Weaponised Autism as the Font of Human Creativity

weaponized autism train pepe face cover image


This post builds on my review of The Beginning of Infinity, although you don’t have to read that one first.


What makes humans special? How is it that we are able to unleash the energy of the atom, transmute handfuls of sand into powerful djinn, and generally manipulate matter in any way not strictly forbidden by the laws of physics, while our hominid ancestors gather dust in the natural history museum? What tectonic forces have torn such a chasm between us and our surviving cousins, still banging rocks together in the animal kingdom?

For most of history this was no great mystery: only humans have an immortal soul, granted to us by a Creator who made us in His own image.

Then Linnaeus and Darwin revealed we were merely naked apes, sparking the great debate that remains unresolved almost three hundred years later: what is our secret sauce as a species?

Some rely on bold conjectures: it’s our big brains. Others make convincing arguments for our language skills, or twiddle their opposable thumbs suggestively.

Wrong, says Simon Baron-Cohen. In fact, the font of all human creativity, the source of all our knowledge, and the key to becoming masters of the universe is… autism.

While you may wonder if this is a skit by the S. Baron-Cohen who is a famous graduate of clown school, this is, to be clear, the opinion of his cousin; the S. Baron-Cohen who is a very serious researcher of developmental psychopathology at Cambridge.

The Pattern Seekers: How Autism Drives Human Invention (2020) is Baron-Cohen’s latest salvo in a long-running battle to publicise his fascinating research, much of which is politically unpalatable and therefore doesn’t receive the quality of criticism it deserves.

The Pattern seekers review, Simon Baron-Cohen: how autism drives human invention

In the course of examining this book’s grand theory, we will discover important clues to several mysteries that our host has previously pondered: from puzzles about autism itself, to the nature of creativity, to whether or not current AIs are on a path to superintelligence.

We’ll also see how Baron-Cohen’s thesis accords with unconventional theories about knowledge creation promulgated by David Deutsch—a revitalisation of Karl Popper’s epistemology which makes surprising predictions about the true nature of human capabilities.


Autism as the Font of Human Creativity

The thesis in short: the secret sauce that only humans possess is the systematising trait

The systematising trait is genetic. It evolved somewhere between 70,000 and 100,000 years ago. It is characterised as the ability to run thought processes with an if-and-then structure: if you plant a seed, and the soil is moist, then the seed will sprout.

If you are familiar with Boolean logic then you will know that these simple operations can be built up into arbitrarily complex calculations, limited only by the point at which your Excel sheet bricks your entire computer.

In humans, this snowballing in complexity required a second catalyst: the theory of mind required to model what other people are feeling and thinking, also known as the empathising trait, which arose in roughly the same period.1

Systematising and empathising are the twin helices that weave throughout Baron-Cohen’s body of work. His research involves coming up with questions that allow us to express the ability to systematise and empathise as quotients (from here on in, ‘SQ’ and ‘EQ’), then running large questionnaire-based studies to see how these traits are distributed amongst various populations, and what else they might be correlated with.

If we combine SQ and EQ in a two-by-two, we get a matrix of neurotypes that looks like this:

Simon Baron-Cohen neurotype matrix of systematising and empathising

Baron-Cohen splits these neurotypes into five categories.

In the top left corner, we have extreme systematisers, accounting for 4 per cent of men and 2 per cent of women.

The next band contains those for whom SQ is higher than EQ but not extremely so, accounting for 40 per cent of men and 26 per cent of women.

The inverse band in which EQ is higher than SQ makes up 40 per cent of women and 24 per cent of men.

In the bottom right corner, we have extreme empathisers, comprising 3 per cent of women and 1 per cent of men.

Finally, everyone in the middle band has a roughly balanced EQ and SQ: 30 per cent of women and 31 per cent of men.

(All of these numbers are kind of tautological, in that e.g. the researchers normed it so everyone above the 2.5th percentile was considered an extreme systematiser, but at least they give us some archetypes to play with.)

Let’s start with the elephant in the room. EQ and SQ are both textbook examples of sexual dimorphism, with overlapping distributions and two distinct peaks. To literally nobody’s surprise, some people are mad about this.

Trying to litigate sex differences in the brain goes way beyond the scope of this review and is incredibly boring for anyone who already accepts that evolution doesn’t magically begin from the neck down. So let’s stick with the connection between creativity and autists of all sexes and genders.

The first interesting finding is that EQ and SQ are pretty much independently distributed, suggesting they have distinct causes and don’t trade off against one another. But they are slightly anti-correlated. Why?

After some cunning research, Baron-Cohen and friends tracked down the likely cause: expression of the relevant genes is mediated by levels of testosterone in the womb. This happens to affect both traits in opposite directions, so a foetal brain soaked in testosterone leads to a higher SQ, but a lower EQ.

Your SQ score predicts whether or not you work in a STEM field (much more so than your sex does!), how well you can visualise mechanical reasoning, and so on. Meanwhile, your EQ predicts how well you perform on things like facial emotion recognition tests.

None of these findings are particularly surprising. In fact, if you stitch them all together, they pretty much add up to the Internet folklore definition of an autist: probably male, works in STEM, struggles to make eye contact, ‘sperges out’ over esoteric subjects, and so on.

As it turns out, there’s a strong overlap in the genes associated with autism and the systematising trait. Autistic men are 50 per cent more likely to have a high or extreme SQ than the general population, while autistic women are almost double the base rate.

But autism and systematising are not synonymous!

One of the reasons we confuse these traits is that there is no foolproof way to bucket them into natural categories. Systematising ability is very much a continuous variable. Autism disorder also falls on a spectrum: the folk idea that everyone is “a little bit autistic” is wrong, but it’s certainly possible to have some autistic traits without receiving a diagnosis. So the categories inevitably get messy, but we can still make an important distinction: the vast majority of high-systematisers are not diagnosed as autistic, and plenty of autistic people are not that great at systematising.

So we now have:

  • a name for the-thing-we-colloquially-call-autism-but-which-isn’t-autism,
  • an understanding of why these two groups tend to overlap,
  • a mechanism for how they arise: a cluster of related genes, mediated by hormone levels in the womb.

This very useful distinction might be the best thing about Baron-Cohen’s book. Unfortunately, his tendency to equivocate between the benefits of systematising and the benefits of actually being autistic might be the worst thing about this book, on which, more later.

First, let’s work through a few obvious counterarguments to the big central claim about autism driving all human creativity and invention.


Obvious Objections

But isn’t autism rigid and repetitive? That sounds like the opposite of creativity.

In short: yes. But that’s not always a bad thing!

Here it might be helpful to break the systematising mechanism into substeps:

  1. Display curiosity about the world; ask a question
  2. Answer the question by hypothesising an if-and-then pattern
  3. Test the pattern in a loop: repeat observations to see if it holds true
  4. Modify the pattern by varying the ‘if’ or the ‘and’ to see what happens to the ‘then’
  5. Test this modified pattern in a loop

…which looks pretty much like the good ole scientific method.

What happens when a set of fusty rules that science students reluctantly learn becomes a drive so instinctive and obsessive that you have to stop yourself from doing it? In Internet-speak, you get weaponized autism: think Edison working 18-hour days, ignoring family obligations, refusing to bathe, napping under his table reeking of sweat and solvents.

Baron-Cohen argues that systems thinking is necessary not only for iterating lightbulb filaments, but for penning essays, practising free-throws, or composing violin concertos. The systematising trait determines how good you are at spotting patterns, which is a precondition for creating them in whatever domain you happen to be drawn towards.

Could the actual medical condition of autism, as distinct from being a high-systematiser, convey any creative advantages? I think so, at least in some circumstances. Great musicians and auteurs of cinema often don’t care about pleasing anyone except themselves: they have a direct connection to the muse, and refuse the design-by-committee approach which would only dilute their work.2

Where autists might be at a disadvantage is in creative acts that are specifically designed to elicit a certain emotional response from others, or involve accurately modelling other people’s motivations and desires, e.g. literary fiction or psychology.

So: systematising is necessary for creating new knowledge and art, but it’s not always sufficient—sometimes you need a strong ability to empathise, too.


But don’t non-human animals use tools too?

Yes. Sometimes this looks quite sophisticated: crows have figured out not only to drop nuts on the road so cars will run them over, but to drop them on the pedestrian crossing where they are safer to retrieve.

But this is mere “associative learning”, says Baron-Cohen. Think Skinner’s pigeons or Pavlov’s dogs: many animals can learn to associate a specific action or series of actions with a reward, and can even pass on the trick to other group members by imitation (mimesis).

Baron-Cohen doesn’t do the best job of explaining exactly what distinguishes associative learning from if-and-then reasoning. As far as I can tell, the key is that animals can’t mess around and iterate in a generative way. If they stumble onto something by chance, great: they might repeat the action, and even pass it on to others. But they lack the drive to see how it generalises.

chimp poking a stick in termite mound
ucumari photography | CC 2.0

The upshot is that in wild animals, any tool use involves essentially one step, or at most two: a sequence of events they could have plausibly stumbled onto by accident.

Chimps are a pretty great test case here, in that we have a recent common ancestor (so they’ve had the same amount of time as us to invent disc golf and nuclear fission) and they have plenty of raw smarts, to the point where they can e.g. memorise nine-digit sequences much faster than humans can. But even in captivity, chimps are almost incapable of understanding causal reasoning, and lack the instinct to experiment that is already present in human infants.

Baron-Cohen works through a bunch of other potential counterexamples, from birdsong to sponge-wielding dolphins to hawks starting fires, arguing that these are all more parsimoniously explained as associative or mimetic learning.

I found his arguments convincing, but then (just before deadline) I talked to an AI researcher who was equally convinced in the opposite direction by Frans de Waal’s book Are We Smart Enough to Know How Smart Animals Are?3

This seems like an important question to resolve. Fortunately, Baron-Cohen has given us a clear path to falsify his theory: all it would take is any example of an animal using a multi-step invention that couldn’t plausibly have arisen by chance.


Didn’t other human species invent tools?

Again: yes, but only the kind of tools you could plausibly stumble across by chance, with no evidence of any drive to experiment. Across a span of millions of years, the only big breakthrough was in banging rocks together in a slightly more sophisticated way.

paleolithic stone tool
literal cutting-edge technology in the palaeolithic

Baron-Cohen’s reading of the archaeological record is that almost nothing happened up until approximately 70,000-100,000 years ago, and then it all happened in an explosion of activity: jewellery, controlled fire, engraving, bow-and-arrow hunting, sewing, art, dwellings, and so on.

In a book this slim and sweeping, there isn’t room to ground these claims in ways that might satisfy subject matter experts. Again, the important thing is that it gives us a clear criterion to reject the theory that homo sapiens were the only hominins capable of systems thinking: all it would take is a neanderthal necklace, or a homo habilis house.


Isn’t the real cause of the cognitive revolution opposable thumbs/bipedalism/larger brains?

No. Or at least, that’s not sufficient: apes have opposable thumbs, neanderthals had bigger brains than we do, several other hominins were bipedal, etc.


Why would a complicated polygenic trait lead to a sudden step change in ability?

Mutations are way less exciting than X-Men led us to believe: you can develop the ability to e.g. use electromagnetic radiation to navigate your environment, but you have to start with a near-useless cluster of photosensitive cells and make a bunch of incremental improvements over 100 million years—you don’t jump straight to a fully-formed eyeball.

eye evolution
wikimedia | CC 3.0

Baron-Cohen acknowledges this. He even says we should expect to see some kind of proto-systematising ability in other species. But then he…spends a lot of time arguing we don’t, in fact, see anything like this?

Perhaps the systematising trait is so hugely advantageous that it evolves rapidly, leaving little trace of proto-versions. Or maybe it’s surprisingly simple to code for, and there are no proto-versions. Perhaps the apparent binary comes from some other variable that allows systematising genes to be expressed at all.

I don’t know. But based on the evidence Baron-Cohen presents, it really does appear that the systematising trait is something that we and only we possess.


Some Satisfying Mysteries Explained

Onwards to several mysteries for which Baron-Cohen’s book gives us either a plausible resolution, or at least some valuable clues and cruxes for further exploration.

1. Why are some humans so much more creative than others?

I didn’t need much convincing to believe there is a binary attribute that divides modern humans from all other species, and that attribute is something like “creativity” and not raw computational horsepower, because I have recently tumbled down the David Deutsch rabbithole.

In The Beginning of Infinity, Deutsch points out that humans have ‘computational universality’: with the help of a pencil and a piece of paper, we can run any program that a general-purpose computer can (i.e. we are Turing-complete). This is not one of those annoying metaphors that compares the brain to whatever technology we’ve invented most recently: it’s based on the fact that all known laws of physics are computable by a series of approximations on a digital computer—something Ada Lovelace discovered in the 19th century.

Of course, your microwave oven is also Turing-complete. What makes humans special is that we are creative, which Deutsch defines as the ability to come up with new explanations.

This is the central pillar of Karl Popper’s epistemology, in which knowledge advances through good explanations, while experiment and empirical observations can only ever help us choose between existing theories (see: the problem of induction).4

Deutsch claims the ability to create explanations is something that you either have or you don’t: once you have it, you are not only Turing-complete but a “universal explainer”, or more simply, a person, regardless of whether you’re instantiated in carbon, silicon, or space alien goop.

This is extremely good news for humans. An AI’s inner machinations might be totally inscrutable to us, but any new theory it brings into the world will necessarily contain at its core a good explanation, which a human (augmented by tools like a pencil or a computer) would theoretically be capable of understanding. To argue otherwise is tantamount to an appeal to the supernatural, as if God had imposed arbitrary limits on what we can know.

Universality flies in the face of our intuitions. Consider that Deutsch is a physicist who founded the field of quantum computing. His insistence that all people with non-damaged brains are, in principle, able to e.g. parse the Schrödinger equation that governs the evolution of the wavefunction in quantum mechanics has always seemed to me if not flat-out wrong, at the very least in massive violation of the evidence right in front of our noses.

Now we come back to the systematising trait, which goes a long way towards resolving my confusion.

There is a binary here, in that a species either develops this cognitive module or it doesn’t. But the fact that it’s polygenic means there is a wide range of systematising ability within the population: some people are very good at doing the kind of thing that leads to inventions and new knowledge, and some people are very bad at it (while still being infinitely better than e.g. apes, who apparently can’t do it at all).

Deutsch would argue that low-SQ people are still capable of solving any given problem if they were sufficiently interested, had unlimited time, excellent tutors, the necessary cultural ‘software’, etc. Instead, we tend to gravitate towards doing things we’re good at: if some people find it especially taxing or boring to think systematically, that could be enough to explain the real-world divergence in problem-solving ability.5


2. Are large language models a path to AGI?

Deutsch thinks AGI is possible, but large language models won’t get us there: they are mind-bogglingly impressive induction machines, but the ‘knowledge’ they produce was in some sense already contained within the dataset, with no ability to make the kind of novel explanatory leaps that e.g. Copernicus took in conjecturing that the Earth rotates around the sun.

As far as I know most AI researchers disagree with this. The meme response is ‘fine, AIs are stochastic parrots, but what makes you think humans aren’t?’. In this view, our ability to model the world, do causal reasoning, and imagine counterfactuals is something that emerges downstream of massive amounts of data (as opposed to starting with the right cognitive module or algorithm, and then feeding data into it).

We could design a test for this by e.g. sandboxing a 2024 model, waiting for humans to make an explanatory leap in some field, and then prodding the sandboxed model to see if it can make the same leap. Infuriatingly, both camps are so confident in their own stance that as far as I can tell no-one is actually interested in trying to operationalise the experiment.

Baron-Cohen’s book helped me realise this is a massive crux. It also gave me another idea for resolving it. If apes and other smart animals really can’t do any causal or explanatory reasoning, then that is strong evidence that Deutsch is in fact right about this: if you don’t start out with nucleotide pairs that code for a specific cognitive process, throwing vast amounts of data and compute at the problem doesn’t help.

If animals can do weak versions of this kind of reasoning, then I would update to thinking Deutsch is wrong: maybe things like causation and counterfactuals really do emerge from running a ton of data through a feedback loop of predictive processing/gradient descent.


3. Will AGI be more creative than lowly humans?

Let’s assume Deutsch is right about universality: if you or I were sufficiently motivated, had the best tutors, and unlimited time, we could make novel contributions in his field of quantum physics.

This might be true in a very narrow technical sense, but in practice, there’s no way it’s going to happen. The gulf between me and Deutsch is already unbridgeable. Now think about the most fuzzy-thinking person you know!

So, what happens if we encounter aliens or superintelligences whose systematising ability towers as far above Deutsch as he towers above a lowly wordcel? They might technically be able to explain things to us, but in practice, we’re obsolete.

This is where Baron-Cohen’s theory gives us some hope (or despair, depending on your point of view).

At the end of the book, Baron-Cohen speculates that the systematising trait might follow something like an inverted U-shaped curve: more is better, but only up to a point.

Inverted U-shape curve for autism/systematising ability

Once you reach the top of the curve, the whole ‘rigidity and repetition’ thing starts to become pathological, until you become the kind of person who is compelled to spend all day arranging toys in rows and counting them over and over. The hyper-focus on one system interferes with learning about the wider world, including very important things like language and speech.

If Deutsch is right about universality, then silicon-based people will not be immune to this problem. But they will very likely have other advantages, like a larger memory and faster processing speeds—which brings us to intelligence.


4. Why do autistic people score worse than average on IQ tests?

Here’s a paradox previously pondered by Scott Alexander: the genes that contribute to autism are associated with higher intelligence, but if you have too many of them, you somehow become less intelligent. Why?

Some cases of autism are caused by straightforwardly bad things that interrupt normal development, e.g. maternal illness or de novo mutations. It isn’t surprising to see harmful effects there. But even cases of autism that are caused entirely by the normal genetic lottery seem to lead to a lower average IQ.

Again, the inverted U-shape of the systematising trait might be the key to solving this puzzle.

Systematising is not the same thing as intelligence, but it must help with some components of IQ tests, up to a certain point at which it impairs performance again.

I couldn’t find data on the exact relationship between the two traits, but if it exists, we can make some predictions about what it should look like:

  • We would expect to see a positive correlation between IQ and SQ that flattens and then becomes negative on the right tail
  • The overall correlation should be zero (otherwise e.g. men would score higher on IQ tests than women do).

If anyone can find this dataset, please share it. If it doesn’t exist, it would make for a great research project.


5. Why is the prevalence of autism increasing?

The conventional answer is vaccines uh, increased awareness leading to more people being diagnosed. This must be true to some extent, but Baron-Cohen offers another intriguing explanation: assortative mating between nerds.

It would be surprising if autism itself was positively-selected for, given the severe comorbidities that come along with it (more on this in the next section). But it’s not at all surprising that systematising is positively-selected for. Nerds really do rule the world, with an unprecedented demand for the skills possessed by high-SQ individuals, and unprecedented mobility allowing those individuals to flock together in hubs like Silicon Valley.

What happens when two hyper-systematisers get together? There is probably some regression to the mean, but they’re each still contributing a mega-dose of genes that are associated with autism.

And so, when Baron-Cohen heard that the children of couples who met at MIT seemed to have 5-10x higher rates of autism than the general population, he saw an opportunity to test the theory. Unfortunately his study was canned by the then-president, Charles Vest, for fear of reputational harm. Boo!

In a stroke of luck, Baron-Cohen was later able to carry out the study in the city of Eindhoven, described as “the Silicon Valley of the Netherlands”. The results were right in line with the experiment’s predictions: there were 229 cases of autism per 10,000 children in Eindhoven, compared to just 84 and 57 in the two control cities.

Compare this against Scott’s attempt to answer the same question based on ACX survey data, in which he tentatively found that assortative mating either doesn’t carry a risk, or the risk is small.

All these data have limitations, but I put more weight on Baron-Cohen’s findings (he has also run the experiment in the opposite direction, by asking what fields the parents of autistic children work in, and found similar results).

My guess is that if you are a very high-SQ person and you marry someone similar, you do in fact have a much higher chance of having an autistic kid: probably at least double the baseline rate. That’s a big relative increase, but in an absolute sense, it only takes you from the baseline 1-2 per cent chance to 2-4 per cent.

That’s all the interesting stuff out of the way. Now we finish up with my only major gripe with the book.


The Great Equivocation

Baron-Cohen argues that the unique human capital tied up in neurodiverse brains represents a huge untapped opportunity, in that the majority of autistic people currently struggle to find jobs and form relationships.

We might fix this by changing some of our societal norms: maybe we stop judging a person’s worth by the firmness of their handshake and ability to make steady eye contact, for example. And in fact this is already happening: there are companies doing affirmative action for neurodiverse hires, and a consulting firm specifically designed for autists.

This seems like a glaring civilisational inadequacy that we should totally fix. Good on Baron-Cohen for championing the cause, which he clearly feels strongly about.

My only complaint is that he is less than fully forthcoming about the difference between high-functioning autists who just need to be given a chance, and the unfortunate people suffering from the kind of severe disability that no amount of societal change is going to help with.

On this you are much better off reading Scott, who gives a laundry list of comorbidities:

Thirty percent of autistic people have comorbid epilepsy, often very severe. Over half of autistics are cognitively disabled. Autistics have three times the risk of Tourette’s Syndrome, five times the risk of cerebral palsy, about a hundred times the risk of tuberous sclerosis, and various balance and coordination disorders, plus an increased rate of other psychiatric disorders like bipolar and schizophrenia. There are treatments for these conditions, both pharmacological and otherwise, but they come with their own set of side-effects and difficulties and none of them are 100% effective. […] Half of autistic children self-injure, and more than half of autistic children and adolescents are physically aggressive. Autistic children are twenty-eight times more likely to be suicidal than other children. Three-quarters have eating problems ranging from “picky eater” to “will not eat food, good luck doing something about this”. About two-thirds have “sleep disorders”, which is sometimes a euphemism for “wakes up screaming in the middle of the night and will not stop”. As best I can tell, all these studies were done on non-institutionalized autistic people who were generally well-treated and still living with their parents.

And his experience working at the coalface of mental health:

A year or so ago, after a particularly bad week when two different nurses had to go to the emergency room, the charge nurse told me in no uncertain terms that the nursing staff was burned out and I was banned from accepting any more autistic patients. This is a nurse who treats homicidal psychopaths and severely psychotic people every day with a smile on her face. When she says “autistic”, it seems worlds apart from the “autistic” that means “good at math and makes cute hand flap motions”. When a mental health professional says “autistic”, the image that comes to mind is someone restrained in a hospital bed, screaming.

Needless to say these people do not feature in any of the book’s inspiring case studies.

Baron-Cohen mentions some of the comorbidities, and says autistic people should demand treatment for any unwanted symptoms. But he says these are not “core” features of autism because they do not occur universally, and do not “characterise the autistic mind”. This is such a weird argument that I have to assume that either his advocacy has muddied his thinking, or he’s fudging his true beliefs for reasons of political expediency. In either case, it feels like he’s bending over backwards to avoid any implication that autism might be a net negative.

And so Baron-Cohen’s great idea—that there is a genetic trait that is both the cause of all of our success as a species, and that is strongly linked to a disorder that causes great suffering—is marred by the tendency to equivocate between the two groups, including in the very subtitle of the book: “How Autism Drives Human Invention”.

Is there something beneficial about having autism, as distinct from the folk definition, or merely being a high-systematiser? I read the book twice and I still couldn’t find a clear answer. It seems plausible to me that we could benefit from the ‘outsider view’ of an autistic anthropologist, or the kind of art made by someone who is naturally immune to social contagion, but these are my own clumsy speculations: Baron-Cohen doesn’t actually make the case!

The answer to this question really matters. For every Temple Grandin giving inspirational TED talks about the benefits of neurodiversity, there is some larger number of people compulsively slamming their heads into a wall. The answer to the question might change how we allocate resources to mental illness. At the margin it might even change some people’s choice of partner.

But most of all it matters because—inshallah—we’re going to get widespread polygenic embryo screening in the near future.

Would ‘curing’ autism ironically doom human civilisation to never make progress again? I am much more confident in answering this weaker version of the question: no.

But still. It would be really great to get some more clarity on this. If we are lucky, then perhaps an autist will examine it with the rigour it deserves.


Thanks to Cam, Benny, Phoebe, and Sonnie for giving feedback on this review. Special thanks to Sara for recommending I read this book.


Notes:

  1. Baron-Cohen makes a distinction between cognitive empathy, which is the classic theory-of-mind stuff, and affective empathy, which is the drive to respond to someone with an appropriate emotion. In this sense autists and sociopaths are inverted: autists tend to be bad at navigating social reality, but still care about other people’s feelings, whereas sociopaths are unable to sympathise, but don’t have problems modelling other minds.
  2. Every time I googled someone who I think fits this mould (Mozart, Aphex Twin, Stanley Kubrick, Ingmar Bergman), the Internet said they were on the spectrum. I don’t put much stock in this kind of ‘diagnosis’ but it sure is suggestive.
  3. Another eleventh-hour addition to my reading list is Robin Hanson’s glowing endorsement of Cecilia Heyes’ Cognitive Gadgets: The Cultural Evolution of Thinking, which seems to argue that causal reasoning can arise from associative learning in combination with sociality.
  4. In my experience the problem of induction is hard to fully internalise. Deutsch’s book gives the beginner-friendly version. For a shorter, more technical explanation, machine learning PhD Vaden Masrani asks “did Popper disprove machine learning, or does machine learning disprove Popper?”.
  5. We might reconcile universality and the IQ literature in a similar way. I have some ideas about how to do this but it deserves a dedicated post.

Not Sure What the Future Holds? Get Your Copy of Optionality Now.

Optionality Book available now
Subscribe
Notify of
guest

3 Comments
Inline Feedbacks
View all comments
Nightingale
Nightingale
17 days ago

The article’s exploration of the intersection between autism and human creativity presents a thought-provoking perspective. However, the discussion on current AI technologies, particularly large language models (LLMs), could benefit from a more nuanced examination.

First and foremost, it’s important to clarify that achieving Artificial General Intelligence (AGI) isn’t contingent solely on LLMs. AI encompasses a diverse range of architectures, each with the potential to lead to AGI. Narrowing the focus solely on LLMs restricts the broader conversation and overlooks other promising pathways.

The article defines creativity primarily in terms of generating novel explanations. However, creativity also includes devising new, unseen solutions to presented problems—a domain where AI has already demonstrated significant promise. Consider AlphaGo’s famous match against Lee Sedol in 2016; AlphaGo’s unexpected and counterintuitive moves surprised even the most seasoned Go experts, exemplifying a form of computational creativity. This rapid advancement in AI capabilities caught the Go community off guard. Top players like Lee Sedol, who had mastered Go through extensive study and practice, faced an unforeseen and unbeatable opponent.

Following this, AlphaZero further amplified this creativity by training solely through self-play, generating strategies devoid of human data input. MuZero, AlphaZero’s successor, advanced even further by mastering a diverse range of games, including Go, chess, and Atari, without explicit knowledge of their rules. This ability to strategize and plan in unfamiliar environments signals a creative problem-solving capability that transcends mere pattern recognition.

To address the current limitations of LLMs, it’s helpful to utilize the psychological framework of System 1 and System 2 thinking. System 1 (“fast thinking”), akin to current LLM models, operates through quick, automatic, and subconscious decision-making. System 2 (“slow thinking”), by contrast, involves deliberate, slow, and conscious reasoning—a cognitive flexibility humans possess but LLMs currently lack. While humans can adjust cognitive efforts based on problem complexity, LLMs maintain a constant computational load regardless of the task’s intricacy. System 2 thinking, which somewhat mirrors the “systematizing trait” discussed in the article, is almost absent in current LLMs and this is a known weakness. However, researchers are actively pursuing solutions to bridge this gap.

Current AI systems, particularly those employing reinforcement learning, already demonstrate creativity within narrow domains. Their achievements should not be dismissed as mere statistical parroting. These systems embody the potential for innovative, original thought processes, which are pivotal stepping stones toward more advanced forms of intelligence.

The article rightly encourages critical scrutiny of AI’s current limitations and potentials. Nonetheless, underestimating AI’s capacity for creativity and novel problem-solving overlooks the considerable advancements already made. While it may be comforting to believe that powerful AIs are a distant prospect, ignoring the possibility of their imminent arrival could prove hazardous, especially if it leads to neglecting necessary safety measures.

Nightingale
Nightingale
14 days ago

> “Was this written with the help of an LLM?”

Yes. I wrote a rought draft with points, which I wanted to say, then let GPT4 to “improve” the text. It was a kind of experiment, whenever the LLM can be helpful and indeed I think, the result is better, than I would write without the “tool”. I did this in few iterations. But it also has drawbacks, as some of my words and expression vanished, and the tone of the message changed and it sound more formal and less casual, as I would write it. As you said, the style feels a bit awkward.

I also tried to ask the model, whenever it can brainstorm more new arguments, in the direction where I wanted to go, but this didn’t bring anything new.

> “I’m guessing you heard Francois Challet on Dwarkesh podcast?”

Not yet, but I’m listening to it now.

I’ll write more later.