a simple warning: beware of anyone who cites a single study to prove their point. Given the natural variation in experimental design, it’s almost guaranteed that some study somewhere will report a statistically significant correlation between any two variables.

And even if someone cites multiple studies, you still need to ask whether they’re presenting a representative sample of the research, or just the papers that support their point.

It’s not so much “beware the man of one study” as “beware the man of any number of studies less than a relatively complete and not-cherry-picked survey of the research”

All of this is very true and good, so much so that the article’s title has become a kind of catchphrase used to ward off cherry-pickers.

5. I Can Tolerate Anything Except the Outgroup

I Can Tolerate Anything Except the Outgroup is a meditation on tolerance and on the uncomfortable fact that we all, often without realizing it, build echo chambers around ourselves. We’re quick to accept distant groups who don’t share our daily lives, yet far less forgiving towards the “near” outgroups: the people within our own society who see the world differently.

This dynamic is especially evident in politics, where people on opposing sides often harbor deep animosity toward each other, yet this is rarely framed as a form of discrimination. Scott invites us to reexamine that assumption, citing studies that compare racial discrimination with partisan discrimination, and reveal that the latter, in some cases, can be even more intense.

Discrimination on the basis of party was much stronger than discrimination on the basis of race. The size of the race effect for white people was only 56-44 (and in the reverse of the expected direction); the size of the party effect was about 80-20 for Democrats and 69-31 for Republicans.

Seeing things through this lens can feel strange, but it offers a very different perspective on politics, one that might help increase civility if it were more widely embraced. Alas, the wider cultural zeitgeist was never really influenced by this article (thanks in part to something discussed in another article on this list) and we are now more divided than ever.

4. The Categories Were Made for Man, Not Man for the Categories

This article teaches what I believe to be an important lesson.

Human categories (like “fish,” “planet,” or “father”) aren’t discoveries etched into nature; they’re conceptual tools, little memes, we invent to serve our purposes. There’s no single, objective fact about where the boundary of these categories really lies, no hidden Platonic realm of perfect fishes—only better or worse ways to draw those boundaries depending on what we’re trying to accomplish.

So, if you find yourself debating whether Pluto is a planet, it’s wise to first recognize that you may simply be using different definitions. Once that’s clear, you can move on to the more productive question: which definition is more useful or makes more sense in context? Otherwise, you risk being stuck forever at the unproductive stage of arguing over arbitrary choices.

This will help you cut through the noise that clouds many of life’s discussions.

3. The Toxoplasma of Rage

Scott Alexander’s Toxoplasma of Rage pinpointed a problem far ahead of its time, a problem that has only grown worse since. This is one of the reasons the appeal to tolerance in I Can Tolerate Anything Except the Outgroup was always destined to fall mostly on death ears.

Posts, memes, and news stories that divide people and provoke anger are far more likely to be shared. When a topic triggers moral outrage or identity-based reactions (“our side” versus “their side”), it spreads faster than calm, nuanced information ever could. The internet, driven by algorithms that reward engagement, amplifies the most emotional and divisive versions of an issue, like a kind of parasitical toxoplasma, rather than the most accurate or constructive ones. And since attention means ad revenue, capitalism itself ends up incentivizing outrage.

It’s probably no coincidence that America is now more polarized than at any point since researchers began tracking polarization in the 1970s and ’80s.

2. The Control Group Is Out of Control

Did you know there are, apparently well-conducted, scientific meta-analyses that report evidence for psychic phenomena?

It’s a reminder of how hard science really is. It’s not enough to avoid mistaking noise for signal; you also have to ensure your methods aren’t subtly confounded. I really like these kinds of discussions concerning the epistemology of science because I want to discover true things in my own line of work.

Inside the article, you’ll find what I believe to be some hard truths, such as:

Studies are going to be confounded by the allegiance of the researcher.

And

The highest level of the Pyramid of Scientific Evidence is meta-analysis. But a lot of meta-analyses are crap.

The ending is pessimistic, but playfully so. It’s clear that we are discovering useful truths about the world and that science, despite all its flaws and difficulties, genuinely works. The pursuit of truth is one of life’s most beautiful challenges, remembering the many pitfalls along the way (as this article does) is what helps us stay on course.

1. Meditations on Moloch

Yes, yes, I know. I’m boring, mainstream, and predictable. I’m putting what is probably Scott Alexander’s most famous article at the very top of my list. What can I say? I think it deserves the spot.

If you aren’t familiar, Moloch represents the god of coordination failures; the force behind how even rational individuals, groups, or nations can become trapped in destructive competitions that leave everyone worse off. Moloch is the arms races. Moloch is the zero-sum scramble for status and credentials. Moloch is the embodiment of the Nash equilibrium in the prisoner’s dilemma: the outcome where both players lose by trying to maximize their own gain. This makes Moloch the arch-nemesis of the rationalist that would like to solve everything through more reason.

One of the great merits of Meditations on Moloch is that it casts a revealing light on an enemy exceptionally skilled at lurking unnoticed in the darkness—an ancient and cunning aeon, forever hiding behind our quarrels, whispering through our hatreds. Too often does Moloch go unnoticed, but once you’ve learned to recognize his presence, you begin to see his grimy tendrils everywhere.

Moloch whose mind is pure machinery! Moloch whose blood is running money! Moloch whose fingers are ten armies! Moloch whose breast is a cannibal dynamo! Moloch whose ear is a smoking tomb!

Subscribe now

The challenge, of course, is that social phenomena like politics are incredibly complex, shaped by countless variables, making it difficult to derive precise, physics-like laws to explain why political factions form and evolve. And yet, it’s hard to resist exploring a few of the more fascinating possibilities.

1. A Model for the Left vs Right Divide

Many have tried to explicate the essence of the divide between the political left and right. Some academics, like Corey Robin, argue that conservatism is fundamentally about defending established hierarchies, while the left seeks to dismantle them in pursuit of greater equality. Thomas Sowell offers a different contrast: the Tragic Vision, associated with conservatism, sees human nature as flawed and unchangeable; thus, social institutions must constrain our worst impulses. The Utopian Vision, associated with the left, views human nature as malleable and perfectible; so by reforming institutions we should be capable of improving both society and individuals. Others frame the divide in moral terms: the left prioritizes reducing harm, while the right emphasizes loyalty and respect for authority.

Still others (as seen in a past post) interpret the left–right divide as a fundamental struggle between progress and tradition. From this view, the left tends to champion innovation and reform. The right, by contrast, values continuity, inherited wisdom, and the preservation of long-standing institutions and cultural norms. Where the left sees tradition as a constraint on justice or liberation, the right often sees it as a stabilizing force, refined through generations of trial and error.

But few have had the audacity to venture into the nebulous question of why this division arises in the first place. Perhaps this is because it’s unclear whether the left–right split is even a stable dichotomy in societies that permit democratic expression (more on this later).

1.1. Scott’s Theory

One internet article that does attempt the perilous leap is The Thrive/Survive Theory of the Political Spectrum by Scott Alexander. Alexander Hypothesizes that

Rightism is what happens when you’re optimizing for surviving an unsafe environment, leftism is what happens when you’re optimized for thriving in a safe environment.

To illustrate this, as is standard practice, he invokes the zombie apocalypse scenario; in a world besieged by an existential threat—where survival is uncertain—people would likely develop a newfound appreciation for guns, the military, and strong hierarchies. They would grow more distrustful of outsiders, more rigid in their social expectations, and more accepting of authority and conformity. In contrast, in a world of abundance and security—such as a post-scarcity utopia— it is natural to suppose that people will instead favor openness, tolerance, egalitarianism, and more fluid social norms. This contrast seems to reveal what might be at the heart of the political divide.

Although Alexander doesn’t state it outright, it follows from his model that much of the reason for the left–right split is experiential. People are shaped by the environments they inhabit: some grow up in contexts that feel secure and full of opportunity, while others are raised in conditions marked by precarity and threat. As a result, they develop fundamentally different intuitions about how society ought to be organized. Put another way, differing life experiences amount to different "training sets," which produce divergent world models in peoples heads. And these divergent mental frameworks create an epistemic chasm, one that ultimately manifests as the enduring divide between left and right.

This theory is appealing, because it would make sense of what thinkers have said about the essence of the left–right divide. Strict hierarchies tend to be more efficient and therefore more useful in times when survival takes precedence over thriving, so we would expect conservatives to favor them over progressives (Corey Robin: check). Similarly, we’d expect those preoccupied with survival to be less inclined toward utopian social experimentation than individuals who enjoy greater security (Thomas Sowell: check). Moreover, thriving 'lefties' should be more likely to question or play with social norms, while conservatives—operating from a 'survival' mindset—would be expected to adopt a more cautious and risk-averse stance. And this, too, is what we observe.

1.2. Academic Corroboration

A notable body of work that both anticipates and supports Alexander’s model can be found in the writings of political scientist Ronald F. Inglehart (which were introduced to me by fellow Substacker ). In the 1970s, Ronald F. Inglehart began exploring how people’s values change as societies experience rising levels of economic security compared to the past. His research found that generations raised in greater affluence tend to prioritize different concerns—such as self-expression, quality of life, environmental protection, and participatory democracy—what he and his colleagues termed postmaterialist values. Impressively, in The Silent Revolution (1977), they even predicted a shift in voting patterns:

Our data on the relationship between value type and one's sense of belonging to the Left or Right have striking implications concerning possible realignment in the social bases of politics. Traditionally the bulk of support for the parties of the Left has come from the working class, with most of the middle class voting for parties of the Right in nearly all Western countries. And the persistence of established party loyalties is such that this pattern still held true among the populations surveyed in 1973. Yet there is a consistent tendency for the predominately middle-class Post-Materialist types to identify with the Left, while the less affluent Materialists are more likely to place themselves on the Right. In the long term, this could lead to the neutralization or even the reversal of familiar class-voting patterns.

Some might say that this is precisely what has happened over the past decade or so, seemingly corroborating Alexander’s model. Did I already say impressive?

Another highly relevant body of work comes from Michele Gelfand. Gelfand’s research on cultural tightness and looseness examines why some societies develop strict social norms and strong enforcement mechanisms, while others embrace more flexibility and tolerance for novelty. Her central finding is that the degree of “tightness” in a culture is strongly shaped by its ecological and historical conditions. Societies that have faced frequent and severe collective threats—such as war, natural disasters, disease outbreaks, or chronic resource scarcity—tend to evolve tighter norms. In these environments, coordinated behavior and social order are essential for survival, so rules become clearer, punishments harsher, and conformity more valued. In contrast, societies with relative safety, stability, and resource abundance can afford looser norms, granting individuals more behavioral freedom and encouraging diversity, experimentation, and nonconformity.

Through large-scale cross-cultural studies spanning dozens of countries, Gelfand and her colleagues found systematic patterns linking historical threat levels to contemporary social behaviors. Tight cultures—such as Singapore, Japan, and South Korea—often feature meticulous rule-following, lower tolerance for behaviors that deviate from the norm, and stronger social sanctions for violations. Looser cultures—such as Brazil, the Netherlands, and New Zealand—tend to prize self-expression, accept a wider range of behaviors, and operate with fewer formal and informal constraints.

Importantly for what concerns us presently, her studies suggest that tight–loose differences correlate with political inclination. Conservatives tend to favor tighter norms—more order, more rules, and stronger sanctions—particularly when they perceive society as being under threat. Liberals, by contrast, are more inclined toward looser norms—greater tolerance for diversity, weaker sanctions, and an emphasis on individual freedom—especially when they perceive society as safe and stable. So, has Scott Alexander’s model pretty much been proven right?

1.3. Objections

In my view, the thrive–survive theory of the political spectrum does a surprisingly good job of explaining in a simple manner why the left-right political divide exists. Especially considering that in a field as messy and tangled as politics, even a half-decent model is already punching way above its weight class. That said, there are a few nagging doubts I’d feel remiss not to toss on the table for your consideration, dear reader.

a. Is the Left–Right Divide Universal?

If these kinds of explanations were correct, we’d expect to find a recognizable Left–Right split in every society that allows for some forms of democratic expression. But is that actually the case?

There are certainly historical examples that seem to fit: the Legalists versus the Confucians in ancient China, the Royalists and Parliamentarians in the English Civil War, Athens and Sparta in classical Greece—each embodying contrasting political temperaments that map, at least loosely, onto our Left–Right concepts. Still, the boundaries are always fuzzy when dealing with these kinds of categorizations.

b. Weren’t Leftists the Working Poor?

If the Left corresponds to the “thrive” side of the thrive/survive theory, how do we square that with the fact that, historically, leftist politics has often been associated with working-class movements?

One possible way to answer this question is to point out that while the base of many left-wing movements has been working-class, the ideology itself has often been articulated, refined, and promoted by people from outside that class. In other words, the intellectual architects of left-wing thought seem to have disproportionately come from educated, relatively comfortable backgrounds (although there are exceptions). This would make sense under the thrive–survive framework: those in more secure positions are freer to imagine and advocate for sweeping social change, even if the changes they champion are meant to benefit less secure groups.

2. A Model for Authoritarianism

Alexander’s theory has some compelling elements, but it obviously can’t cover the full carnival of political phenomena, with all the quirks and outright absurdities. One particularly relevant aspect that seems worth folding into any political model is the authoritarian axis of the political spectrum. We know, after all, that democracies sometimes like to hand power to charismatic authoritarians who tear down institutions and, in the worst cases, racks up body counts in the millions. This feels like the kind of thing it would be nice to explain. Can we augment our model to capture these observations?

Luckily, authoritarianism has been extensively studied since World War II, and on this substack we have already poked at some of the relevant research (see here and here) so I’ll resist the urge to repeat myself too much.

The short version of what impresses me in the data we have collected is that authoritarian inclinations seem to correlate strongly with cognitive rigidity and black-and-white thinking, both fueled by our brain’s deep craving for conserving mental resources. Simpler models of the world are cheaper to run, so our minds crave them, and nothing hits the spot like reframing reality as a clean ‘good guys versus corrupted hellspawn’ comic book story. That’s why populist rhetoric is the authoritarian’s favorite toy.

Once you’ve cast the world that way, the next move is obvious: crush the villains, watch them scatter before you, and install your flawless government that all the good people will naturally adore. From inside that story, authoritarianism doesn’t just seem tempting, it seems downright logical.

(There is, however, a novel thought I had recently regarding populism that I wanted to share. It seems like there could be a time and place where an “us versus them” mentality, coupled with the rhetoric utilized by populism, is quite appropriate, and that is…when you are actually living under a tyranny.

In a tyranny, the government thrives on lies. Elites seek to control or silence dissent through violent means, while a small circle clings to nearly all the political power. In that environment, communication must be fast, simple, and often clandestine. Here populism suddenly stops looking like cheap sloganeering and starts looking like smart messaging strategy. “The corrupt elite versus the people” isn’t just rhetoric anymore, it’s a pretty accurate map of reality. For North Koreans, it seems like exactly the right tactic.

So, it wouldn’t surprise me if this kind of response had some evolutionary roots. Our primate cousins have often had to deal with oppressive dominance structures. Among species like capuchins and chimpanzees, coalitions form to resist overbearing alphas, sometimes by ganging up and killing them. Humans might carry a similar instinct—a psychological predisposition that flips when conditions demand it, priming us to rally together and overthrow oppressive rulers by whatever means necessary.)

2.1. Shouldn’t Authoritarianism Be Mostly Right-Wing?

Conserving mental resources doesn’t seem like a thriving thing; it seems more like surviving. So if Scott’s model is mostly right, how do we explain that we’ve definitely seen authoritarianism coming from the left too?

I’m not too sure. Maybe the left really does dabble less in overt authoritarianism, like some scholars like to claim. But another possibility is this: left-wing authoritarianism seems to arrive through the promise of paradise after a transition period. It conjectures an extraordinary payoff and does so by brutally simplifying the expected value of forcing societal progress at all costs.

Framed this way, the straightforward message isn’t just tempting to people who are struggling—it also resonates with the idealistic thrivers, who love imagining a renewed society free of injustice and are not immune to the appeal of simple world models. Right-wing authoritarianism, by contrast, is less about building paradise and more about preventing collapse. It offers order, stability, and a return to traditions and laws.

So, if the right simplifies in the direction of avoiding dystopia, the left seems to simplify in the direction of achieving utopia. And this may be how the authoritarian dimension ties into the left–right spectrum.

3. Utilizing Our Little Model

If we accept the premises of this simple model at a normative level—that it makes sense to lean conservative when things are going poorly and progressive when we’re thriving—then we can use it to generate rough estimates of which political positions might be most appropriate for different societies.

In times of rapid technological progress that enhance human flourishing, our model tells us the time is ripe to question existing norms and explore how society could be reshaped.

Conversely, in a society facing collapse, the model-recommended approach would be to adopt conservative measures, act with greater caution, and prioritize swift decision-making.

This line of thought can also shape how we think about politics in other countries. In some Middle-Eastern or African nations, there may be too little “thriving” for Western norms like democracy to take root, as Democracy could depend on a baseline of safety and prosperity before it can gain widespread acceptance, and without that foundation, trying to force it may simply be premature.

What’s striking is that none of these judgments seem to require strong ideological commitments or partisan leanings. We can just look at the data—using rough proxies for technological progress like GDP—and let the ‘theory’ guide us.

It can feel jarring to think that in some situations the right stance should be more authoritarian, while in others it might be more progressive. It runs against the usual fixed identities people adopt in politics. I suspect it’s a completely different perspective to what most people are accustomed to, which seems reason enough to lay it on the table for your consideration.

At this point, I’d feel remiss if I didn’t speculate about what our theory suggests for Western countries. Most of them today have modest economic growth and reasonably stable democratic governments. So if we use this approach as a rough guide, I think we get something like the following.

All standard contraindications apply.

Subscribe now

Unfortunately most critiques of shrimp welfare so far have been pretty bad, some just lashing out at Bentham’s Bulldog with little to no true argumentation. I’ll try to do better by presenting what I think is a very strong objection, one I’ve been turning over in my mind for a long time. But first, let’s lay out the argument for the shrimps.

1. The Argument for Shrimp Welfare

The essence of the argument for shrimp welfare is straightforward: if we accept that pain and suffering are bad, and that shrimp are capable of suffering, then the scale of their suffering is staggering since billions are killed each year in brutal ways. From this perspective, donating to reduce shrimp suffering may be one of the most effective charities possible—perhaps even the most effective—since it could be the cheapest way to alleviate the greatest amount of suffering.

Along the same lines, Bentham’s Bulldog—real name Matthew Adelstein, an Oxford student and leading advocate for shrimp welfare, who has charmed many on Substack (myself included)—has recently argued that insect suffering is “the worst thing in the world by far.” His reasoning (in his own words) runs as follows:

1. Pain and suffering are bad.

2. If something is responsible for the overwhelming majority of the world’s pain and suffering in expectation, that thing is very bad.

3. Insect suffering is, in expectation, responsible for the overwhelming majority of the world’s pain and suffering.

Matthew points out that there are an estimated 10^18 insects in the world—a number far larger than even the shrimps—with most living lives filled with suffering before meeting an early death. And since any suffering is bad, the sheer scale makes this the worst thing in the world. He takes this reasoning extremely seriously, so much so that he has even implied it might be better to save a trillion insects from extreme agony than a human being.

At this point, it is reasonable to wonder whether insects and shrimp are even capable of feeling actual pain, but this won’t be the focus of my objection. I think Matthew is right that there’s a fair chance they do, at least to some degree. And keep in mind, in the case of insects, even if we assume they feel only the tiniest fraction of pain—say, as little as 0.001% with respect to humans—their sheer numbers means that, once all their suffering is added up, it will still easily outweigh human suffering by a wide margin.

Of course, there are other objections that could shift the numbers by several orders of magnitude: for example, when a human dies or suffers, other humans usually suffer too, as K-selected species likely have greater capacity for empathy than r-selected ones, and it seems like we may be ignoring this magnifying effect in our aggregate calculations. Still, I want to broadly accept all of Matthew’s assumptions about shrimp and insect consciousness and their capacity to suffer, because I believe there is a much stronger challenge to his argument, one that comes directly from the utilitarian framework itself.

2. The Utilitarian Case for Why a Human is Worth More than a Gazillion Shrimps

An omission I think Matthew sometimes makes in presenting his calculations—and one that many critics of utilitarianism also fall prey to in other cases—is treating the felicific calculus as if it simply stops at adding up the suffering of different beings, weighted by their capacity to feel it. But if we are reasoning in a genuinely consequentialist way about the morality of actions, what matters are the full consequences of actions taken in their entirety.

For a consequentialist the question isn’t just how much suffering or happiness a human experiences compared to a shrimp on average. What matters is the full range of first- and second-order consequences arising from a human’s suffering or death relative to that of a shrimp. And once we account for these broader effects, I believe it’s fairly easy to make an argument for why humans play a far greater role in moral calculations than shrimp or insects. Allow me to get specific.

For utilitarians like Matthew and me, much of the universe looks like a terrible battle against suffering. Suffering is everywhere and overwhelming. As Matthew might put it, quadrillions of creatures are screeching in pain across trillions of galaxies. Animals are eaten alive every day. Entire species have come and gone, living lives of hardship before going extinct without ever managing to improve their condition. Even most of human history has been marked by misery, disease, and death—and in many places, it still is.

Yet, well-being and happiness exist. It is possible to experience consciousness as something joyful and worth having, and in an astonishing and still somewhat mysterious turn of events—driven by contingencies and idiosyncrasies we don’t yet fully understand—humans appear to be the only species (as far as we know) that has managed to meaningfully reduce the aggregate suffering of its own kind. Indeed, over time, and across many parts of the world, we've seen forms of progress: lower infant mortality, longer life expectancy, fewer deaths from disease, and far less extreme poverty. We've developed pain relief, mental health care, safer working conditions, and social safety nets. While suffering hasn’t vanished, and many forms persist or evolve, the average human today probably endures less pain and hardship than in most of recorded history.

This is absolutely incredible. Indeed, so much so that it’s contentious and some don’t believe it. But it looks to me like it’s true, and I think many (counting among them Matthew himself) would be inclined to agree. At the very least, it seems that through our ingenuity we have the potential to make the world a much better place, and, crucially, we are the only beings we know of capable of such an astonishing feat.

In my wildest dreams, I see a future where we carry happiness to distant worlds, unshackle innocent creatures from the suffering they are bound by, and slowly wipe agony from the face of a grateful universe. Within humanity lies the possibility of intergalactic utopia, and as a utilitarian, I can’t help but think this is where we should be headed.

But what does all of this have to do with the shrimps?

The point is that if we’re going to make a felicific calculation comparing the utility of a human with that of a shrimp, a central part of it must reflect the fact that humanity has the chance to significantly reduce future suffering. Shrimps and insects, unfortunately, do not. More concretely, inflicting pain or death on a human prevents them from contributing to the advancement of society—whether by raising children, doing useful work, or helping create the conditions for technological innovation. And so far, human society is the only force we know of that has ever managed to combat suffering and meaningfully increase well-being on a massive scale.

So, how would this consideration change the consequentialist calculus? It’s going to be hard to pin down expected values when we’re talking about something as vast as a shot at intergalactic utopia. Still, I think we can attempt a rough calculation to make the impact of this shift clearer by focusing on a narrower question. How much expected suffering would the universe have to undergo in order for shrimps to occupy the privileged position humans now hold—the position of being able to meaningfully affect aggregate suffering on a global scale?

Let’s suppose that intelligence of the kind humans possess arises about once every 100 million years. There are roughly 2 million species on the planet right now, so a rough chance that the evolutionary descendants of shrimp would reach human-level intelligence is about 1 in 2 million. That means it could take another 100 million years of the same kind of suffering the universe currently endures just to have a 1-in-2-million chance that shrimp would ever reach the position humans now occupy, with the ability to meaningfully reduce suffering. It wouldn’t surprise me if, following this line of reasoning, the average human turned out to be ‘worth’ a googolplex more than a shrimp in consequentialist terms.

In other words, once we account for significant second-order effects of actions, the worst thing in the world by far doesn’t seem to be shrimp or insect suffering. It seems to be the extinction of the only beings we know of who are capable of doing anything about suffering at a large scale in millions of years: humans.

3. An Incidental Justification for a Common Intuition

I believe the line of reasoning I’ve just outlined provides a justification for a common intuition: that intelligence matters in moral calculations (notably, this point has been raised on Substack in a pair of articles by and , albeit without a thorough justification).

In the opposing camp, (one of my favorite living philosophers) has famously argued that this view is deeply implausible. To make his case, he constructed the following graph, intended to illustrate how improbable it is that intelligence carries great moral weight.

For Huemer, it seems highly implausible that things could work this way—so much so that he thinks it would be very challenging to explain why such a graph could be true.

Can someone even think of a possible explanation? If you just think about this theory on its own (without considering, for example, how it impacts your own interests
or what it implies about your own behavior), would anyone have thought
this was how it worked? Would anyone find this intuitively obvious? As a
famous ethical intuiter, I must say that this doesn’t strike me as intuitive at all.

Section 1 of this article provides a possible explanation: when a human suffers intensely or dies, their ability to participate towards the advancement of society in a meaningful way is severely impaired. And since human society is the only force we know of with the potential of reducing aggregate suffering on a massive scale, this consequence carries significant ramifications with profound moral weight.

Additionally, this explanation addresses two aspects of the graph that Huemer finds puzzling;

i. The graph starts out almost horizontal. But somewhere between the
intelligence of a typical cow and that of a typical human, the graph
takes a sharp upturn, soaring up about a million times higher than where
it was for the cow IQ. This is required in order to say that the pain
of billions of farm animals is unimportant, and yet also claim that
similar pain for (a much smaller number of) humans is very important.

ii. But then the graph very quickly turns almost horizontal again. This
is required in order to make it so that the interests of a very smart
human, such as Albert Einstein, don’t wind up being vastly more
important than those of the rest of us. Also, so that even smarter
aliens can’t inflict great pain on us for the sake of minor amusements
for themselves.

Sure, this is a logically possible (not contradictory) view. But it is very odd and (to me) hard to believe.

But both of these points, I believe, can be accounted for without much difficulty by our explanation: cows would need millions of years of evolution before they could even have the chance of affecting suffering on a massive scale, whereas even an average human has a shot at giving birth to the next Albert Einstein (not to mention all the other significant ways an average human contributes to society).

I would also add that I find this explanation highly intuitive—and I suspect many others will too. In fact, I’d conjecture that if shrimps all had Einstein-level IQs, were exchanging complex ideas, and producing world-changing technological innovations, humanity’s view of them would be radically different. I highly doubt we’d be eating them, and I suspect even Bryan Caplan and Lyman Stone would immediately grant them moral consideration on par with humans.

We value intelligence because we intuitively understand it holds the potential to lead to more future well-being and this matters greatly in any kind of felicific calculation.

4. Objections and Doubts

I’m rarely extremely convinced by any argument, and that includes my own. So what should we make of the argument I’ve just put forward? One objection that comes to mind is that it rests on a very optimistic view of humanity (so optimistic, in fact, that I surprise myself, since I usually think I’m too cynical). The assumption I’m talking about is that humans have a real chance to gradually make things much better in terms of suffering and happiness for all. Why would I find this plausible?

Well, as already mentioned, it seems we may indeed be in the process of making things better—at least for ourselves. Secondly, I think we may possess innate moral instincts that incline us to care about the suffering of other beings, so we will want to make things better for sentient creatures (see my articles and videos on utilitarian naturalism for a more in-depth view).

Relevantly, I also think these instincts manifest most strongly with regards to animals among individuals who grow up in loving, thriving environments, where they need not be preoccupied much with primary needs. So, from my point of view, if humanity continues to improve its own condition, then more and more people should eventually become more empathetic toward animals. I would conjecture that if we reach a post-scarcity society, humanity will devote mountains of additional attention to the problem of animal suffering. Indeed, ultimately, I believe that, if socio-technological progress continues to advance, the shrimp gang and Matthew himself will be vindicated: at some point, most of humanity will take animal suffering very seriously indeed.

What I mean to say is that not only shrimp and insect welfare may not be the most important issues in the world, but also that, in practice, the shortest road to improving their welfare could be by reducing human suffering.

In my mind (and I believe much academic research broadly supports the view) this would not be too dissimilar to how the Western world abandoned slavery or extended more rights to women. It wasn’t a sudden spiritual illumination or moment of divine insight, but rather the fact that technological and intellectual progress made the environment ripe for change. Once the right conditions were in place, societies were correctly able to recognize the old norms as immoral—and, just as importantly, to bear the costs of doing better.

At this point, I think there’s room for some confusion. One might object that my argument itself betrays the fact that humans will always care more about their own suffering than that of other species, because they are more intelligent, and so shrimp suffering will never truly end. But I don’t think that follows. My view is consistent with the idea that as human suffering gradually decreases, the gratuitous suffering of animals will rise in relative importance.

4.1 On Felicific Calculations

At one point, while writing the argument in Section 1, I had a long paragraph of calculations that I eventually cut in favor of a shorter, more intuitive approach. It turns out that estimating the full ramifications of actions can be difficult: consequences beget further consequences, and tracing and weighing them all appropriately may prove to be very complex.

This made me realize that the challenge I’m raising to shrimp welfare can be reformulated in a weaker, but perhaps more compelling way to some: consequentialist calculations are rarely straightforward, and it is often quite difficult to determine with certainty which action is truly best when considering all its possible ramifications.

5. Why I Decided to Write this Article

As some of you may have gleaned, I don’t think shrimp suffering is morally neutral. In fact, nothing in the arguments or explanations in Sections 1–2 implies so. I agree that shrimp, and insect suffering are morally significant—assuming, as I think is plausible, that they are capable of suffering. So why did I feel the need to write a critique to shrimp welfare?

First, because I believe what I’m saying may be true, and trying to say something true is usually worthwhile. Second, because I think it’s important to highlight the possibility that the shortest road to reducing shrimp suffering may actually run through reducing human suffering. And third, because I sometimes worry about the direction the conversation around shrimp and insect welfare seems to be going.

Despite never having attended an Effective Altruism meeting or participated in their forums, I deeply sympathize with the idea of helping others effectively. More than that, I see in the movement a kernel of something humanity desperately needs: a noble goal.

The way I see it, utilitarianism offers us that worthy objective—one that can give meaning to anyone’s life, that gives a lot of meaning to mine, and that I believe could guide many others toward well-being and purpose, especially in today’s world.

So, when I come across discussions about whether we should sacrifice a single human to save a trillion insects, I worry that people will come away thinking this is what utilitarianism ultimately leads to, or that this is what it’s all about. (Indeed, there are already critics using shrimp welfare as a way to try to discredit utilitarianism altogether.) And in the process, people who might otherwise benefit greatly from engaging with the ideas behind effective altruism could be discouraged from ever doing so.

Worse, I worry that some might even conclude that the most moral course of action is to blow up the world in order to stop insect suffering. Forgetting that performing crude utilitarian calculations like these is the pastime of comic-book villains.

Lest I be misunderstood, I don’t think advocacy against animal suffering is wrong. For what it’s worth, it makes my stomach turn to know that we engage in practices like eyestalk ablation, which involves crushing one (or both) of a female shrimp’s eyestalks to increase egg production, or that so many shrimp die slow, agonizing deaths at harvest.

In fact, I felt so guilty writing this article that I donated $100 to shrimp welfare. Enough, I’m told, to help approximately 150,000 shrimp experience less suffering.

What I’m saying is that I think there’s a needle to be threaded here—and if we overstate how important the case for shrimp and insect welfare is, I worry we risk tipping into a kind of radicalism that has the chance of becoming counterproductive.

Subscribe now

It so happens that, in this meme, I find myself sympathizing with the crying nerd at the top of the bell curve. Let me try to explain why.

1. Types of Moral Intuitions

A conceptual taxonomy of moral intuitions was attempted over a century ago by Henry Sidgwick in The Methods of Ethics.

Sidgwick proposed three categories: perceptual, dogmatic, and philosophical intuitions.

• Perceptual intuitions are immediate, emotional reactions to particular actions. If you witness a murder and feel a sick twist in your stomach, that's a perceptual intuition at work; a kind of moral feeling, direct and unmediated.

• Dogmatic intuitions are common-sense moral rules that most people treat as self-evident: statements like “stealing is wrong” or “we should keep our promises.” They apply to types of actions rather than specific cases and form the backbone of what people usually call common-sense morality.

• Philosophical intuitions are more abstract in nature. They don’t refer to particular actions or categories of actions, but instead express principles that are meant to be universalizable — applying to all agents, in all situations. Also, they appear true or important upon self reflection. These are things like “suffering is bad”, or as Sidgwick might exemplify:

  “It cannot be right for A to treat B in a manner in which it would be wrong for B to treat A, merely on the ground that they are two different individuals, and without there being any difference between the natures or circumstances of the two which can be stated as a reasonable ground for difference of treatment”.

Sidgwick regarded philosophical intuitions as the foundation of moral reasoning. These were supposed to be principles that, upon careful reflection, no reasonable person would reject. So, he believed, they should serve as the starting point for building a systematic moral theory. And, when pursued to its logical conclusion, Sidgwick thought that, this method of ethical reflection culminated in utilitarianism.

2. What to Make of This?

As a rule of thumb, I never trust philosophers. Which naturally raises the question: what are we to make of this seemingly arbitrary taxonomy?

Well, for starters, it's intelligible. Which is already pretty good by the standards of 19th-century philosophy. It does seem possible to coherently categorize moral thoughts in the way Sidgwick proposes.

Still, even if the classification makes sense, it’s not immediately clear why we should consider it especially valuable. Why, for instance, would we have different types of moral intuitions? Why should we treat philosophical intuitions as more foundational than others? And why would some moral principles be “self-evident” upon reflection?

Also, Sidgwick offers little in the way of argument for why we should all have the same philosophical intuitions; he mostly takes it for granted. Perhaps because, to him, it was self-evident (ha). But what about the rest of us mere mortals?

3. The Latest on Our Brain

Contemporary research in evolutionary psychology and cognitive neuroscience suggests that the human mind is not a single, unified block, but rather a collection of specialized cognitive systems—semi-independent, domain-specific networks shaped by natural selection to solve distinct adaptive problems. These systems govern different aspects of behavior, such as threat detection, social reasoning, language, mating, and foraging. Each system operates according to its own functional logic, processing specific types of information and producing outputs tailored to its evolutionary role. While these systems often work in coordination, they can also compete for our limited cognitive resources such as attention, working memory, and executive control.

Importantly, these cognitive systems are not monolithic; rather, they are composed of smaller, specialized subcomponents that carry out specific functions in service of the broader system. For example, the threat detection system—which monitors and responds to potential danger—relies on lower-level mechanisms that rapidly evaluate sensory input for signs of imminent harm. One such mechanism is the fear response, which automatically mobilizes attention, physiological arousal, and motor readiness when a threat is detected. This fear response operates largely outside conscious awareness, prioritizing immediate survival by preparing the organism to freeze, flee, or fight. In this way, subcomponents act as functional building blocks within larger cognitive architectures, allowing complex behavior to emerge. They are biological heuristics embedded in us by millions of years of evolution.

4. A Modern Reappraisal of Sidgwick

One possibility that would begin to shed light on some of the questions Sidgwick leaves us with is that moral intuitions (intuitions regarding how we ought to behave towards others) emerge from a cognitive system that governs our normative standards of good and bad cooperation. Rooted in our social nature, this moral system would have the goal of informing and guiding our conduct towards other beings.

Working under this assumption, philosophical intuitions would be natural candidates for some of the most ancient cognitive dispositions within our moral system, reflecting its core function. This helps clarify why such intuitions would serve as fitting foundations for a normative ethical theory: given that they would be universally shared among humans.

When comparing different moral intuitions, this idea doesn't seem that far-fetched. For instance, consider the contrast between the intuition that “causing unnecessary harm to others is wrong” and the intuition that “not bowing when you meet your elders is wrong.” The former appears more deeply rooted, a candidate for an innate philosophical intuition, something likely to be recognized as wrong across cultures—while the latter seems more culturally specific. But I think we can push the analysis even further.

If you’ve been following some of my other articles on morality, you’ll know that Harvard philosopher Joshua Greene places significant emphasis on what he calls deontological intuitions; those related to duty, rules, and rights. He argues that these intuitions function as a kind of System 1 morality: fast, efficient, and often unconscious heuristics that help us navigate the complexities of social life.

Sidgwick’s dogmatic intuitions appear to closely resemble the deontological intuitions Greene describes, leading one to suspect they might serve a similar functional role. If that's the case, it would offer a plausible explanation for their origin: they are cognitively efficient heuristics (or Schelling points) designed to help us act in line with our deeper, foundational philosophical intuitions, while accounting for our bounded cognitive capacities. Without such shortcuts, we might be paralyzed by moral deliberation—forced, for instance, to evaluate in every situation whether there are “reasonable grounds to differentiate treatment” between person A and person B, as in the example of a plausible philosophical intuition we used from Sidgwick.

We might even go further and suggest that many dogmatic intuitions such as 'failing to bow when meeting one's elders is wrong' seem like man-made solutions to practical social problems, and therefore appear more learned than innate.

And what to say about perceptual intuitions? Since perceptual intuitions, as defined by Sidgwick, arise as automatic emotional responses, they seem well-suited to refer to subcomponents of our moral cognitive system: functional components of our multi-layered moral architecture.

Take, for instance, the perceptual intuition of feeling guilty. Guilt likely evolved as a mechanism to promote prosocial behavior and maintain social cohesion. By causing individuals to feel bad when they violate social norms or harm others, guilt motivates reparative actions—such as apologizing or making amends—which in turn strengthen group bonds and cooperation. In this sense, guilt operates as a moral mechanism, a functional subcomponent of our broader moral system.

To recap, standing on the shoulder of giants, we have just reconceptualized Sidgwick in the following way:

Perceptual Intuitions
• Origin: Innate
• Role: Subcomponents of the moral system (biological heuristics)

Deontological Intuitions
• Origin: Learned
• Role: Memetic heuristics due to bounded cognition

Philosophical Intuitions
• Origin: Innate
• Role: First principles of the moral system

We have thus attempted to outline a neuroscientific account of why we appear to experience different tiers of moral intuitions, an account under which Sidgwick’s categorization begins to make more sense.

At this point, we should ask ourselves an important question: Are we in over our heads? The answer, of course, is yes. Nonetheless, sometimes we must have the courage to hypothesize, and to share those hypotheses with the world. Undoubtedly, our model will be flawed, but it might help or inspire others to get closer to a more accurate understanding of what’s really going on. For what it’s worth, our explanation seems reasonable to me. And importantly, it appears falsifiable through cross-cultural studies, which is a definite plus. So, I propose we press on.

5. Self-evident Upon Reflection

Armed with our new reconceptualization of Sidgwick, we can begin to sketch what the process of identifying a claim as “self-evident upon reflection” might actually involve. This schematic method should serve as a way of evaluating moral intuitions to determine how they should be categorized, whether they are the foundational philosophical intuitions or merely dogmatic/perceptual heuristics.

• We begin by selecting a specific moral intuition—a normative statement that comes to mind concerning how people ought to act, or what is good or bad to do.

• Next, we assess its universalizability: we ask ourselves whether the principle appears to holds consistently across cases, or whether counterexamples arise in certain scenarios. Is it merely a context-dependent heuristic, or does it reflect a guiding feature of our moral psychology?

• We then consider whether it plausibly reflects a philosophical intuition tied to our innate understanding of morality, or whether it is instead a self-serving rationalization, shaped by competing impulses or a desire to prosper without much concern for others. Is it a thought provided to us by our ethical system evolved to promote social cooperation and fairness or by our self-interested system aimed at maximizing personal survival, status, and reproductive success?

• Finally, we evaluate how well it fits with other candidate philosophical intuitions. Does it cohere with them, or is there tension?

Given the assumption we are working under — that healthy humans share an evolved moral module whose purpose is to provide a normative understanding of how to behave towards other beings — it appears plausible that this method could bring us closer to uncovering its governing principles.

(The method of identifying what is self-evident upon reflection represents the armchair approach to uncovering our philosophical intuitions. This is possible because we are in a rare case were we are both the inquirers and the object of inquiry. Of course, there is another route to investigate whether philosophical intuitions exist: scientific, cross-cultural studies—though this, of course, requires getting up from the armchair (boo). As more of a scientist than a philosopher, I tend to prefer this second approach, though it is undeniably more effortful and costly.)

6. Some Examples

To understand if a methodology makes sense it is always useful to look at some examples.

a) Thou Shall not Kill

Let’s take the sixth commandment in the Bible: “Thou shalt not kill.” Is this a philosophical intuition, or merely a perceptual/dogmatic one?

Well, is it universalizable in the sense we have discussed before? How does it hold up in edge cases? Consider, for example, an extreme case of self-defense: suppose that during a school shooting you manage to seize a weapon and kill the shooter before he harms others. Is this to be considered immoral because it violates the intuition that killing is bad? This appears quite dubious to me and I’m willing to bet for many other people too.

What seems more plausible is that “thou shalt not kill” was introduced as a moral norm in early civilizations because it functions as a very precious heuristic — a simple rule that generally promotes moral behavior. It reliably guides people toward good actions in the vast majority of situations, since it is very rarely a good idea to kill someone. So, it appears appropriate to categorize the moral intuition that killing is wrong as a dogmatic intuition.

b) Caring About What’s Physically Proximal

It seems like we have more ethical concerns for beings who are in our physical proximity rather than if they are far away, even if we don’t know them. Is caring about people on the basis of how near they are to us in space a philosophical intuition?

Up to a point this behavior makes sense. Our brains don’t have the cognitive capacity to keep in mind every person on the planet, so it’s reasonable that we prioritize what we know better and can more directly influence. However, we also recognize that if a close friend travels to the other side of the world, we wouldn’t suddenly believe it doesn’t matter if something bad happens to them just because they're far away.

This suggests we’re dealing with a perceptual intuition, a subcomponent of our moral system that downregulates concern for distant events. It likely evolved as a way to help us manage cognitive complexity, nudging us to focus on what’s known and actionable. Here, our self-interested systems are likely also at play, competing for our attention; since it pays to cooperate with those who can reciprocate.

In general, I suppose that caring more about what’s physically close is a candidate for a perceptual intuition. It appears to be a biologically shared, functional response rather than a core moral principle governing our moral system. Indeed, there appear to be counterexamples and humans often view helping those who can’t reciprocate as deeply moral (example abound: consider, for example, someone sacrificing their life to save a bunch of strangers).

(caring more for people who are close would also clash with other plausible philosophical intuitions, like the one proposed by Sidgwick previously).

c) Partiality Towards Family

What about the near universal human impulse to prefer family members over strangers?

Again, this makes some sense. In moral decision-making, prioritizing people we know to be good and loving over complete unknowns appears justifiable. But it seems possible to take this too far. For example, if we absolutize the principle to something like 'one should always prioritize family above all else' we can conjure up scenarios were such a principle seems highly dubious. Consider, for instance, a choice between saving the life of a child you don’t know and preventing a small bruise on your own child’s foot, I would bet that practically everybody would recognize that the moral choice is to save the unknown child.

As before this is another case were we would expect there to be tension between our evolved moral system that governs how we should behave towards others and different modules driven to want to propagate our genes at any cost. This makes the ‘partiality towards family’ intuition a candidate for subcomponent debunking.

In general, within this framework, there appear to be three ways of debunking intuitions: the intuition may be a learned heuristic, a biologically evolved heuristic, or a product of a cognitive system that competes with the moral one.

7. Some Objections

a) Our Moral System Is Itself a Heuristic for Gene Propagation

We did a lot of talking regarding our moral system and its core principles, but isn’t our moral system, under this framework, simply an evolved way to enhance gene progagation? In effect it seems like the whole moral system is a “heuristic” for helping us enhance replication success as social creatures. If we carry the logic we've used so far to its ultimate conclusion, it seems we’re led to debunk morality itself.

I don’t think that’s quite right. We are, in a very real sense, physically bound to care about morality in a way that we are not bound to care about gene propagation. Our morality operates according to certain laws that govern its functioning, and we are simply trying to uncover them. Philosophical intuitions serve as a way to axiomatize a model of this morality we intrinsically care about.

b) This Still Seems Arbitrary

Yes, the armchair approach carries a degree of arbitrariness, as it involves me talking to myself and betting on what healthy humans might think about various scenarios. It’s easy to fall into post-hoc rationalizations. What we’ve done here is simply proposed a sketch of a model, that aligns with Sidgwick’s categorization, that explains why we experience different and often conflicting moral intuitions.

7. Back to Common-Sense Morality

But why are we even trying so hard to obtain a coherent model of human morality? We have our common-sense morality, let’s just use that and be done with it.

First of all, if we were to adopt the common-sense morality of ancient societies, we can be fairly certain that our present-day common-sense would be horrified. Indeed, what is common-sense shifts across time and cultures (by the way, our model offers an explanation for this, suggesting that different environments give rise to different learned heuristics). This already exposes the shallowness of relying on common-sense—one is left wondering: which one, exactly?

Secondly, If we could explain how common-sense morality evolves across time and cultures by uncovering its underlying principles, we may be better equipped to avoid ethical missteps and promote more meaningful cross-cultural dialogue.

Thirdly, the lack of curiosity about the structure that governs our morality smacks of scientific surrender. Humanity didn’t come this far by shying away from complex problems. We progressed by striving to understand, and I suggest we continue to do just that.

Subscribe now

I wandered through their codexes and scattered fragments of scripture. Some of it was noise. Pebbles I couldn’t extract much from. Yet amid the gravel, there were glints of gold. Some incantations I’d heard before, just wearing different names. Others hit me sideways: small, sharp insights that I eagerly stored away, in the memeplex folder of my brain.

I’m sure my voyages in rationalist land aren’t over. For now, though, I’d like to offer a modest list of the core concepts I’ve found most useful in practice. But first, a few words on what they don’t tell you.

1. Prerequisites

Before setting out to explore the rationalists’ musings, it helps to have a passing familiarity with a few ideas, just enough to decode the jargon and see where their arguments lead.

a. Basic Probability Theory

A grasp of basic math and probability theory helps, especially Bayes theorem. Not because you’ll be doing any calculations (you won’t), but because they love talking in terms of priors and posteriors. Familiarity with the terminology keeps the writings intelligible when technical terminology appears.

b. Philosophical Vocabulary

Ditto for philosophy. Terms like epistemology, ontology, and instrumental rationality appear frequently. You don’t need a philosophy degree, but having a basic grasp of these concepts helps prevent confusion.

c. Constructing Argument Maps

I’ve mentioned before that it took me an embarrassingly long time to realize you can create an argument map for practically any argument you come across. This skill is especially useful when reading anything with a reasoned point—particularly from rationalists and their cousins, the analytic philosophers—who have a deep love for systematizing reasoning. Their structured style makes argument mapping especially straightforward. This can be jarring if you’re used to a more discursive approach, but it’s a highly efficient way to communicate. There’s little repetition, precise language, and no Straussian weirdness, resulting in far less room for misinterpretation, a deeper issue in written communication than may seem at first glance.

2. Main Concepts

But enough with these tedious preliminaries, here are concepts the rationalists helped popularize, that I’ve found most important up to now. I’ll try to arrange them in order of personal usefulness, from least to most.

13. Bayesian Updating

I remember doing a lot of exercises on Bayes’ Rule during my degree. I found it remarkable that, starting from Kolmogorov’s axioms, we could quickly arrive at a simple rule that already produced counterintuitive results.

Rationalists, too, are fascinated by Bayes’ Rule—but I suspect for a slightly different reason. Bayes shows how to update beliefs while preserving uncertainty, which can be a revelation if you’re used to thinking in dichotomous terms: true or false. As such Bayesian reasoning serves as a kind of heuristic reminder: don’t be too sure of anything, think probabilistically, and remember that it’s rational to slightly adjust your beliefs in light of new evidence, even if that evidence is weak.

I’ll admit that, probably due to my education, this way of thinking was already ingrained in me before I ever embarked on my rationalist voyages. So, while I also admire Bayes’ Rule, it hasn’t changed how I think or behave as much as some other ideas on this list have.

12. Occam's Razor

Like with many other ideas, I first learned about Occam’s Razor on the internet. The principle is simple: when faced with multiple explanations, favor the simplest one, the one that makes the fewest assumptions.

This makes sense because each additional assumption carries its own probability of being true, so unless it meaningfully improves the theory’s accuracy, adding more assumptions lowers the overall likelihood that a theory is correct.

Occam’s Razor can also be grounded in a more rigorous framework: algorithmic information theory. In this framework we can define something known as Solomonoff induction, that states that simpler explanations of the observations (those that can be described by shorter programs) should be assigned higher prior probabilities. In my naïveté, I think that this comes remarkably close to a formal resolution of Hume’s problem of induction, making Occam’s Razor a valuable principle for evaluating competing hypothesis.

11. Knowing your Cognitive Biases

I was already aware of cognitive biases before my adventures in Rationalistlandia, but I was fairly nonchalant about them. I saw them as quirky mental glitches that popped up in specific situations. What I didn’t grasp was just how deeply and pervasively they shape our understanding of the world.

Now, thanks mostly to decision theorists, I’ve come to see that many biases aren’t just bugs, they’re often adaptive shortcuts for reducing complexity and making quick judgments. They’re not just occasional distortions; they permeate our cognition precisely because many likely evolved to enable efficient decision-making.

Although I’m not sure how often I manage to transcend my own biases, I’ve found rationalist culture helpful in this regard, mainly because it treats bias-spotting as a community sport—often directed at others rather than oneself, but still, it keeps the problem in mind.

10. Schelling Points

A Schelling point is an option that people tend to choose by default in the absence of communication, because it seems natural, special, or obvious to all involved.

Imagine two people are told they’ll win a prize if they can pick the same number from the following list without any exchange of information: 5, 12, 9, 35682934324, 2. Many people would intuitively choose the large number. Why? Because it feels special or more ‘focal’, it’s a solution people gravitate towards in the absence of communication.

Schelling points have interesting game theoretic implications as they seem to suggest that cooperation is possible even without exchanging information. They are also sometimes interpreted more abstractly. For instance, some may argue that moral norms function as Schelling points, serving as default expectations for behavior in the absence of explicit rules or agreements. The knowledge of this concept could help in combating one of rationality’s greatest enemies, who will appear later on this list: Moloch.

9. The Proving Too Much Technique

Once you start mapping arguments, you’ll notice that most people rarely make blatantly invalid ones, where the conclusion doesn’t follow from the premises. More often, the structure is logically valid, and the real challenge lies in evaluating whether the premises themselves are true. This can be especially difficult when those premises rely on empirical claims, requiring inferential estimates regarding how likely they are to hold.

However, there’s a powerful shortcut that can sometimes dismantle an argument quickly, without need for additional effort. It’s by finding a reductio ad absurdum, which rationalist jedi master Scott Alexander sometimes refers to as “proving too much”. The idea is simple: if a line of reasoning leads to conclusions that are absurd, or wildly overreaching when applied to another setting, then something is likely wrong with the reasoning itself.

Consider the following argument:

We shouldn’t allow this new medical drug because we shouldn’t permit things that carry a risk of death, even a small one.

By the logic of the argument, we’d also have to ban childbirth, driving to the hospital, or even taking Tylenol, all of which carry small but nonzero risks of death. Unless we're prepared to endorse all those bans, the principle behind the argument collapses under its own weight: the argument needs to be amended because it proves too much.

8. Beware Isolated Demands for Rigor

Beware isolated demands for rigor is the title of another well-known post by Scott Alexander (who, by the way, runs a substack called Astral Codex Ten). The article has achieved near-normative status within rationalist communities, and deservedly so, in my view. The idea is to be mindful of how much rigor we demand from arguments or explanations we dislike, especially when we’re far more lenient with arguments that support our own views.

This was a trap I used to fall into often: if a study reached a conclusion I didn’t agree with, I’d dig into the methodology, trying to uncover potential flaws, biases, or hidden assumptions that could undermine the results. But when a study supported my preexisting views, I barely questioned it. I just accepted it as strong confirmation of my perennial correctness.

The reality is, you can always ask for more rigor, clearer definitions, more precise measurements, better statistical methods. No analysis is perfect. But if we apply that scrutiny only to arguments we dislike, we’re not being consistent. To keep our epistemology fair and not view the world through distorted lens, we have to hold all claims, even the ones we’re inclined to agree with, to the same standards.

7. The Inside vs Outside View

The inside view vs. outside view refers to a conceptual framework for thinking about prediction and decision-making under uncertainty, popularized—as far as I know—by over at his blog Overcoming Bias.

The inside view refers to analyzing a situation by focusing on its unique features: examining the specific details, considering chains of causation, and reasoning about how events will unfold. It’s the natural way most people plan projects. You break the project down into subcomponents, estimate how long each task will take, and then sum those times to produce an overall forecast based on how you expect things to go.

The outside view, by contrast, asks: what usually happens in situations like this? Instead of relying on internal reasoning, you find a reference class—a group of similar past cases—and base your prediction on how things typically turned out for them.

Hanson and other rationalists argue that the outside view is often more accurate. We tend to be overly optimistic when reasoning from the inside, partly because we’re emotionally invested in the success of our plans, and partly because (unless we have expertise in the task at hand) we overlook unknown unknowns—things we didn’t think to factor in because we didn’t know they existed.

At its core, the inside vs. outside view is a tool for combating the planning fallacy: the tendency to underestimate how long things will take, how much they’ll cost, or how likely they are to go wrong. The solution (inspired from probabilistic thinking) is to anchor your expectations to base rates drawn from similar cases in the real world.

6. Moloch

Moloch is a metaphor for systems where individual actions, driven by self-interest or survival, lead to outcomes that are bad for everyone, even when no one wants that result. Moloch is what happens when everyone does what's best for them, and it makes the world worse for all. This is an especially terrifying prospect for rationalists, because it suggests that even if everyone acts rationally, things can still go horribly wrong, casting Moloch as an arch-nemesis of rationality itself.

Examples of Moloch in action abound. Arms races are one of his favorite pastimes: if your neighbor starts stockpiling weapons because he’s afraid of you, and you can’t know his intentions, then it’s rational for you to start arming too, after all, maybe he wants to invade you. And just like that, you’re both worse off, trapped in a destructive loop neither of you wanted, while Moloch cackles from the shadows.

Moloch’s favorite feast, however, may be authoritarian regimes, where even the dictator’s closest allies secretly loathe the system. Yet under the suffocating grip of total surveillance, where every word and whisper is monitored, even the inner circle cannot coordinate resistance. As a result, isolated acts of defiance are swiftly and brutally extinguished, rendering rebellion not only perilous, but irrational. And so, despite near-universal hatred, the system endures, nourishing a gleeful Moloch all the while.

But don’t be fooled into thinking Moloch is always easy to spot, he is crafty and insidious. He slips into systems where we might not expect him. For instance, some say that the academic publishing system has fallen to Moloch: no one truly likes it, yet everyone must play along, because stepping outside the system is professional suicide. Others point to dating apps, claiming they too bear his mark. Some even argue that sex itself has been twisted by his invisible, putrid hand.

Moloch the incomprehensible prison! Moloch the crossbone soulless jailhouse and Congress of sorrows! Moloch whose buildings are judgment! Moloch the vast stone of war! Moloch the stunned governments!

5. The Map Is Not the Territory

This is another idea I first encountered in statistics, during an introductory lecture where I learned that all statisticians are legally obligated to repeat ten times a day the mantra: “All models are wrong, but some are useful”. In fact, no statistical model can fully capture the complexity of the reality it tries to describe; there will practically always be simplifications, assumptions, and edge cases where the model breaks down. But that doesn’t make modeling useless. On the contrary, models abstract away detail to help us understand and predict the world more effectively.

A map is also a model—albeit of a geographical area—and, like all models, it’s necessarily imperfect. A perfectly accurate map, one that includes every rock, blade of grass, and crack in the pavement, would be as large and unwieldy as the territory itself, and therefore useless. “The Map is Not the Territory” is the preferred metaphor in rationalist circles for reminding ourselves not to confuse a model with the reality it is meant to represent.

As an example of the importance of this concept take the 2008 financial crisis, which many believe was exacerbated by people mistaking their financial models for the real-world phenomena they were meant to represent. On paper, the lending practices seemed sustainable, but the numbers were built on oversimplified assumptions: the map was confused with the territory and the results were catastrophic.

4. The Categories Were Made for Man

The Categories Were Made For Man, Not Man For The Categories is another frequently cited post by rationalist grandmaster Scott Alexander, centered around a simple but powerful idea. We tend to treat categories—and the words that define them—as if they refer to fixed, objective entities out in the world. But in reality, we come up with words to categorize things because it’s efficient to do so. Words aren’t divine labels pointing to Platonic ideals; they’re functional sounds we use to gesture at fuzzy clusters of things we find similar or relevant.

Problems arise when we get too attached to the idea that a word must point to a precise phenomenon in reality or if we start thinking we can’t redefine words. In practice, many categories cover a range of overlapping cases and are often used in a context dependent manner. As we learn more or our priorities shift, the best way to group and label things may change. We can even invent entirely new words, and in fact, we do this all the time in science; the categories were made to serve us, not the other way around.

I think there’s a connection here to “The Map Is Not the Territory.” Categories can be seen as tiny descriptive models of reality. And just as mistaking a map for the actual landscape can lead us astray, mistaking our labels for immutable truths can blind us to better ways of slicing reality. Seen through this lens, “The Categories Were Made for Man” becomes a special case of “The Map Is Not the Territory”, and they should occupy the same position in this list.

3. The Principle of Charity

The principle of charity asks us to interpret other people’s arguments in their strongest, most reasonable form, especially when they are unclear or ambiguous. It’s closely related to the idea of steelmanning: taking an opposing position and actively improving it, making it even stronger than how it is being presented, and then engaging with that version.

Steelmanning arguments that go against your own position is often psychologically taxing, but I’ve found it to be extraordinarily helpful for understanding the world more accurately. If your goal is to find truth rather than simply “winning debates”, this approach is super useful. After all, dismantling a weak or poorly expressed argument doesn’t mean much. And, I apologize in advance, but now I’m going to have to hit you with the obligatory quote.

He who knows only his own side of the case, knows little of that.

— John Stuart Mill, On Liberty (1859)

2. Having a Coherent Moral Theory

I initially came to appreciate the value of having a coherent moral theory through internet debates. Rationalists don’t always discuss this aspect explicitly, but it seems implicitly embedded in their approach towards ethical conundrums.

We usually walk around with a jumble of moral norms bouncing around in our heads. Some come from religion, others from our friends and family, and other still from our favored political tribe—sometimes, they contradict one another. I suspect this lack of coherence caused me a good deal of cognitive dissonance when I was younger. Ideas would clash without resolution, and I had no clear framework to sort them out.

Actually taking the time to define what we mean by “good,” coherently and explicitly, can orient our lives. It helps resolve internal contradictions and gives us a stable direction, rather than leaving us stuck reacting to a tangle of conflicting intuitions. Plus, this kind of moral clarity can also lead to deeper realizations.

In politics, we often let parties tell us what’s right, largely because each movement comes with its own culture and embedded moral norms. But ideally, it should work the other way around: we should first clarify what is good—what we’re aiming for—and only then support the political structures that best align with that vision. Not the other way around. This was one of the central points I made in a post arguing against the idea of political allegiance.

1. Calibration

Calibration is the degree to which the confidence in our beliefs or predictions matches how often those beliefs or predictions turn out to be correct. It’s about how well our judgments align with reality: If we say we’re 90% confident about something, then ideally we should be right about 90% of the time.

Crucially, this isn’t just some abstract epistemic virtue, it’s something you can actually test (here’s a free one). I found the concept so useful I ended up writing a short article about it. It really is a great way to keep your confidence tethered to reality.

Subscribe now

Post Scriptum

I’m sure there must be some other great rationalist concepts out there I’ve missed. If one comes to mind, and you feel like sharing, be a good epistemic Samaritan and let us know in the comment section.

A few recurring themes run through his writing. One is his deep skepticism toward the academic field that studies misinformation, which he argues often misleads more than it illuminates. Another is his consistent reminder that naive realism, the belief that we perceive the world exactly as it is, free from distortion or interpretation, is simply false.

Notably, a recent article in Time magazine alleged that he is an example of someone being cultivated by authoritarian leaders because of his critiques of misinformation studies. This is wrong, and quite tragically so, since he’s not shy about condemning authoritarianism and Trumpism in basically every other post.

Personally, I don’t always agree with Williams, but I’ve learned a lot from him. His articles are clear, precise, and beautifully sourced. It’s a pleasure to read them.

Anyway, here are his top 10 11 substack articles according to me. But first, an honorable mention.

HM. Misinformation researchers are wrong: There can't be a science of misleading content

I’m not a fan of the claim that there can’t be a science of something, it seems to me that, eventually, people who claim such things will be proven wrong. But this was the first real essay Williams published on Substack, and it offers a great introduction to his critique of misinformation studies, so I wanted to include it as an honorable mention.

In it, he outlines a fundamental dilemma central to the study of misinformation.

- On the one hand, if researchers focus on clear-cut cases of misinformation, misinformation is rare and largely symptomatic of other problems, at least in Western democracies.

- On the other hand, if researchers focus on subtler ways in which communication can be misleading even when it is not demonstrably false, the concept of misinformation becomes so broad, amorphous, and value-laden that we should not delegate the task of identifying misinformation to a class of experts.

As Williams points out, outright false news is actually quite rare, so the dilemma he presents is real and foundational to his entire way of thinking about the subject.

11. Can democracy work?

Professor Williams wants us to appreciate just how far ahead of his time American writer and political commentator Walter Lippmann was, and oh boy, was he.

Back in 1922, Lippmann had already realized that humans carry with them simplified internal models of the world, that often diverge significantly from actual reality, especially in the political realm. This surprised me, because we talk a great deal about biases and world models in cognitive science and machine learning today, so hearing that kind of thinking articulated over a century ago felt extraordinarily prescient.

Because of this insight, Lippmann believed democracy was in trouble. In his view, only trained experts, those with a clearer grasp of reality, should be making society’s big decisions. Williams recounts his argument, and then offers a thoughtful critique of it.

10. People embrace beliefs that signal their traits and loyalties

Williams titles his pieces like someone who's been shouting into the void for years and just wants to get the message out, if only through the title.

In People embrace beliefs that signal their traits and loyalties, he tells us exactly that.

- People endorse happy beliefs that make them seem nice and approachable - for example, that good things happen to those who attract positive energy.

- People affirm supernatural beliefs that make them seem rule-following and cooperative - for example, that there’s an all-knowing God who punishes bad behaviour, or a crocodile-like spirit that attacks people for violating social norms.

- People endorse paranoid, conspiratorial worldviews to signal to others that they’re highly vigilant against social manipulation - for example, that they will blame apparent coincidences on others and jump to conspiratorial conclusions on little evidence.

I was already aware of the powerful insight that beliefs aren’t always about tracking truth, but can function as signals that bring social rewards. Still, some of the possibilities explored were new to me, like the idea that believing in conspiracy theories can serve as a signal of vigilance, a way of projecting epistemic distrust so others don’t mess with you. Fascinating stuff.

9. The world outside and the pictures in our heads

This article revisits a highly interesting historical debate between journalist Walter Lippmann and pragmatist philosopher John Dewey.

Lippmann argued for technocracy, believing that the complexities of reality are too great for the average person to accurately grasp, given their limited time and resources. So, it’s better to let experts make the big decisions.

Dewey, by contrast, was more optimistic about democracy. He believed that ordinary citizens could make enough sense of reality to participate meaningfully in collective decision-making.

Williams, evaluating their debate a century later, thinks Lippmann was too optimistic about the ability of experts to see reality clearly, while Dewey, was kind of vague and unsatisfying. Still, both thinkers were prescient in recognizing that the limitations of our epistemic faculties pose the most fundamental political challenge to society.

Neither Lippmann nor Dewey had particularly satisfying proposals about which institutions and practices are most consistent with such epistemic limitations. Nevertheless, they both saw—correctly—that they pose one of the most profound political challenges we confront in the modern world.

8. In politics, the truth is not self-evident. So why do we act as if it is?

Dan Williams isn’t sure but he proposes 6 possible culprits:

1. Ignorance

2. Evolutionary mismatch

3. No skin in the game

4. Social benefits

5. Politics demands action

6. The problem of unilateral reasonableness

I would add one more factor to the list: cognitive resources. Constantly doubting and calibrating our internal world model likely requires significant mental effort, which can slow down decision-making. In contrast, greater confidence in our world model should make decisions faster and more efficient.

Because of this tradeoff, I sometimes worry that stepping outside naive realism is a kind of luxury, one that only privileged people who have enough time to think about the world (like me and Professor Williams) can afford.

7. People are persuaded by rational arguments. Is that a good thing?

Have you ever thought that rational arguments could be bad? Neither did I, but Williams did, and he has good reasons.

Studies show that people are generally persuaded by rational arguments, even if just a little bit. So what’s stopping a bad hombre from spamming rational arguments with terrible conclusions, slowly moving people in the wrong direction? Are we at the mercy of the best rational argument producers?

My first instinct would be to say it’s harder to craft compelling arguments around falsehoods. But Williams takes a different path. He explores some fascinating academic theories suggesting that, through the social exchange of reasons, we tend—on average—to converge toward more accurate beliefs.

In this light, he concludes, the exchange of rational arguments appears to be a force for good.

6. The media very rarely makes things up

Contrary to what some people think, Williams tells us that the media rarely fabricate stories. With so much happening each day, it's far less risky to their reputation to select true stories that fit the narrative they want to promote than to make something up. Which, of course, makes the already difficult task of understanding the world clearly even harder.

Also, Williams can’t resist returning to one of his favorite critiques: that the people evaluating media bias often lack the very objective standards they accuse the media of missing.

When woke people write about the media, they argue it is biased in favour of white supremacist, patriarchal, heteronormative, etc., ideas. When Noam Chomsky writes about media, he argues that it is biased in favour of capitalist, militarist, and right-wing ideas. When Steven Pinker writes about media bias, he argues it is biased against the acknowledgement of progress. And when liberal social scientists evaluate the bias in media outlets run by liberal journalists, they develop methods showing that media bias does not exist.

5. The puzzle of populist devotion: How alliances and propaganda explain right-wing populism

Williams seeks to apply a novel framework called Alliance Theory (which explains how political belief systems are formed) to better understand the phenomenon of right-wing populism.

What explains the contents of political ideologies? Alliance Theory attempts to answer this question with only two core assumptions: (1) humans possess abilities for forming and detecting alliances, and (2) instinctively use propagandistic tactics to support their allies against their rivals in conflicts.

Applying this line of thought helps explain the spontaneous, emergent alliance between groups that don’t typically interact: less formally educated and socially conservative voters; business elites who resent academia; and those psychologically inclined toward conspiratorial, paranoid, and anti-establishment politics, regardless of who or what the “establishment” is.

But in the final section, Williams has a crisis of faith. As a liberal elite himself, a natural rival of the populist right, he begins to question whether his own analysis is hopelessly biased. Is he just engaging in the process of producing propaganda for his own tribe?

4. Misinformation is often the symptom, not the disease

Misinformation is not something that happens to the mass public but rather something its members are complicit in producing.

This piece is Williams in condensed form. In it he argues that political bias isn’t a virus to be cured, it’s part of the human condition.

Many of his core insights are here: people aren’t gullible; they’re just not disinterested truth-seekers. Beliefs carry social weight. We’re drawn to those that signal loyalty to our tribe and portray our opponents in a negative light. And this isn’t exactly irrational, there are real social rewards for doing so. That’s why we’re often motivated not to seek the truth, but to seek out information that justifies the beliefs we already want to hold.

The result is a marketplace of rationalisations that rewards the production and dissemination of content that supports favoured narratives in society. We tend to view the super-spreaders of misinformation as master manipulators, orchestrating mass delusion from their keyboards and podcast appearances, but they are often better understood as entrepreneurs who use their rhetorical skills to affirm and justify in-demand beliefs in exchange for social and financial rewards. Beyond Merchants of Doubt they are merchants of affirmation, and for the right price they'll validate and rationalise anything.

3. Fashionable ideas

While Williams sees value in the concept of luxury beliefs, he disagrees with the way it is frequently presented.

First, he challenges the common historical framing: the idea that, as expensive material goods became more accessible to the masses, elites turned to ideas to differentiate themselves. He argues this overlooks the fact that elites have always used norms, manners, and beliefs as markers of distinction, and that many luxury items still remain available (only) to the elites.

Second, he critiques the tendency to contrast elites with their erroneous luxury beliefs against a working class portrayed as holding more accurate political views. Because we have evidence that:

On average, more affluent, more educated people overwhelmingly have more accurate, better-informed beliefs—less conspiratorial, more evidence-based, more scientific, and so on—than poorer, less-educated people.

Third,

When it comes to typical luxury goods, the poor literally cannot afford them. This is why they honestly signal wealth. But what prevents poor, uneducated people from adopting the policy beliefs or preferences of the “elite”?

Indeed, as Williams points out, the working poor often support policies that go against their own interests, so what is stopping them from also parroting the, so-called, luxury beliefs?

Williams suggests that it may be more accurate to reinterpret the concept as fashionable beliefs, rooted in privileged access to high-status social networks.

However, once reframed this way, it seems like the analysis should be “divorced from any close analogy to the concept of luxury goods and the idea of ‘elite’ status markers.” This is because there are many different social networks where status games are played, each with its own fashions.

Highly-educated, affluent, progressive members of the professional class might be a kind of elite, and their beliefs are no doubt trendy among certain segments of society. But the Tucker Carlsons and Elon Musks of the world—whose beliefs are, in my view, much less evidence-based and reality-oriented— are also elite trendsetters in the marketplace of ideas, albeit among very different communities.

2. Contra critical theory

Williams tells us that

Science is a credit economy in which scientists compete to win prestige (i.e., get credit) for making novel discoveries. This reward system is science’s engine, motivating scientists to invest time, energy, and ingenuity into advancing the frontier of human knowledge.

As usual, he's being a bit cynical (I still hold on to the naive belief that scientists can be genuinely motivated by ethical concerns, aimed at improving the human condition). Still, he makes a valid point: different disciplines reward different things. And critical theory seems to value the development of novel, “sophisticated” critiques of Western society.

Williams continues by acknowledging that, although there is “a lot of truth and insight in much critical scholarship,” it's still problematic that we've created a one-sided marketplace for self critical content.

Humans are, after all, prone to motivated reasoning, we tend to ignore counterarguments and lower our standards of evidence to support conclusions we're incentivized to believe. And, within the incentive structure of critical theory, this can lead to critiques that are excessively negative, at times resembling high-IQ forms of conspiracy theorizing.

1. The "everyone is biased" bias

Dan Williams at his most fragile.

lately I’ve been wondering whether I’m wrong about everything.

This piece was written shortly after Trump took office, and the political moment seeps into every line. Musk is (deservedly, in my view) portrayed as sociopathic in his relationship to truth and honesty.

As Williams reflects on his life’s work, amid a crisis of confidence, he would like to develop and refine his thought. While it's true everyone is biased, as his writings have often reminded us, it’s also crucially true that some are more biased than others. Some distortions run deeper, and pretending otherwise can lead to its own kind of intellectual defeatism.

The implication being that intellectual humility has limits, at some point, even the most self-aware thinkers should be willing to take a stand.

Looking at how most people approach politics, the problem of intellectual arrogance seems much more widespread than misplaced humility. Nevertheless, a rare vice is still a vice—and one worth taking seriously.

Subscribe now

Post Scriptum

In his writings, Williams raises very good criticisms of the field of misinformation, and I’m sure the field suffers from all the usual academic degeneracies. Still, I’m skeptical of abandoning the project of trying to study misinformation and bias in a more rigorous, scientific way.

I’m inclined to think there are promising directions worth exploring, particularly in developing quantitative measures of partisan bias. For example, we could analyze ideological word clusters and word frequency, or apply algorithmic sentiment analysis to assess the tone and direction of articles and videos across different publications, thereby generating aggregate measures of partisanship. We might also track how often partisan think tanks are cited, or examine user behavior in comment sections to identify ideological clustering.

These kinds of measures won’t be perfect, but I dont think they will be useless either. They offer a potential first step toward reducing the influence of human bias in how we assess misinformation, something I agree with Williams is a major obstacle.

(After a quick chat with ChatGPT, I’ve gathered that many of these things are already happening.)

Other, potentially less noisy, indicators could include political donations by media ownership, editorial endorsements, and long-term patterns of coverage. Large language models also seem well positioned to assist in identifying bias and distortion at scale. I’m cautiously optimistic about the potential usefulness of these kinds of approaches.

In the bible there is a story in Genesis about Cain and Abel. And look, it’s not just a story, it’s a meta-story. It’s the story behind every story you will ever hear.

See the ‘I’, that thing you call you, it’s not just sitting there in your head like some bureaucrat pulling levers. It conceptualizes itself as a story. Or more precisely, it’s a nested hierarchy of stories within stories. The ego doesn’t emerge in a vacuum. It emerges in narrative space, and that’s not optional! You’re not deciding to be in a story—you’re already in one. The only question is whether it's a tragedy, a farce, or, God willing, something redemptive

Now, people say, ‘Well, the ego is just a product of brain activity.’ That’s like saying Hamlet is just ink on a page. So when *I* ask myself what do I really think of Peterson’s books what I am really asking is what does the story in a nested hierarchy of stories think of the stories that are in Peterson’s books.

The only way to properly answer this question is to understand what should be at the top of the hierarchy of stories. What we should be aiming for. See, this is not just a descriptive question, it’s a moral one. It’s about value, about the ideal that orients all subordinate narratives. The Greeks called it the Logos. Christians see it as the divine principle of reasoned, truthful speech, that separates order from chaos.

So in a narrative sense, that is truer than mere descriptive truth itself, the real answer to what I think of Peterson’s books is found in acting out the divine Logos of having read them, and behaving as if I’ve had intelligent thoughts about them.

This is what I really think of Jordan Peterson’s books.

Subscribe now

And these posts got me thinking about a rather embarrassing gap in my own mental landscape, something I somehow managed to miss until halfway through my PhD. It wasn’t an emotion, and it wasn’t quite a developmental milestone either. It was more like an insight—an idea that, once learned, makes certain parts of reality easier to grasp.

Lately, I’ve been increasingly aware that there are these things, let’s call them “useful memeplexes”, that function a bit like mental software updates. Install one, and suddenly the world makes a bit more sense or becomes simpler to navigate. This is all pretty obvious, I’m not talking about anything transcendent here. To give a concrete example, consider numbers.

We take the numbers memeplex for granted, but not every culture has it. The Pirahã people of Brazil, for example, apparently operate with only the concepts of “one,” “two,” and “many,” seemingly getting along just fine. (Though I wouldn’t want to be their accountant.)

Of course, lacking the numbers memeplex does come with some serious disadvantages. It makes information sharing... imprecise. If an enemy is attacking their scouts might report, “many are coming,” leaving everyone guessing whether that means 100 soldiers or 100,000. Not great for strategic planning.

In fact, the numbers memeplex is so overwhelmingly useful that multiple successful civilizations independently reinvented it. And today, we make sure it gets forcibly installed during childhood via the mandatory software update known as formal education.

But not all useful memeplexes are so widespread. Some slip through the cracks. Here’s the one I managed to miss:

1. The Universality of Argument Maps

I vaguely remember being taught about syllogisms in high school during a brief detour into Aristotle, but their significance was lost on me at the time. The big thing I didn’t grasp was that they weren’t just about showcasing logical implications regarding how mortal Socrates was, they were archetypal representations of every argument you’ll ever hear.

And once you realize this fact you should be able to take almost any argument (short of the illegible) and reconstruct it into this skeletal form—a structure known as an argument map.

Say your significant other comes home, sees the trash still sitting by the door, and says, “You forgot to take out the trash again,” before disappearing into the bedroom with the kind of silence that hums ominously.

That’s a (compressed) argument. And if you unpack it, the logic looks something like this:

Premises:

1. You didn’t take the trash out.

2. It was your duty to take out the trash.

3. Not doing your duty is bad.

Therefore: You did something bad.

At this point, your options are pretty limited, you can either object to one of the premises or apologize.

Argument maps are tools to help us engage in cognitive reflection. They allow us to reflect more deeply (engage in System 2 thinking for dual-process theory enthusiasts) by laying out the structure of an argument clearly and explicitly.

When I first discovered this, I began trying to form argument maps in my head during disagreements. At first, it was awkward and clunky but over time, it became second nature, almost like adding numbers. It’s been incredibly useful for keeping my objections focused and avoiding common logical fallacies. Just as importantly, it’s helped me recognize when I don’t actually have a real objection, when I’m simply uncomfortable with the conclusion and might just be embarrassingly wrong about something.

I see reconstructing arguments into maps as analogous to formalizing a problem using mathematics: it takes more cognitive effort upfront, but in return, you get clarity, precision, and a better chance of actually being right.

2. What Are the Rationality Memeplexes?

I often wonder what other useful memeplexes I might be missing, and I can’t help but wish there were a giant, well-curated list somewhere, ideally sorted by lifetime utility.

This feels like a part of the thing the rationalist community might be trying to do. Rationalists have spent a fair amount of time trying to define what rationality actually is, and while I think I understand what they’re gesturing toward, I’ve never found their descriptions entirely satisfying at an applied level. How in practice does one become more rational?

It seems to me that at least part of rationality is about acquiring a set of useful memeplexes, mastering them, and learning when and where to deploy them. I feel like this, in a way, should be a similar process to becoming a good statistician or a good plumber. The hope is that it would offer a more pragmatic picture of what rationality actually looks like in practice.

Perhaps in a future post, I’ll try to sketch out a list of useful memeplexes, just to see how far (if anywhere at all) this approach can take us.

Subscribe now

I had to open this image in a photo editor, pull out the color picker, click on the two faces and double-check—yep, they really are the same color. And yet my brain still refuses to believe it.

Why does this happen?

One of the leading explanations is that our visual system has to make sense of an overwhelming stream of high-dimensional data. And to allow us to navigate the world efficiently it has to compresses this information, filtering out what seems irrelevant and amplifying patterns that matter for survival and evolutionary success.

So, instead of giving us a raw data dump—the color value of every single pixel in a 258,100-element grid—our brain processes the input using heuristics, focusing on what it interprets as patterns: light sources, shadows, surrounding objects, expected textures, and materials. From this, it constructs our perception.

This is why the same wavelength of light (i.e., the same color) can look different depending on its context. The brain seems to be effectively "guessing", using prior experience and environmental cues to speed things up for us.

1. A Far Reaching Explanation

This observation—that it’s advantageous to simplify complex, high-dimensional for the sake of efficiency—turns out to be incredibly far-reaching. It starts as a neat little idea and grows like an Old God inevitably claiming another galaxy, until it ends as a totalizing paradigm that describes the whole realm of perception.

I’m now going to list a wide range of domains where this same pattern shows up, building gradually toward a crescendo of near-absurdity. After the first few examples, dear reader, you are officially granted permission to skim guilt-free. After all, efficiency is key for your fitness.

a. Optics

We've already seen one example in the visual domain, but to drive the point home, here are two more.

In this image, Einstein’s face is actually concave—the nose is furthest from the camera. But most viewers perceive it as convex, with the nose appearing to protrude toward them. The brain sees something that resembles a face and serves up the standard, familiar perception, without bothering to check the fine print.

The image is in greyscale, with a few colored lines sketched over objects and faces—but our brain, relying on perceptual shortcuts, fills in the rest and convinces us it's in full color.

b. Acoustics

Ever heard of the McGurk effect? It’s what happens when our cerebrum fuses visual and auditory information, leading us to hear different sounds—even though the actual audio never changes.

The brain, which has a need for speed, knows that when people move their mouths in a certain way, a specific sound usually follows, so it compresses visual cues with auditory input to quickly give us an approximation of reality that should be right, more often than not.

c. Gustation

Multiple studies have shown this kind of thing happens with taste, too. Take, for example, this one where researchers added red or green food dye to perfectly normal orange juice—no change in flavor, just a visual tweak. Somehow, people started reporting that the greenish juice tasted more sour than the original, and the reddish one didn’t.

Once again, the brain wants to cut corners and simplifies raw input using some kind of heuristic to generate our perception (I’m going to stop saying this part out loud from now on).

d. Olfaction

Olfactory sensations don’t seem to be immune either. Studies find that people’s descriptions of a wine’s “nose” are dramatically influenced by its color—even when the wine itself hasn’t changed.

e. Haptics

Haptics is the study of touch. I didn’t know that either until my brain decided I needed to write this article. Anyway, there’s an experiment in haptics where researchers place a fake arm in front of you, stroke it with a brush, and somehow, you feel it in your real arm, even though it’s not being touched at all.

This appears like another case of faulty misapprehension due to fast input compression, just like in the other examples we've seen. But the brain, being the efficiency addict it is, demands even more heuristics.

f. Epistemology

Constantly re-evaluating beliefs and processing new information is cognitively demanding. To conserve mental energy, the brain filters for information that aligns with existing beliefs, reducing mental load and enabling quicker decisions. This mechanism is thought to underlie confirmation bias (an expression of cognitive ease) where familiar ideas are processed more fluently and feel more intuitively true. It might also explain why, when we’re not putting in cognitive effort, we like to fall for fake news that flatters our preconceptions.

Similarly, research suggests that people are drawn to conspiracy theories because they offer simplified narratives that make a chaotic world feel more understandable. In both cases, the brain favors models that minimize complexity, even at the cost of accuracy.

g. Sociology

Some researchers think stereotypes are automatically activated when we encounter someone from a stereotyped group. It’s not that low-prejudice individuals don’t have stereotypical intuitions—they’ve just learned to inhibit them through conscious control.

From this perspective, stereotypes appear to resemble a form of dimensionality reduction: the brain condenses a complex individual into a simplified set of expected traits, enabling rapid judgments in the moment.

As is well known, in social settings, this can backfire. Generalizations can oversimplify human complexity in harmful or unfair ways, leading to prejudice and discrimination. This is the flip side of generalized heuristics: the same mental shortcuts that make us think quicker can backfire — just like in all the other cases we've been discussing.

h. Mathematics

It’s well known that our mathematical intuitions can be misleading. Take this classic question (courtesy of Shane Frederick):

A bat and a ball together cost $1.10. The bat costs $1.00 more than the ball. How much does the ball cost?

Many people instinctively answer 10 cents, but that’s wrong. The correct answer is 5 cents (solve the system given by x+y=1.10 and x=y+1, to confirm it yourself).

Or take this other example:

If 5 machines take 5 hours to make 5 widgets, how long would it take 100 machines to make 100 widgets?

Many answer 100 hours—but that’s incorrect. The correct answer is 5 hours.

The rate of production stays constant: each machine makes 1 widget in 5 hours. So if you have 100 machines working simultaneously, each producing 1 widget in 5 hours, you end up with 100 widgets in the same 5 hours.

Why do our intuitions sometimes get these answers wrong? It seems we’re not really solving the problem—we’re pattern matching. We're applying familiar mental templates that work in most familiar situations but fail here.

For example, in the bat-and-ball problem, many people latch onto the structure: 'total cost + one item’s cost means I need to solve by subtraction.' That’s a good heuristic for everyday life, since we are usually given problems in that form, but it doesn't fit the bat-and-ball setup.

Likewise, when all the numbers in a problem match—like “5 machines, 5 hours, 5 widgets”—our brains often assume the answer is probably just the repeated number again. It’s the kind of shortcut that works fine when someone says, 'If 5 people need 5 plates to eat, how many plates are needed for 10 people?' But it can also lead us straight to Wrongistan if we aren’t careful.

i. Inference and Probability

It's well known that we greatly simplify probabilistic thinking. For example, we often judge how likely something is based on how easily examples come to mind (the availability heuristic).

And when we're told we've tested positive for a rare disease, using a test with a 99% chance of correctly identifying someone with the disease, we often ignore the rarity of the disease itself and mistakenly believe there's a 99% chance we have the disease.

Instead of doing:

we do

Our brain just drops the other terms in Bayes’ Rule like they’re optional. Talk about a cognitive shortcut! But we’re still not fast enough.

j. Physics

Research shows that we naturally develop intuitive theories of motion that are computationally simpler—things like 'push = motion'—because they’re easier for our brains to grasp.

These mental models end up resembling the medieval theory of impetus, which, as it’s naming might give away, is not how motion actually works. So when we are asked certain basic physics question we get them quite wrong, a bit like the mathematical questions from before.

k. Rhetoric and Persuasion

Much of rhetoric and charisma hinges on punchy slogans, wit, and speech patterns that get the message across fast. Great orators don't waste time with “hums” and “ums”—they cut straight to the point.

Some of our best models of persuasion, like the Elaboration Likelihood Model, show that we can convince others either through the cognitively demanding central route of logical argument—or the far more efficient peripheral route, which relies on cues like tone, authority, or emotional resonance. In other words: the brain seems to have a route dedicated to being convinced by whoever efficiently offers a simple model of reality, regardless of whether they’re actually right.

l. AI

It’s well known that people who work in AI—or statistics, like me—are often trying to make models that reduce complexity just enough to act meaningfully in a brutally complex world.

But, perhaps more surprisingly, after being optimized neural networks naturally learn compressed internal representations that reflect useful aspects of the environment. Lower layers might detect edges or textures, mid-level layers group those into shapes, and higher layers might abstract those shapes into objects or even concepts. Even for our silicon brothers and sisters, it’s dimensionality reduction all the way up: from raw pixels to “cat.”

m. Logic and Argumentation

Have you ever stopped to wonder why our brains love committing logical fallacies?Turns out, one can view a great deal of logical fallacies as inductive judgments about arguments.

Take the ad hominem: if someone’s untrustworthy, it’s a decent heuristic to discount what they say. Appeal to authority? If experts agree it probably has a higher-than-random chance of being true. The post hoc ergo propter hoc fallacy is not deductively sound but it makes inductive sense: If you eat a weird berry and get sick right after, your brain says, “not eating that again.”

The brain figured out that ditching deduction for induction is a massive time-saver and because it’s addicted to efficiency, we end up with logical fallacies.

n. Morality

What? surely not morality too.

Yes, my cherished reader—morality might be one more domain where the brain applies its favorite fiendish trick.

2. The Master Problem of Politics

It’s now time to take stock and try (the apparently near impossible feat of) reflecting more deeply to see things more clearly.

Our brain is a freak on crack for compression and speed. This idea recurs in many different forms. During millennia of evolution it has undoubtedly discovered that making ten thousand almost-right decisions a day is superior to making two flawless ones.

Now drop that brain into politics—one of the most complex systems we can study—and let us ask ourselves: do we really think cognitive shortcuts are going to vanish?

If anything, we would expect them to ramp up.

Is this what we see? Seems so. Dimensionality reduction appears to be working overtime in the chaotic terrain of politics, and, these days, it feels most at home in the realms of political extremes and populism.

On this Substack, we’ve already discussed the case of radical and extreme beliefs at length. While, the connection to populism has recently been explored in a flurry of new studies, which suggest that lower levels of cognitive reflection are associated with stronger populist tendencies (see also this related work).

(Interestingly these studies came out after I made a YouTube video on the topic—proving that scientists from all over the world were clearly watching my channel, and that deferring to my cognition is an efficient way to conserve your own mental resources.)

In general, I don’t think this connection to populism was that hard to hypothesize. Many aspects of populism scream cognitive shortcut: the charismatic leader bypassing rational argument and appealing directly to the electorate through the peripheral route of persuasion; the black and white narrative of good, ordinary people versus a shadowy, evil elite; the conspiratorial inclinations; the disdain for checks and balances.

For example, I suspect it’s no coincidence that North Korea, the Soviet Union, Nazi Germany, Fascist Italy, Maoist China and Trump’s America all share a distrust of free trade and a fondness for economic nationalism. It seems like they’re leaders operated off similar intuitive models of the global economy as a giant zero-sum game.

Totalitarianism, in my view, is this logic taken to its endpoint, the ultimate act of dimensionality reduction. All complexity flattened into a single narrative, all dissent erased, all thought aligned. One great leader to control the most efficient (but also the highest variance) form of government: dictatorship.

If this analysis holds, it could also explain why populists appear to perform poorly once they're actually in charge.

Indeed, up to now, the argument has been largely non-normative but it's not a stretch to suggest that relying on fast-and-loose heuristics in making policy decisions is uniquely ill-advised. Cognitive shortcuts are great for everyday life, where speed often matters more than precision, but they were never built to handle the dense, multi-variable, game-theoretical complexity of governance and economics.

This doesn’t mean non-populists will always be right and populists will always be wrong. But that, on average, non-populists will be more likely to consult experts, weigh trade-offs, and throw more cognition at problems. While populists leaders, surrounded by people too afraid to voice strong criticism, will be more likely to trust their gut and just wing it.

Thus, non-populists usually fare better, as their decisions tend to bear a little more resemblance to those made by a prediction market. In this sense, the typical “strongman” charisma is part of the populist problem. It risks turning every decision into a loyalty test, where deference matters more than insight.

3. How to Solve the Problem?

The answer is both simple and hard.

In all the examples from Section 1, we can slow down and try harder to get closer to the truth.

For the McGurk effect, we can close our eyes and replay the audio to isolate the sound.
For probabilistic reasoning, we can pause and carefully apply Bayes’ Rule.
For visual illusions, we can use tools like Photoshop and a color picker to check what’s really there.
For the math problem, we can actually solve the system of equations instead of relying on gut intuition, and so on and so forth.

The lesson is that effort—combined with deliberate cognitive reflection—can help us see reality more clearly. So on a personal level the answer to our problem seems fairly simple: think more.

On the other hand how do we raise global cognitive reflection scores? Perhaps we can’t, and we just all have to play our part; calling out the worst oversimplifications, injecting reason into conversations, and practicing patience and charity when engaging with others.

In future posts, I’ll go into some more concret methods. But for now, that’s what I’ll continue trying to do.

Subscribe now

I.

Imagine you live in a small town, and one afternoon on your usual walk, a man named Ross Rose corners you. He seems visibly upset and launches into a passionate rant about how the gardening firm hired by the local council has completely botched the rose beds in the public park.

He pulls out his phone and shows you photos. As he talks, he mentions things like pruning angles, bud unions, and the importance of timing during the dormant season. It seems quite clear to you that Ross Rose knows a great deal of things about roses.

Even if some of the technical details goes over your head, the photos speak for themselves. You're no horticulturist, but you’re pretty sure some of them won’t make it. Ross even tells you that many local beekeepers (the town has a thriving beekeeper economy) rely on those roses as one of the few remaining nectar sources. Without them, their livelihoods could be at risk.

So you tell Ross you see his point and agree. You promise to mention it to a friend of yours who works with the local authorities, then continue on with your day.

Later, you do bring it up to your friend—and learn that Ross had already spoken to him. The council has agreed to fund new rose plantings and is offering short-term financial assistance to local beekeepers affected. With a bit of time, they say, things should be more or less back to normal.

II.

A week later, there’s a town meeting to address issues facing the community. One of the most urgent is the recent uncovering of a fentanyl trafficking ring that had been targeting local youth. They say around a dozen underage kids are now struggling with fentanyl addiction. The atmosphere is somber.

Partway through the meeting, Ross Rose stands up and begins talking about another problem the town is facing: the roses in the public park. He recounts the pruning incident and criticizes the council’s response as insufficient. Only 80% of the original roses, he says, are expected to recover within a year.

You don’t doubt Ross’s sincerity and you suspect he’s right. But as he speaks, you can’t help but question the timing. With the community facing something as serious as fentanyl trafficking you’re not sure this was the right moment for the roses. Maybe it would’ve been better to talk about them at the end of the meeting, or in some other moment.

III.

The next day, news breaks: a giant meteor is hurtling toward Earth, and world governments aren’t sure they can stop it. Scientists estimate there’s a 60% chance of impact sometime next month. An emergency town meeting is called.

People scramble to discuss contingency plans. Some talk about a government evacuation program involving spacecrafts. Others speculate about surviving underground. It’s all very frantic and chaotic.

Then, amid the chaos, Ross Rose stands up. He reminds everyone that the meteor might not even hit Earth and then starts talking about the roses in the public park, he insists that more needs to be done to allow the roses to growth as strong as they were before.

And that’s when it hits you: Ross Rose is crazy.

2. The Well-Ordering of Societal Problems

The story of Ross Rose illustrates what I believe is a true pragmatic principle: in politics, it’s not enough to be right about an issue. You also have to place that issue within a hierarchy of importance, relative to everything else society is facing, to decide how much attention it actually warrants. After all, public attention is a finite resource, so we should try to allocate it in a way that is proportion to the relative importance of each issue.1

Ross’s mistake isn’t that he’s wrong—he’s probably right. His failure is that he hasn’t well-ordered societal problems, so when he repeatedly tries to draw attention to the mispruning of the roses, he ends up looking like a loon.

Of course, as with most principles, no one can fully live up to the ideal. First, because measuring the importance of an issue is highly non-trivial; it’s not always as straightforward as it was in Ross’s case. And second, because speaking about something in proportion to its importance is not an easy feat for humans to achieve.

Still, if a principle is true, we should try our best to live by it.

3. Ideologies Suffering from Success

Political movements and ideologies are often held together by shared factual claims or core objectives: “The workers are oppressed,” “Our nation is the greatest,” “We need to dismantle the deep state.”

An interesting phenomenon can emerge when a movement that has correctly identified a societal problem actually succeeds in substantially addressing it. Paradoxically, that very success can make its members at risk of looking a bit like Ross Rose.

Why? Because as the movement achieves its goals, the original issue tends to shrink in urgency. Its relative importance fades, pushed lower in the evolving hierarchy of societal problems. Yet the movement’s rhetorical energy and emotional investment can remain frozen in time, misaligned with the new reality.

I suspect this is part of the reason why, in democracies, political parties from different orientations tend to alternate in power—success erodes urgency, and eventually, the public shifts its attention elsewhere.

This dynamic can affect any movement, from anti-communist crusades to social justice activism. One reason it happens may be that movements aren’t just tools for solving problems, they’re also social ecosystems. They bring people together, offering identity, purpose, and a sense of belonging. And when a movement starts to succeed, no one wants to see their tribe shrink or drift into irrelevance—especially after finally making a tangible impact on the world.

This is the strange irony of successful political movements: their very effectiveness can destroy their relevancy. But that’s not a tragedy. On the contrary, it’s usually a sign that progress has been made.

Subscribe now

Most of the time, I end up wanting to bring up parts of my background, because my life has been pretty weird and certain things seem like they'd need some explaining.

I've settled on giving a brief overview of some of the most important bits here, so that I can just refer back to this article whenever I need to.

1. My Pre-Teens

I was born in the 90s and spent most of my early life happily in a lovely European city of over 100,000 people.

Unfortunately, my dad died of a heart attack when I was nine, and my mom was diagnosed with cancer soon after. She passed away when I was twelve.

I found this gesture on their part to be quite inconsiderate in my regards and I didn’t appreciate it much.

My dad was big and tall, a veterinarian, and (from what I can remember) a center-left moderate. He was also a heavy smoker who loved reading and programming.

I felt very loved by him and loved him very much.

My mom had a slightly more complicated résumé: she had been a hippie and was a very progressive feminist. She was also an evangelical Christian, a New Testament type who liked giving money to people and talking about compassion and Jesus (we had immigrants ringing our doorbell every other week).

I felt very loved by her and loved her very much.

When you lose your parents young the state assigns you a legal guardian. In my case, it was my aunt (my father’s sister) someone I had only met a handful of times before.

Later I discovered there had been big arguments about it. My mother had left a written note expressing her wish that I be raised by some distant cousins I knew a bit better. But the note had no legal weight.

2. My Teens

My new home was in a small town of less than 6000 souls.

My aunt had "broken up" with her husband, though they were still living together and still technically married. They had two sons, both quite a bit older than me.

In my new family nobody believed in God (my naive habit of praying before meals was met with mockery), and dinner table conversation didn’t focus on sharing how our days went, let alone discussions around the systemic alienation of late-capitalist hegemony as refracted through the rhizomic intersectionality of marginalized identities. You mostly ate a big plate of shut-the-fuck-up and watched the news.

You were allowed to talk about the family business, though. Of which my uncle, the (self declared) head of the household, was the owner. To be fair, everything did seem to gravitate around him, since the other three members of my new family all worked for him. Politically, he was very much on the right, but he also had a generalized distrust of politicians, bordering at times on outright hatred.

On a practical level this different family culture meant that many of what were previously virtues had now become vices, and many of my vices had become virtues.

In my first home, I had been encouraged not to care about how I dressed, to spend long hours with books, to avoid saying no, and to be kind and empathetic to everyone.

Now, I was told to preoccupy myself greatly with appearances and presentation, to always stay in order, and to be strong, even forceful, when I wanted something in life.

It took me a while to understand this, and longer still for my brain to learn how to cope with it. Though, to be honest, I’m not sure it ever did.

I didn’t like my new existence much. At first, I tried to escape into books when I could — one of my oldest habits — but eventually I discovered the world offered stronger coping mechanisms. In my early twenties, I found a powerful one: mathematics.

But that’s a story for another time.

3. Dysthymia

When I was around 22 or 23, I started to suspect there might be something wrong with me. Outside the house, I enjoyed decent success with the opposite sex and had plenty of friends. One might, in a moment of entomological poetic license, have described me as a resplendent social butterfly. And yet, I had a hard time wanting to get out of bed in the morning.

For a long time, I assumed this was normal. After all, the great philosopher Schopenhauer had declared that life was a pendulum swinging between boredom and suffering. But, at a certain point, I began to suspect that this 'insight' into the human condition might not have been telling the whole story

So I went to see a psychologist, in the hope that a trained professional might give me some additional information regarding the nature of my predicament.

I wasn’t happy about doing this, there was still some social stigma around these things at the time, and It felt like admitting I was some kind of loon. The psychologist asked me a series of kind questions, I cried a little, and at the end of the session I asked if everything was okay. He said probably not.

I pressed him for a diagnosis, and he told me it was too early to say — but if he had to guess, I might be cyclothymic (the little brother of bipolar disorder).

I never went back to him and didn’t tell anyone about it.

About six months later, I repeated the procedure with a different psychologist. She refused to offer a diagnosis but encouraged me to keep seeing her. I didn’t go back to her either.

I wasn’t entirely sold on this sort of thing becoming a hobby of mine, so there was no third attempt — at least not for a long time. I remember wondering if I might have some kind of split personality disorder. A small part of me thought that was actually kind of cool; I was like Two-Face from Batman! Looking back, I think I’d taught myself to laugh in the face of suffering as a kind of survival tactic, and somewhere along the way, I’d started romanticizing my condition.

Finally, at 27-28, during my PhD, I started seeing a psychiatrist on a regular basis for about six months. Very expensive. After a while, he told me he thought I had dysthymia.

He prescribed me some Xanax to take as needed, and an SSRI whose name I can’t recall but, to be honest, I was too anxious to take them properly. I’d read one too many stories about side effects on the internet. So I took the SSRI every other day for about a month, then quit entirely because I was too scared to keep going. Sorry doc.

4. Dysthymia Free

Now I’m in my mid-thirties, and I’ve been free of dysthymia — or whatever the hell it was — for three and a half years! Life feels completely different from anything I’d ever known. I never, ever imagined it could feel this serene, this pleasant. Hurrah!

In my attempts to get better, I tried a lot of things. I changed my diet, fixed my sleep patterns, got more natural light, and made major adjustments to both my social and professional life. I even reevaluated my moral perspective on the world.

I often wonder which of these changes—if any—actually made the difference. I have my suspicions. But I can’t know, given the causal ambiguity introduced by confounding variables and the lack of experimental control.

I sure hope everything keeps going well though! (Please, dear God)

And that’s it for now, I guess. I think there may be some insights lurking in how different parts of my life unfolded, so I might talk about them in more detail in the future.

Subscribe now

In 2013, Scott wrote a lot. Towards the end of the year, he got accepted into medical school and things cooled off a bit. His writing started off much stronger than I expected — meaning, I was bracing for a bunch of naive, half-baked ideas from his first year of blogging — but instead I found a lot of very engaging articles some of which were surprisingly relevant to our times, all of it written with a refreshing casual clarity.

His critique of neo-reaction especially caught me off guard — a part of it was basically a much better version of a brief piece I had written, to the point where I felt a little embarrassed and started worrying that people would think I’d just copied him.

But enough with my whining, here are the final rankings of the ten eleven twelve thirteen articles I liked the most:

13. Reactionary Philosophy In An Enormous, Planet-Sized Nutshell

I like good summaries. Top 10 lists are a sort of summary. Reactionary Philosophy In An Enormous Planet-Sized Nutshell is a great summary of what neo-reaction is all about. Reading it helped me understand the reactionary perspective better, which is definitely a plus.

But it’s not only a summary it’s also a steelman, Scott was building up neo-reaction before criticizing it, hard.

12. I Myself Am a Scientismist

Scott seems to set out to defend ‘Scientism’ the “purported fallacy in which people naively believe that science can solve everything.”

Although he kind of cheats because at the end he says:

By holding scientific theories, which can be and are disproven, they trained themselves in Doubt. And that Doubt continues to serve them when they branch into other areas where theories cannot be disproven so easily. And maybe they will be less easily swayed by attractive verbal arguments.

And that is why I consider myself a scientismist. I know it is supposed to be a perjorative, but I am reclaiming it. And I know it has many definitions, but this one is mine:

A view of hypothesis-space that accounts for human fallibilities, as revealed by past experiences.

And a very, very high burden of proof before zeroing in on any one area of that space.

Which seems a little different from actual Scientism — but I agree with it, and I like it, so I’m not complaining too hard.

11. Scientific Freud

In this article, Scott points out that cognitive behavioral therapy and psychoanalysis have been found to have roughly the same level of evidentiary support across several large meta-analyses. This surprised me greatly. So Scott wonders if perhaps “psychotherapies work by having a charismatic, caring person listen to your problems and then do ritualistic psychotherapy-sounding things to you, but not by any of the exercises or theories of the specific therapy itself.”

An interesting discussion on how to evaluate different kinds of evidence follows.

10. Thank You for Doing Something Ambiguously Between Smoking and Not Smoking

In this article, Scott explores the concept of fungibility — the idea that certain things can substitute for one another. As an example, he uses e-cigarettes to illustrate how fungibility plays out in practice.

So let’s accept that using e-cigarettes will get you addicted and set you back a lot of money and otherwise be annoying but probably not deadlier than anything else you do on a daily basis. What then?

Well, in that case, it’s worse than not smoking but much much better than smoking. And whether or not their existence is a good thing depends on what they funge against. Do they funge against smoking tobacco or not smoking at all?

The main point is that people sometimes get too caught up in how bad something is (like the e-cigarettes), without realizing that it "funges" against something even worse — and that this tradeoff should be taken into account when making decisions. True.

E-cigarettes are literally the exact same thing as something that’s given out to anyone who asks in convenience stores, except without the cancer. To suddenly hold them to an extremely high standard of safety seems like a fallacy of fungibility.

9. All Debates Are Bravery Debates

This article is titled the way it is because Scott had previously criticized what he called "bravery debates" — arguments about which group is more persecuted, oppressed, or holds a minority opinion (and is therefore braver for expressing it). He found these kinds of debates toxic and unproductive.

But now he’s reconsidering, suggesting that maybe many debates have to involve this element — because it’s kind of necessary to know which group is suffering more, since it’s hard to micro-target advice or policy to the level of the individual.

It’s really hard to target advice at exactly the people who need it. You can’t go around giving everyone surveys to see how selfish they are, and give half of them Atlas Shrugged and half of them the collected works of Peter Singer. […] To a first approximation, all you can do is saturate society with pro-selfishness or anti-selfishness messages, and realize you’ll be hurting a select few people while helping the majority.

Thus, if you can only saturate society with one kind of advice, it makes sense to figure out which of two “opposing groups” is more oppressed, to target their issues.

I’m not sold on the idea that we can only broadcast one kind of advice throughout society, but the part of the article I found most interesting was the observation that different people need different advice depending on what they’ve over-indexed on growing up.

One person might grow up learning to be too selfish, another to be too self-sacrificing, and so they end up needing very different kinds of counsel. When they eventually find their favorite thinkers, they might land on opposite sides of some cultural debate — not because they disagree at a foundational level, but because they’ve gravitated toward opposite solutions to their problems, without realizing both sets of advice are valid.

It’s much easier to be charitable in political debates when you view the two participants as coming from two different cultures that err on opposite sides, each trying to propose advice that would help their own culture, each being tragically unaware that the other culture exists.

8. Valleys-Have-Two-Sides

I really appreciated this short article; first because it’s short (I like my signal-to-noise ratio high), and second because it says something I think is true.

Scott points out that political ideologies have a bad habit of comparing "our gritty reality to their beautiful thought experiment." They promise utopias without accounting for the friction of the real world.

He then tells us that neoreactionaries fall into exactly this trap: they imagine a benevolent monarch ruling wisely over society, just like the communists imagined their worker’s paradise. But instead of comparing their utopia to other utopias, they compare it directly to our messy, imperfect reality — and of course it comes out looking better than it should.

The opposite of perfect unimpeded progress up the right side of the valley is perfect unimpeded progress up the left side of the valley. The opposite of attempted progress up the left side dragged down by lingering rightists is attempted progress up the right side dragged down by lingering leftists.

7. Nature Is Not a Slate It’s a Series of Levers

In this article, Scott says, “being raised by criminals has no effect on anyone’s personality.” I was shocked.

I couldn’t imagine how growing up around criminals wouldn’t affect someone's personality. But then I realized there was a catch: I wasn’t thinking about the right definition of "personality."

Parenting can massively influence specific outcomes like academic achievement, religious beliefs, criminal behavior, and mental health. But Scott — being a psychiatrist in training — was thinking about the technically correct definition of personality, involving traits like introversion vs. extroversion, or agreeableness vs. disagreeableness.

Research suggests that when it comes to these traits, parenting matters a lot less than almost everyone thinks, if it matters at all. Which was still very surprising to me but less so.

Anyway, the main point of the article is that using our knowledge of human nature to design good policies is a smart strategy — and that people who ignore human nature usually have a much harder time making their ideas work.

But my point is that if I’m progressive – a label I am not entirely comfortable with but which people keep trying to pin on me – this is my progressivism. The idea of using knowledge of human nature to create a structure with a few clever little lever taps that encourage people to perform in effective and prosocial ways. It’s a lot less ambitious than “LET’S TOTALLY REMAKE EVERY ASPECT OF SOCIETY AS A UTOPIA”, but it’s a lot more practical.

6. We Wrestle Not with Flesh and Blood but Against Powers and Principalities

This article starts by wondering what would happen if we granted the neoreactionary premise that the past really was better, and concludes that it probably wouldn’t matter anyway, because we can’t turn back the clock.

For Scott, "the biggest changes in history have been predetermined reactions to different technological conditions", not the results of political movements. The political movements are just the puppets of technological progress, with “Vast Formless Things” like the Industrial Revolution pulling the strings behind the scenes.

But trying to reverse massive shifts like the Industrial Revolution is both basically impossible and extremely risky. If that’s what the neoreactionaries are aiming for, they’re tilting at windmills.

So this is my first beef with Reactionaries. They see someone identifying as Progressive saying something – Gloria Steinem pushing for women’s rights or something – and they say “Oh no, that awful Progressive Gloria Steinem is screwing up our traditional gender roles. If only she would be quiet, everything would go back to normal!”

Gloria Steinem is a puppet. If she’s part of some movement, even a large saecular movement calling itself Progressivism, they, too, are puppets. It is stupid to get upset at puppets. If you rip them up, the puppeteer will get new ones.

If you don’t like women’s lib, your enemy isn’t Gloria Steinem. Your enemy is the Vast Formless Thing controlling Gloria Steinem. In this case, that would be the demographic transition.

[…]

I suspect that the most valuable features of past societies – the ones that we read fantasy books to recapture, the ones that make Renaissance Faires and Medieval Times so attractive – have nothing to do with politics and cannot be restored through politics. In order to regain them, you’re going to have to roll back the Industrial Revolution. Needless to say, that makes fighting against the demographic transition look easy.

5. Who by Very Slow Decay

Scott Alexander describes the absolutely horrific state of end-of-life care — how we prolong life until it becomes a sick caricature of itself. It’s a harrowing read.

He explains that most doctors, when facing their own deaths, choose to engage as little as possible with the dystopian ‘end of life’ healthcare system they've spent their lives working within.

While reading the piece, I couldn’t stop thinking about Daniel Kahneman — one of the fathers of decision theory — who chose assisted suicide in Switzerland in 2024.

4. If It’s Worth Doing, It’s Worth Doing With Made-Up Statistics

Scott notices that sometimes our initial intuitions about the world are very very wrong

Remember the Bayes mammogram problem? The correct answer is 7.8%; most doctors (and others) intuitively feel like the answer should be about 80%. So doctors – who are specifically trained in having good intuitive judgment about diseases – are wrong by an order of magnitude. And it “only” being one order of magnitude is not to the doctors’ credit: by changing the numbers in the problem we can make doctors’ answers as wrong as we want.

So the doctors probably would be better off explicitly doing the Bayesian calculation.

He then argues that in some cases, it's better to just make up some rough numbers and run calculations on them, rather than relying on intuition alone. Doing so often leads to more accurate answers than we might suspect.

I like this idea. Math is a tool that makes System 2 thinking easier. At the very least, it forces one to formalize a problem more clearly, and this can help in understanding it better. Good advice for getting a quick handle on things.

Some things work okay on System 1 reasoning. Other things work badly. Really really badly. Factor of a hundred badly

3. The Poor You Will Always Have With You

In this response article, Scott Alexander explains how he sees societal change happening: for him values shift in response to economic, material, and technological conditions — it's not the values themselves that drive social change.

It's Urbanization + Growth -> Social Change -> Progressive Values

not

Urbanization + Growth -> Progressive Values -> Social Change

He argues that neoreactionaries adopt the latter mistaken framework — believing that values drive societal change — and as a result, they make big errors in their lucubrations (yes, 2013 was the year of articles against neoreaction).

I think this is a really useful way of looking at social change, and it lines up with how I think about the evolution of norms under "utilitarian naturalism." So of course I’m going to like this article — it flatters my preconceptions.

2. 90% of all Claims About the Problems with Medical Studies are Wrong

Scott is rightly preoccupied with how people interpret the true fact that much research is wrong or fails replication. Some say “this proves the medical establishment is clueless and hopelessly irrational and that two smart people working in a basement for five minutes can discover a new medical science far better than what all doctors could have produced in seventy years.”

So he takes some time to explain how scientific consensus actually work in practice.

Doctors are mostly reading famous influential studies like the ones mentioned above, which are at worst 40% and at best 5% wrong.

[…]

When doctors say that, for example, iron supplements help anaemia, it’s not because they hit iron on their Big Chart O’ Human Metabolic Pathways, then ran a single study, got p = .05, and rushed off to publish a medical textbook. It’s because they knew hemoglobin had iron in it, there are at least 21 randomized controlled studies, probably some had p-values closer to .001 than to .05 even though I don’t have any of them in front of me to check, and eventually some really really smart statisticians at the Cochrane Collaboration gave it their seal of approval.

Understanding the practical epistemology surrounding scientific consensus is of great importance, not just for interpreting medical research, but for making sense of academic findings in general.

Unfortunately, this article did not exactly spark a golden age of enlightenment, since twelve years later the situation has seemingly managed to get even worse.

1. A Thrive Survive Theory of the Political Spectrum

Scott floats a hypothesis about the origins of the two souls of the political spectrum.

My hypothesis is that rightism is what happens when you’re optimizing for surviving an unsafe environment, leftism is what happens when you’re optimized for thriving in a safe environment.

I have a soft spot for hypotheses like these, the kind that offer a different perspective on things we usually take for granted.

Scott walks through a list of reasons why he thinks his theory might be true — though he also has to admit there’s some counterevidence.

I admit some confusions. For example, it seems weird that poor people, the people who are actually desperate and insecure, are often leftist, whereas rich people, the ones who are actually completely safe, are often rightist.

But twelve years later, that counterevidence has more or less crumbled. If anything, the voting trends among poorer voters are starting to look like evidence in his favor. Some might even go so far as to say that Scott’s "theory" predicted the shift before it happened.

Maybe he really is as insightful as the legends say.

Subscribe now

I was one of those, sometimes a few at once. Now I think it’s the wrong way to go about things for atleast a couple of reasons.

1. A Renowned Problem

It’s a well-documented psychological phenomenon that affiliating with a group—even informally, or just “in spirit”—tends to erode one’s independent judgment. The identification gradually morphs into a kind of cognitive fusion until eventually you're not entirely sure whether you hold a position because you reasoned it through, or because everyone on your favorite podcast nodded vigorously when someone said it.

In the context of politics, the problem is significantly amplified. Unlike, say, a neighborhood book club, political organizations are structured around the pursuit and retention of power. This creates systematic incentives to promote loyalty and suppress dissent, even at the expense of truth or reason. The result is a kind of epistemic degradation: individuals become less concerned with whether a claim is true and more concerned with whether it advances the interests of “their side.”

2. A Peculiar Problem

Aside from the usual stuff, I suspect there’s a deeper—or at least weirder—issue. When people “pick a side” (or better: when a side is handed down to them by their parents) something strange starts to happen. People begin to believe that their morals come from their political ideology. That the ethical beliefs they hold aren’t just aligned with their political group, but are actually transmitted by the political movement itself.

It’s not “I believe in this party because it reflects my values.” It becomes “these are my values because this is what my party believes.” There’s a kind of moral inversion—where instead of using your ethics to judge politics, politics becomes the thing that tells you what’s ethical.

This is completely backwards.

Politics is supposed to be downstream from morality—not the other way around. A political movement is not supposed to produce your values but to reflect them.

The right order should be: first, work out what you think is good. Try to arrive at a coherent sense of what “goodness” even means to you. Then you adopt political positions that align with that definition.

3. Upsides from Refusing Political Allegiance

In general, I suspect that refusing political allegiance has some benefits:

1. It tends to make political discussions less vitriolic and more productive. You’re not defending a team—you’re engaging with ideas. In my experience that’s an upgrade.

2. During life we make many mistakes in things much simpler than politics. Refusing political allegiance makes it easier to walk away from bad political choices and stick with good ones, because you’re not emotionally tied to a party’s identity. You can actually change your mind based on evidence without feeling like a traitor.

3. You can pick and choose the good parts of different ideologies. Think free markets can do amazing things? Great—take a bite. Think inequality is a serious problem, and there’s solid evidence for a left-wing policy with reasonable trade-offs that might help? You can take a slice of that too. There’s no need to sign a loyalty oath to an entire worldview just to support one good idea. You can read widely and be willing to mix, match, or discard ideas based on how well they hold up under scrutiny.

4. It forces you to think about what morality actually is. When you’re not outsourcing your ethics to a party platform, you have to ask: where do my values come from? What do I actually believe? And since I suspect there is widespread moral confusion this seems like a plus.

In any case I no longer declare a political affiliation. When people ask, I just say, “I try to support whatever increases human well-being and flourishing.”

This tends to fluster people a little. Sometimes they’ll follow up with, “Okay, but who did you vote for?” I think what they’re really asking is: please, just give me a label so I know what script to run. It's the mind reaching for a comfortable cognitive shortcut. Just tell me what team you're on, goddammit!

But that’s exactly what I’m trying to avoid. I don’t want to know what group you belong to, either. I want to start from your values—what you care about, what you think a good society looks like—and go from there. Any political position I hold is an empirical matter that is downstream from my moral position, and therefore open to revision. I find this approach leads to more charitable and interesting conversations.

4. Objections

a) Belonging to a group has real benefits—companionship, protection, a sense of identity. Humans are social creatures; we like having a tribe. There’s comfort in knowing someone has your back.

b) A more interesting objection is this: “Aren’t you just pushing the in-group problem one level down?” If you reject political parties and instead base everything on your moral framework, isn’t that just... choosing a different kind of party?

Point a) is just true, there are positives that come with political allegiance. As for b)—it’s a valid worry, but I don’t think that’s how it actually plays out. When conversations begin with values instead of team loyalty, something shifts. There is more common ground, less vitriol. It also sets up the right pathway for talking about politics—begin with values, then move to the evidence that supports your political stance.

Although there might be something I am missing since many (most?) smart people declare some sort of political allegiance. I suppose I just need the right empirical evidence to change my mind on this stance too.

Subscribe now

In Study 1, we found that rates of ‘utilitarian’ judgment were associated with a broadly immoral outlook concerning clear ethical transgressions in a business context, as well as with sub-clinical psychopathy. In Study 2, we found that ‘utilitarian’ judgment was associated with greater endorsement of rational egoism, less donation of money to a charity, and less identification with the whole of humanity, a core feature of classical utilitarianism.

A 2023 study also replicated these results. This might seem odd to anyone who follows utilitarians online as they are generally bleeding-heart vegans who talk about their favourite charity and want to extend moral consideration to shrimps. The last people on planet earth you would accuse of lacking empathy. So what gives?

Harvard’s heavy hitter professor Joshua Greene thinks he has an answer.

1. Greene to the Clinical Rescue

In a previous article, we mentioned a study in which Greene explains that deontological type intuitions—such as 'incest is wrong' or 'murder is wrong'—appear to arise from quick, emotional, and automatic System 1 thinking. In contrast, consequentialist judgments tend to rely on slower, more effortful, and cognitively demanding manual System 2 reasoning.

We then suggested that this framing appears to imply a hierarchy of moral judgments, with deontological-type intuitions functioning as rules of thumb that typically lead to good outcomes without requiring the computationally costly moral calculus typical of consequentialism.

Greene agrees that System 1 heuristics are designed to work well on average, so he suggests a principle: we should be more willing to trust them in familiar, everyday contexts.

The No Cognitive Miracles Principle: When we are dealing with unfamiliar moral problems, we ought to rely less on automatic settings automatic emotional responses and more on manual mode conscious, controlled reasoning, lest we bank on cognitive miracles.

And what are these familiar scenarios? The ones we've encountered repeatedly, with plenty of trial and error over time.

Automatic settings can function well only when they have been shaped by trial-and-error experience . This may be the experience of our biological ancestors, as reflected in, for example, a genetic predisposition to fear snakes. Our automatic settings may be shaped by the experience of our cultural “ancestors,” as reflected in a fear of guns, despite one’s having never been harmed by one. Finally, our automatic settings are shaped by our own trial and error, as when one learns to fear hot stoves by touching them. These three mechanisms—genetic transmission, cultural transmission, and learning from personal experience—are the only mechanisms known to endow human automatic cognitive processes with the information they need to function well. For one of our automatic settings to function well, its design must be informed by someone’s trial-and-error experience. Any other way, and it’s a cognitive miracle.

According to Greene, this helps explain why psychopaths sometimes arrive at the 'right' answer (by which he means the consequentialist one) in moral dilemmas; in edge cases, the System 1 deontological intuitions that guide most people can misfire—leading them astray.

For Greene, this not only resolves the apparent paradox but also serves as evidence that his theory is on the right track.

A corollary of the no cognitive miracles principle is that we should expect certain pathological individuals—VMPFC patients? Psychopaths? Alexithymics?—to make better decisions than healthy people in some cases. (This is why such individuals are no embarrassmen to the view I will defend in the next section.)

But at this point, a deontologist might object: are you seriously handwaving away the fact that, in some cases, psychopaths give “better” moral answers than the average person?

Let a philosophical vermin (me) offer a slightly more empathetic reconciliation.

2. Learn to Love the Deontological Intuitions

There’s a tendency among consequentialist philosophers to try to debunk all deontological type intuitions. But if Greene is right—and if consequentialism is where the foundation of morality lies—then these quick moral heuristics, whether learned or evolved, aren’t just frequently reliable; they’re indispensable. After all, we can’t go around calculating every possible branch of reality our next action might set in motion like Doctor Strange.

a) The Organ Transplant “Counterexample”

Take the classic organ transplant thought experiment, often used as a challenge to utilitarianism. A healthy person walks into a hospital for a routine checkup. Meanwhile, five other patients in the same hospital are dying—each in need of a different organ to survive. Should the doctor sacrifice the healthy individual, harvest their organs, and save the five?

Every one of my lovely normie friends, upon being asked this question (I’m fun at parties), has a strong, immediate visceral reaction: absolutely not. You don’t kill people in hospitals. In other words what they’re expressing is a classic deontological type intuition: 'killing an innocent person is wrong’. Not one of them responds in the classical utilitarian manner to resolve the dilemma—by highlighting that the second-order effects of killing innocents in hospitals would be disastrous. Unanticipated consequences would include the collapse of trust in doctors, the implosion of the healthcare system, and potentially, the unraveling of society itself.

For my lovable normie friends (who are no doubt thrilled to be brought up as the representatives of folk morality, but with whom I have a healthy relationship built on making fun of each other), it’s their rule based intuitions that are saving them from collapsing society, and that’s great and healthy.

And it’s not just them. I use quick, rule-of-thumb heuristics all the time when making moral decisions. So does everyone else. Every sane person, anyway. Nobody’s running a full utilitarian simulation every time they decide whether to tip the barista or hold the door open. The sheer time sink alone would produce massive disutility. That’s not moral philosophy—that’s a mental health crisis.

b) Unconscious Realism

Another phenomenon I’ve noticed is that rules based intuitions often give the 'wrong' answer (read: non-consequentialist) in abstract counterexamples. But as soon as you reframe those examples in a more realistic setting, the 'wrong' answer starts to look very much like the right one.

The clearest case of this might be the footbridge dilemma. In this scenario, you're asked whether it would be moral to push a large person off a bridge to stop a runaway trolley headed toward five people tied to the tracks. The 'correct' consequentialist answer is yes—but most people instinctively say no.

Now, if we imagine this scenario playing out in the real world—where we don’t have perfect, oracle-level knowledge of the future—it becomes obviously immoral (and frankly, moronic) to try pushing someone off a bridge to stop a trolley. There are about 10,000 ways things could go horribly wrong.

So, in the realistic setting, our deontological type intuitions actually get the answer exactly right. Why? Because their purpose is to help us navigate the real world, where information is messy and outcomes are uncertain. It’s not all that surprising that they misfire when applied to abstract, idealized thought experiments. They weren’t built for deterministic hypotheticals; they were built for our highly complex probabilistic universe.

This isn’t something a philosophical vermin like me is just making up—it relates to a concept that has been statistically tested called unconscious realism. As Greene himself explains, unconscious realism is

a tendency to unconsciously replace a moral dilemma’s unrealistic assumptions with more realistic ones.

Unconscious realism leads people to project their intuitions onto subtly reconstructed moral dilemmas with more realistic assumptions, potentially distorting the intended analysis of abstract hypotheticals. Greene’s own experiments provide evidence of this phenomenon occurring.

There was a significant main effect of WORSE ( F (1, 417) = 5.80, p = 0.02) with actions expected to be less successful eliciting lower moral acceptability ratings, consistent with unconscious realism.

But Greene understandably isn’t thrilled that his own studies on abstract moral dilemmas reveal evidence of unconscious realism, as it undermines the strength of conclusions drawn from the entire field of trolleyology in cases where such effects were not controlled for. So he comes up with a couple of reasons for why we shouldn’t take this evidence too seriously.

The statistical significance of the ‘‘unconscious realism” covariates included in Experiments 1a and 2a provides limited support for the unconscious realism hypothesis. This support is limited for at least two reasons. First, subjects’ assessments of the likely real-world effects of the actions in question may be post–hoc rationalizations (Haidt, 2001). Second, a correlation between real-world expectations and moral judgments is not sufficient to establish a causal relationship. Nevertheless, these results indicate that effects of unconscious realism may be real and that researchers who use hypothetical cases to study decision-making should consider controlling for such effects as done here.

Nonetheless, unconscious realism helps explain why people often give 'incorrect' (non-consequentialist) responses to abstract moral dilemmas that would be entirely appropriate in real-world scenarios. It’s not that our intuitions are inherently flawed, it’s that they aren’t designed to function well in artificial hypotheticals under assumptions of perfect future information.

3. Psychopathy Leads to Bad Consequences

The considerations I’ve offered in sections 2 are meant to highlight the fact that rule based intuitions and consequentialist considerations can, descriptively, be seen as happily married together (“now kiss”). Viewing deontology as applied consequentialism under cognitive constraints, it becomes clearer why deontological intuitions are valuable. Conversely, the absence of such intuitions represents a form of moral degeneracy that may lead a person to commit actions with harmful consequences.

We might even hazard a guess: the key difference between psychopaths and shrimp-hugging, bleeding-heart utilitarians is that psychopaths lack emotional, deontological-type intuitions and thus give colder, more inhuman answers. While utilitarian thinkers do have the “folk” intuitions—but they’ve spent years obsessively analyzing them, realizing they don’t build a coherent moral theory. So they override them, take the abstract thought experiments like an exam question, and respond like the principled, analytic creatures they are, knowing very well that in more realistic scenarios they would often act differently.

After all, it would be quite strange if utilitarian philosophy itself—unlike utilitarian responses to moral dilemmas—was associated with psychopathy, given that its founding figures, like Bentham, Mill, and Sidgwick, were passionate social reformers that advocated for the abolition of slavery, public education, environmental protection, prison reform, animal rights, workers’ rights, and women’s rights.

Indeed, through a process of dissociation analysis, the tireless Greene finds that utilitarian judgments to moral dilemmas aren't driven solely by antisocial tendencies—it would be very weird indeed if they were—but also by a prosocial concern regarding outcomes with greater collective well-being.

The results of Study 2 indicate that individual differences in the tendency to make sacrificial utilitarian judgments are not driven entirely by antisocial tendencies (low D), as Kahane and colleagues suggest. Instead, sacrificial utilitarian judgments appear to reflect two independent motivational components which vary across people: (an antisocial reduced concern for causing harm (i.e., low D) and an increased prosocial concern for maximizing good outcomes (i.e., high U).

4. When to Go Full Utilitarian?

After this heartfelt defense of our quick, rule-of-thumb moral intuitions from a consequentialist standpoint, you might start to wonder if we can just stop here. Why try to debunk any intuition at all? Maybe it’s fine to trust the little moral voice evolution (or grandma) gave us. After all, these intuitions have apparently been passed down—genetically or memetically—to help us make decent choices most of the time without having to go rain-man mode every morning over breakfast.

However, with a bit of reflection, we can see that there are situations where relying solely on snap intuitions just isn’t enough. For instance, these kind of moral intuitions often clash—sometimes dramatically—across cultures. This isn’t because some people are wired differently from others (the evolution of prosociality and cooperation is though to be extremely ancient) but because different societies, shaped by complex historical and environmental forces, pass down different moral heuristics, rules, and values. In short: they inherit different rule-of-thumb intuitions.

Utilitarianism offers a way to navigate this pluralism—not by discarding quick intuition altogether, but by providing what Greene calls a metamorality: a higher-order framework, accessible through deeper deliberation, for evaluating which moral norms (that shape our intuitions) best promote collective well-being

Along similar lines, an area Greene doesn’t spend much time on—but one that seems especially well-suited for the full utilitarian treatment—is the question of whether our current norms are still doing their job, or if we should replace them with new policies. This is unfamiliar territory for our rule-of-thumb intuitions, which are generally passed down to navigate small-scale, familiar social situations—not abstract policy decisions or large-scale governance.

Utilitarian thinkers, to their credit, are often quite good at identifying when folk moral rules have quietly outlived their usefulness. Traditional gender roles regarding labor, for instance, made more cultural sense thousands of years ago given certain social and economic conditions, but at some point, their usefulness diminished drastically. This is part of what John Stuart Mill recognized when he argued that excluding women from the workforce was unjust. Today, many utilitarians believe that our norms around eating animals fall into a similar suspicious category.

All of this suggests that utilitarian moral reasoning shouldn’t be about tossing out our rule-of-thumb intuitions altogether, but about complementing, challenging, and revising them when they fall short—figuring out which intuitive moral norms are still broadly beneficial, and which ones we might be better off not passing down to future generations. Utilitarianism, in this sense, can tend to the garden of our necessary deontological intuitions, continually refining and updating our cultural norms in service of the greater good.

Subscribe now

Joshua Greene’s The Secret Joke of Kant’s Soul is a heavily cited paper, which means that most people into moral philosophy have read it, as opposed to the majority of other papers that exist.

Greene, a famous philosopher-neuroscientist hybrid sets out to answer a question: why was Kant so against masturbation?

To do so, he examines the psychological foundations of moral judgment by comparing two major ethical theories: deontology and consequentialism.

Greene argues that deontological judgments (morality based on norms, rights and duties) are primarily driven by automatic emotional responses—quick, intuitive reactions that don’t require much conscious thought. While consequentialist judgments (morality based on outcomes and consequences) rely on more cognitively expensive processes, such as deliberate reasoning and cost-benefit analysis.

To convince us of this Greene whips out a bunch of empirical evidence from neuroimaging studies, behavioral experiments, and psychological tests, demonstrating distinct neural correlates for each type of moral judgment. Regions of the brain linked to higher-order cognition—like abstract reasoning and deliberation—light up during consequentialist thinking, while “emotional centers” are more active when people consider duties, values, and moral norms.

But wait what has all this to do with masturbation and Kant?

Well, Kant is the most famous deontologist, and one of the moral rules he took seriously—very seriously—was the no masturbation rule: that one should never, under any circumstances, engage in that kind of self-pleasure. Naturally, being a philosopher, he wasn’t content to simply declare masturbation immoral, he had to justify it.

But this, according to Greene, is where Kant reveals the secret of his soul.

2. Kant’s Argument Against Masturbation

So what is Kant’s argument against genital self-pleasure? Or, as Greene might be tempted to say, Kant’s “argument.” Here’s how Kant himself puts it:

That such an unnatural use (and so misuse) of one’s sexual attributes is a violation of one’s duty to himself and is certainly in the highest degree opposed to morality strikes everyone upon his thinking of it... However, it is not so easy to produce a rational demonstration of the inadmissability of that unnatural use, and even the mere unpurposive use, of one’s sexual attributes as being a violation of one’s duty to himself (and indeed in the highest degree where the unnatural use is concerned). The ground of proof surely lies in the fact that a man gives up his personality (throws it away) when he uses himself merely as a means for the gratification of an animal drive.

First Kant declares that masturbation strikes everyone as being wrong then he says that this must surely be because “a man gives up his personality when he uses himself merely as a means for the gratification of an animal drive.”

If this justification seems a bit strange to you, it’s because it is. What does using oneself as a means for the gratification of an animal drive mean exactly?

What about quenching thirst by drinking water—is that also immoral because we are using our mouth as a means to satisfy an animal drive? What about scratching an itch? It’s something purely physical, done for immediate relief and pleasure. What about humming a song because it feels nice?

More importantly, why is Kant, this philosophical titan of the past, offering a justification so prone to obvious objections and counterexamples?

Well, according to contemporary philosophy star Michael Huemer, it might be because many great philosophers of the past (Kant included) were “bad at thinking”.

My introduction to philosophy was largely through the great philosophers of the past -- the likes of Plato, Aristotle, Descartes, Hume, Kant. From the beginning, I was struck by how bad they were at thinking. Sometimes, they just seemed to be bad at logic, committing fallacies and non sequiturs that even an undergraduate such as myself could quickly see. Other times (almost always!), they seemed to have extremely poor judgment, happily proclaiming absurd conclusions to the world, rather than going back and questioning their starting points.

As a brief aside: when I first came across this idea, I was completely and utterly flabbergasted. After all, isn’t thinking about things literally the job of philosophers? In the following quote from Huemer, I was very much part of most people:

Not everyone realizes all this. Most people, I suspect, believe that the Great Philosophers are actually good at philosophy.

Other contemporary thinkers go even further than Huemer, suggesting that some of history’s celebrated thinkers may have, at times, been less than timeless geniuses and more peddlers of pseudo-profound nonsense.

When it comes to Kant, Greene is more gallant in tone than Huemer—but in terms of substance, his take isn’t all that different.

3. Greene Exposes the Secret of Kant’s Soul

Greene argues that while Kant tries to ground his prohibition in lofty appeals to personal dignity and self-respect, he admits that it's hard to produce a convincing rational demonstration for why masturbation is morally wrong. For Greene, this isn’t a trivial aside—it’s the giveaway. It’s the “emotional centers” lighting up in Kant’s brain generating the deontological intuition, not the regions dedicated to abstract reasoning. Indeed, what Kant sees as pure moral reasoning is, in fact, a post-hoc rationalization of emotional moral intuitions that Kant views as deriving from God.

You can’t really blame Kant for trying to rationalize his moral intuitions. His intuitions derive from his human nature (“the moral law within”), and ultimately from God. God’s a smart guy, Kant must have thought. He wouldn’t give people moral intuitions willy nilly. Instead, we must have the intuitions we have for good reasons. And so Kant set out to discover those reasons, if not by force of reason then by feat of imagination.

Greene doesn’t quite come out and say he’s psychoanalyzing Kant, but that’s effectively what’s happening. For him Kant is trying really hard to post-hoc rationalize the no-fap rule. The emotional intuition that masturbation is wrong comes first, and the high-minded rational justification gets built afterward to make it all look respectable—but it doesn’t hold up to much scrutiny.

In other words, Greene quietly and slightly more politely joins Huemer in the corner of those who think Kant is doing bad reasoning.

a) Nietzsche’s Remark on Kant

Greene’s backup doesn’t end with Huemer. To support his thesis he references the time Nietzsche claimed that the secret joke of Kant’s soul was that he wrote not against popular prejudices, but on their behalf.

“Kant wanted to prove, in a way that would dumbfound the common man, that the common man was right: that was the secret joke of his soul. He wrote against the scholars in support of popular prejudice, but for scholars and not for the people.” — Friedrich Nietzsche

According to Greene Nietzsche got it right: Kant’s complex and confusing moral reasoning was designed to defend common emotional intuitions rather than genuinely deriving moral conclusions from rational first principles. He took intuitive emotional reactions, polished them up, and presented them as if they were the product of pure reason. But they weren’t.

4. Greene Exposes Deontology

Empowered by his novel findings, Greene doesn’t want to limit himself at just critiquing Kant, his ambitions are greater. He suggests that because humans are emotional creatures prone to post-hoc rationalization, the way Kant approached morality isn’t some isolated quirk—it’s the mechanism by which much of deontological moral philosophy is generated in the first place.

We are now ready to put two and two together. What should we expect from creatures who (1) exhibit social/moral behavior that is driven largely by intuitive emotional responses, and (2) are prone to rationalization of their behaviors? The answer, I believe, is deontological moral philosophy.

Put less kindly, for Greene deontology stems from our remarkable ability to fool ourselves. We have a strong emotional reaction, and then we invent a principle to match it. Instead of admitting it's a baseless gut feeling, we construct a post-hoc rationalization and call it moral philosophy.

Deontology, then, is a kind of moral confabulation. We have strong feelings that tell us in clear and uncertain terms that some things simply cannot be done and that other things simply must be done. But it’s not obvious how to make sense of these feelings, and so we, with the help of some especially creative philosophers, make up a rationally appealing story: There are these things called “rights” which people have, and when someone has a right you can’t do anything that would take it away.

Needless to say, in philosophical circles, these are fighting words. Some deontologists were not amused. To which, some say, Greene replied: “Don’t get all emotional about it.”

5. Normative Implications

Toward the end of the paper, Greene asks himself whether his analysis has any normative implications. Do these scientific findings actually tell us something about how we ought to act? Greene thinks they do.

I believe that science does matter for ethics, not because one can derive moral truths from scientific truths, but because scientific, information can undermine factual assumptions on which moral thinking implicitly depends […] The relevance of science, then, is that it can tell us how our moral intuitions work and where they come from. And once we understand our intuitions a bit better we may view them rather differently.

So where do deontological intuitions come from? Greene suggests they originate from evolved psychological drives—like the impulse to punish wrongdoers, or the tendency to care more for those who share our genetic background.

We have inferred on the basis of the available evidence that the phenomenon of rationalist deontological philosophy is best explained as a rationalization of evolved emotional intuition.

According to a philosophical vermin (me), this is one of Greene’s weaker points. It seems pretty clear that many deontological intuitions aren’t just evolved—they’re learned. Take Kant, for example. Greene himself recognizes that his moral instincts are likely due to his strict Lutheran upbringing, not evolution.

But Greene wants to argue that because deontological intuitions often stem from blunt biological facts, this spells trouble for their credibility as moral judgments. They’re not glimpses into moral truth; they’re artifacts of evolutionary wiring. In Greene’s view, that makes them epistemically suspect—products of natural selection, not reason. In other words, evolutionarily debunked.

While the mighty consequentialism appears to carry more moral weight

This is in contrast to consequentialism, which, I will argue, arises from rather different psychological processes, ones that are more “cognitive,” and more likely to involve genuine moral reasoning.

6. The Secret of Greene’s Soul

In the final paragraphs of the paper Greene becomes self-aware and troubled by the evolutionary debunking argument he just used to critique deontology.

Taking these arguments seriously, however, threatens to put us on a second slippery slope (in addition to the one leading to altruistic destitution): How far can the empirical debunking of human moral nature go? If science tells me that I love my children more than other children only because they share my genes (Hamilton, 1964), should I feel uneasy about loving them extra? If science tells me that I’m nice to other people only because a disposition to be nice ultimately helped my ancestors spread their genes (Trivers, 1971), should I stop being nice to people? If I care about my self only because I am biologically programmed to carry my genes into the future, should I stop caring about myself? It seems that one who is unwilling to act on human tendencies with amoral evolutionary causes is ultimately unwilling to be human. Where does one draw the line between correcting the near-sightedness of human moral nature and obliterating it completely?

Greene would like to immunize consequentialist moral reasoning from the same evolutionary debunking argument he uses to vanquish deontology. But he’s too smart not to see where that road leads. It doesn’t matter much that consequentialist intuitions involve “genuine moral reasoning” or that they arise from different cognitive processes. They’re still cognitive processes—still shaped by evolution—and therefore, just as vulnerable to the same kind of debunking.

Greene seems to recognize this, which is probably why he chooses to close the paper with the final passage quoted above—leaving the question hanging, rather than stepping into the same trap he just sprung on deontology.

Nonetheless, Greene is right to note that evolutionary debunking arguments raise questions of their own. One that I find especially pressing is this: if evolution endowed us with certain moral predispositions, why should we trust—or care about—other ones?

The evolutionary debunking point of view seem to smuggle in the assumption that there are alternative moral facts out there, somehow more “real” or trustworthy. But how exactly are we supposed to access them? Through what—some kind of non-evolved moral perception module? A built-in truth-tracking moral compass evolution forgot to tamper with?

There’s something a little suspicious about this setup.

But unfortunately for Greene, without the evolutionary debunking card, he loses the very tool he was using to undermine deontology. The upper hand slips away.

7. A Stronger Argument for Consequentialism?

After finishing his paper, Greene must have sensed that his neuroscience research pointed toward an even stronger argument. Indeed, in a subsequent article, he expands on his ideas and suggests that it’s reasonable to think of deontological intuitions as fast, automatic System 1 heuristics, while consequentialist judgments are the slower, more deliberate product of System 2 thinking—the brain’s manual mode.

This leads him to argue that “we should distrust our automatic settings and rely more on manual mode when attempting to resolve practical moral disagreements.” But perhaps we can take this one step further still.

If we recognize that, according to modern psychological theories, many cognitive biases arise from system 1 processes as quick heuristics to save mental resources, we can begin to sketch a hierarchy of moral judgment.

Given that evaluating the full consequences of our actions is computationally expensive, deontological instincts appear to have evolved as fast, efficient shortcuts—rules of thumb that generally lead to good outcomes without requiring complex moral calculus. But what are they shortcuts towards? Well, they would be shortcuts towards the more cognitively demanding “manual mode” given by consequentialism.

Under this view, deontology isn’t a rival to consequentialism, but a practical proxy. Deontological principles act as quick and efficient ways to achieve good outcomes in practice—rules, norms and values that operate in service of the deeper moral system they’re approximating: consequentialism.

In the end, to highlight the significance of consequentialism, Greene didn’t need to debunk deontological intuitions, just show their actual purpose.

Subscribe now

Sigmund Freud famously claimed that humanity has suffered three great humiliations.

The first was Copernican: the moment we learned that Earth isn’t the center of the universe.

The second was Darwinian: the realization that we weren’t created in a flash of divine craftsmanship, but rather slowly and messily evolved from other animals. Our great-great-…-great-grandfather was a fish.

And then Freud proclaimed himself as the bringer of the third humiliation: the idea that we’re not masters of our own minds but we are governed by subconscious urges.

When I was younger—and significantly more edgy—all this talk of cosmic humiliation got to me. I started to wonder: what if the true answers to humanity’s biggest questions are just... the ones we least want to be true? That humanity engages in century long cope sessions and that the real path to truth is simply giving up the need to feel special?

2. The Epistemic Theory of Cosmic Humiliation

Enter The Epistemic Theory of Cosmic Humiliation (ETCH). The ETCH holds that, when faced with any grand philosophical question, the correct answer is whichever one is most profoundly insulting to human self-importance.

Do we have free will, or are we just meat puppets dancing to the tune of physics? According to the ETCH we’re the meat puppets.

Does God exist and care about us? Apologies—according to the ETCH, we’re just the accidental side effect of some overly friendly amino acids in a lukewarm tide pool.

Are we special because of our unique, ineffable consciousness? The ETCH suggests otherwise. There’s no magical line dividing “mind” from “not-mind”—just a slow gradient from bacteria to Bach.

Ah, but human intelligence! Surely that’s the line. Our reasoning, our logic—totally unreplicable, right? Enter the ETCH holding up a 2032-era AI that just proved the Riemann Hypothesis.

Well then, surely we’re uniquely artistic and creative. We can’t just be the amniotic fluid sustaining the birth of a new being. Right…right?

3. The Epistemic Theory of Cosmic Humiliation is Wrong

There are many ways to show that the ETCH is wrong. But my favorite way to prove it’s wrong is that it’s wrong because it contradicts one of my beliefs.

See, I think Utilitarianism, in some real sense, captures moral truth. But under ETCH, moral relativism or even nihilism would be more humiliating—and therefore more likely to be “true.” So by ETCH logic, my belief in Utilitarianism must be false. Which is obviously unacceptable. Therefore, ETCH is wrong.

If you're not convinced that the passage above thoroughly dismantles the ETCH, consider some more “mundane” physical theory—like the theory of magnetism. It’s true, and yet, not particularly humiliating.

In any case, the ETCH is worse than wrong. It’s depressing. It wraps the universe in an unnecessarily edgy, pessimistic filter. Moreover, I believe in some of these so-called “humiliations,” and I don’t find them all that humiliating to be honest.

4. What Remains of the ETCH?

Despite being ridicolous and false I think there is a tiny, annoying kernel of truth to the ETCH. We do like stories that flatter us. We do have a habit of post-hoc rationalizing beliefs that we like, and some intellectual traditions can fall prey to this psychological dynamic.

I find it plausible that when faced with uncertainty, humanity has often gravitated toward theories that make us feel better. So maybe some of these big questions linger—not because they’re unsolvable, but because we don’t like the most plausible answers. This reasoning would predict that, as our understanding of the universe deepens, the “humbling” of humanity will continue.

So while I don’t buy the ETCH as a theory, I sometimes use it as a kind of quick check: Do I believe a certain thing because it’s true, or because I want it to be true?

Subscribe now