The first section is well-written and reliable, because I didn't write it, it's a book excerpt. The rest of the post is more prone to errors and speculations; you can skip it if you're not into that kind of stuff—I read a few books and articles about autism, but I'm definitely not an expert, and I didn't spend a lot of time checking everything I wrote. Please comment if anything seems wrong (or dumb, or offensive, etc.) so I can fix it.

Summary / Table of Contents

The first section is mostly a book excerpt introducing my favorite theory of how autism works in the brain, namely "Intense World Theory".
Then, I'll flesh out how I'm thinking about that theory, by distinguishing "intense world" and "different learning algorithm hyperparameters" as things that happen in different parts of the brain and have different consequences, even if they tend to go together. I'll talk a bit more about each separately, and try to relate them more specifically and mechanistically to the algorithms going on in different parts of the brain.
- I'll also include how I think this theory connects to other famous aspects/theories of autism, like cerebellar abnormalities, memory, "weak central coherence", etc.
At the end, I'll speculate that there's an exact-opposite "dim world theory of psychopathy". Fun!

Introduction to the Intense World Theory of Autism

There are various theories of the root cause of autism in the brain. I'll mention a few as I go. But I'll start right in with the positive case for the theory I like: The Intense World Theory of Autism.

The Intense World Theory of Autism dates to this 2007 article. I first heard about it from the excellent book The Myth of Mirror Neurons, which devotes a whole chapter to it, excerpted below. Then I read Temple Grandin's The Autistic Brain: Thinking Across the Spectrum and found that she also brought it up, and seemed very enthusiastic about it. I've also read a couple review articles on Intense World Theory, including this one by some of the inventors of the theory, and this one and a few others. Everything I've found is pretty positive the theory—please comment if you've seen any substantive criticism.

I can't dream of describing the theory as well as Gregory Hickok does in The Myth of Mirror Neurons, so I'll just excerpt from that. First some motivation:

…I'm going to suggest the possibility that the dominant neurocognitive theories of autism, which assume that behavioral deficits result from lack of or diminished social sensitivity, have it wrong and in fact have it backward.
"Deficit theories" of dysfunction are reasonable and intuitive. If an individual fails to respond normally to sound, it’s a good bet that the person has a diminished capacity to process and hear sound. He simply isn’t capable of perceiving the signal. Likewise, if another individual fails to respond normally to social stimulation, it’s a reasonable bet that the person has a diminished capacity to process social information. But consider the following thought experiment. Imagine you had a stadium rock concert–type sound system hooked up to your living room television and you attempted to watch the evening news with the sound cranked up all the way. Most likely, you would cover your ears and quickly leave. If you forced yourself to stay, you would run into at least one of three problems as you tried to listen and watch. One, the physical pain would be so extreme that you wouldn’t be able to concentrate on the message. Two, attempts to dampen the sound and ease the pain, say by sticking your fingers in your ears, would filter out many of the fine details you need to hear normally. You would perceive less well. Three, if you did manage to listen, the extreme volume would excite so many nerve fibers that it would drown out the details of the signal itself and again you would miss many things. Excess can be as detrimental to normal function as paucity.…
Behavior does not automatically reveal its cause and can be misleading….

Then he switches to rats:

In fact, Henry Markram, Tania Rinaldi, and Kamila Markram proposed such a theory in a 2007 article aptly titled, "The Intense World Syndrome—An Alternative Hypothesis for Autism." The theory is grounded, oddly enough, in a rat model of autism. I say "oddly enough" because autism has traditionally been considered a uniquely human disorder, with defining symptoms showing up in high-level social and language domains. But it turns out that rats who are exposed prenatally to valproic acid—a compound used in human medications to control seizures and bipolar disorders—develop some key features of autism both neurally and behaviorally, including loss of cerebellar neurons, abnormalities in the serotonergic system, decreased social interactions, increased repetitive behaviors, enhanced anxiety, motor abnormalities, and sensory hypersensitivity. Curiously, the prevalence of autism in humans who are prenatally exposed to valproic acid through maternal use of the medication is substantially higher (one estimate is 11–100 times) than in the general population…
The existence of a rat model makes it possible to explore, in substantial detail, the neural bases of an autistic-like phenotype. The Markrams and their team did precisely that.…
They determined that local neuronal networks in the three brain regions tested in rats are hyperreactive….
The driving force behind the micro network-level hyperreactivity appears to be the direct connections between neurons. Valproic acid-treated animals exhibit a 50 percent increase in the number of direct connections between neurons within a local circuit….
Further, the Markrams and their team found that neural networks in valproic acid–treated rats are also hyperplastic... When connectivity patterns between neurons in local and expanded networks were examined before and after widespread and prolonged (overnight) activation of the entire network, valproic acid–treated rats exhibited an increase in the rate of rewiring, mostly evident in the nonlocal networks, compared to controls…

And back to humans:

Given these kinds of neural changes in the rat model of autism, a plausible story can be told about the neural basis of the range of autistic behaviors. Hyperabilities, such as increased sensory sensitivity or memory in specialized domains, can be explained by hyperreactivity and hyperplasticity of neural circuits. Hyporesponse to social stimuli can be explained in terms of the emotional intensity of the signal, which triggers anxiety and avoidance responses, which means less information is acquired in individual social situations and over time reduced opportunities to learn in the social domain. Theory of mind performance would also be expected to suffer with a hyperactive response to social signals—even if there is no fundamental deficit in mentalizing—because of increased anxiety when interacting with others and/or because avoidance behavior decreases the amount of information perceived or learned. Repetitive behaviors can be viewed as a coping mechanism aimed at regulating the child’s intense world. Motor deficits can be explained by hyperexcitability of the response to sensory stimulation, which has motor consequences, as we’ve discussed, or from hyperreactivity of motor systems themselves. And because language is at some levels a sensorimotor task and at other levels a highly social behavior, abnormalities in the sensorimotor or social domains can be expected to affect language.
All of this is interesting—suggestive even—but how relevant is it, really, to autism?…
Hyper-responsivity leading to avoidance…is observed regularly and uncontroversially in the sensory domain. Autistic individuals often cover their ears when even moderately loud sounds are present in the environment and exhibit other forms of avoidance behavior….
At least one study has confirmed that alternative explanations of the face processing "dysfunctions" in autism may be on the right track. Autistic and nonautistic individuals were scanned using fMRI while they looked at pictures of faces that were either emotionally neutral or emotionally charged. Crucially, using eye-tracking technology, the researchers also monitored which parts of the images their participants were looking at during the experiment. Overall, autistic participants activated their fusiform face region less vigorously than nonautistic controls, replicating previous work. But the eye-tracking data showed that this was simply because they spent less time looking at the most informative region of the faces, the eyes. In fact, when the researchers looked at fusiform activation as a function of time spent fixating on the eyes in the photos, they found a strong positive correlation in the autistic group. This means that the autistic brain is responding quite well to face stimuli, if one takes into account the amount of time spent looking at them...the same study reported that amygdala activation was stronger in the autistic compared to the nonautistic group while looking at faces.
Also consistent with the alternative, emotional hyperreactivity hypothesis are statements from autistic individuals themselves. Here’s a sample gleaned from a paper covering face processing in autism:
It’s painful for me to look at other people’s faces. Other people’s eyes and mouths are especially hard for me to look at. My lack of eye contact sometimes makes people, especially my teachers and professors, think that I’m not paying attention to them. —Matthew Ward, student, University of Wisconsin
Eyes are very intense and show emotions. It can feel creepy to be searched with the eyes. Some autistic people don’t even look at the eyes of actors or news reporters on television.—Jasmine Lee O'Neill, Author
For all my life, my brothers and everyone up 'til very recently, have been trying to make me look at them straight in the face. And that is about the hardest thing that I, as an autistic person, can do, because it’s like hypnosis. And you’re looking at each other square in the eye, and it’s very draining. —Lars Perner, Professor, San Diego State University
These are revealing statements for two reasons. First, they provide a clear indication of an intact theory of mind in these individuals ("my lack of eye contact…makes people…think that…"). And second, active avoidance of eye contact provides just as much evidence for sensitivity to the information contained therein as does active engagement of eye contact. If you can’t recognize that there is information in the eyes, why avoid them?
…

There's a lot more in the book chapter, but I'm probably already pushing the limits of copyright law. I'll leave it at that. Again, this is from the book Myth of Mirror Neurons, which I heartily endorse.

Honestly, I should probably just end the blog post here. But I couldn't resist adding some more commentary and speculation, particularly trying to shift the level of discussion towards lower-level neuroscience and algorithms.

Deeper dive part 1: "Intense world"

Sensory sensitivity

Here's a chart:

Unexpected sound stimulus	Neurotypical	Autistic
Very quiet chirp	Not noticed	Bit of a startle
Moderately loud bang	Bit of a startle	Jump-out-of-your-seat surprise!
Firecracker next to your ear	Jump-out-of-your-seat surprise!	Aaaaaaaaaa!

I could make a similar chart for vision, touch, or any other sense. Scott Alexander brings up here the example of scratchy tags on the back of shirts, which might repeatedly draw the attention of a person with autism, yet be below the threshold where they draw the attention of most neurotypical people.

What's going on in the brain here?

The starting point has to be innate, not learned. We're talking here about startle reflexes, orienting reactions (e.g. where you turn your head and body towards a sound), flinching, releasing cortisol, etc. Nobody teaches you how to do those things! Moreover, there's nothing in your sensory input data that says that an unexpected 27 dB sound is sufficiently threatening to warrant startle and cortisol, but an unexpected 19 dB sound is not. This is a heuristic threshold set by the genome. No wonder it's different in different people! (It could also be modified over time by experience, I'm not sure. But anyway, it has to start somewhere.)

I believe that the brainstem (specifically, superior colliculus & inferior colliculus) is where you'll find these innate circuits—circuits that take a sound (or flash or touch etc.), figure out where it's coming from, and if it's sufficiently loud, execute startle reactions and orienting reactions and so on.

But that doesn't mean that these brainstem circuits are necessarily the underlying cause of sensory sensitivity in any particular person. I think other parts of the brain (amygdala, medial prefrontal cortex, cerebellum) function sorta like "eager servants" of the brainstem in this task—they learn to anticipate when the brainstem is going to execute a startle, and they jump the gun, issuing those commands themselves. So if any of those systems are really trigger-happy—i.e. they issue a command in any circumstance where they think there's even the slightest chance that the brainstem would issue that command—you could get the same kind of sensory sensitivity, seems to me.

By the way, this is how I'm thinking about the (apparently) causal role of the cerebellum in autism.

Eye contact

I think eye contact is basically the same as sensory sensitivity. I believe the brainstem (superior colliculus) detects eye contact, and issues (among other things) arousal reactions ("arousal" in the psychology jargon sense, not the sexual sense—e.g. cortisol release, higher heart rate, etc.). There seems to be an arousal sweet spot: too little arousal is boring, too much arousal is overwhelming. (There are many more dimensions to your feelings than just arousal, of course! But it does seem to me that arousal is of central importance.) As in the book excerpt at the top, people with autism apparently often find that eye contact flies way past the sweet spot into "overwhelming" territory. It's aversive, and therefore they avoid it.

This is my model, which I'll be using throughout this post. I'm not proud of it; it's a lousy model. Valence in fact depends on much more than just arousal; they can vary independently. But I'm using this model anyway, because I think it captures some important kernel of truth, and I don't know any better way to express that kernel of truth.

As in the sensory sensitivity section, the higher-than-typical arousal could come from several other sources besides the brainstem itself—particularly the amygdala, medial prefrontal cortex, and cerebellum.

Social interactions more generally

I think eye-contact-detection is just one example: my hunch is that the brainstem (again, the superior & inferior colliculus) has a whole suite of heuristics for a person being near to and interacting with other people—for example, human-speech-sound-detection, and human-touch-detection, human-smell-detection, and so on. All of them create arousal, I assume, and therefore all of them might potentially create an overwhelming level of arousal in people with autism, who may correspondingly try to avoid those things.

Next up is empathetic simulation. As I discussed (speculatively) here, I think the main neurotypical way of understanding and interacting with other people is by empathetically simulating them, constantly and without any deliberate effort. (I call this "little glimpses of empathy"—to be strongly distinguished with the deliberate, slow, effortful "empathy" that people are usually talking about when they say "empathy" in everyday contexts.)

This kind of empathetic simulation is tied up with social instincts, and their corresponding arousal, which as above may be overwhelming and aversive for some people with autism.

I figure what happens as a consequence, at least sometimes, is: the person with autism gradually learns to interact with and understand people, but without using empathetic simulation. Instead, they just take their general intelligence, and leverage it to build a new human model from the ground up, just as people can model any complicated system from the ground up (e.g. a complicated piece of software). So there's still a human model, but it's built on a different foundation—a foundation without that strong innate connection to brainstem circuits.

I think this explains why (IIUC) there are intelligent people with autism who are able to do theory-of-mind-type reasoning (as in the book excerpt at the top), and able to understand social interactions and conventions, but need to deliberately learn aspects of these things that neurotypical people might find intuitive and effortless. The intuitive-and-effortless part is the pathway that goes

quick effortless empathetic simulation
→ social instincts hardwired into the brainstem
→ resulting "feeling"

…again as discussed here. (Further clarification in this comment.)

Deeper dive part 2: "Different learning algorithm hyperparameters"

I believe that a very large fraction of the human brain (96% by volume) exists for the purpose of within-lifetime learning (see discussion of “learning-from-scratch” here). They thus house learning algorithms, or more specifically “learning-and-inference algorithms”. (The "learning" part is editing synaptic connections for future use; the "inference" part is using previously-learned content for taking better actions now.)

These learning-and-inference algorithms, like any learning-and-inference algorithms, have what we call "hyperparameters" in machine learning.

Here's an example: If you've seen a pattern "A then B then C" recur 10 times in a row, you will start unconsciously expecting AB to be followed by C. But "should" you expect AB to be followed by C after seeing ABC only 2 times? Or what if you've seen the pattern ABC recur 72 times in a row, but then saw AB(not C) twice? What "should" a learning algorithm expect in those cases?

You can imagine a continuous family of learning algorithms, that operate on the same underlying principles, but have different "settings" for deciding the answer to these types of questions.

And I emphasize that this is one of many examples. "How long should the algorithm hold onto memories (other things equal)?" "How similar do two situations need to be before you reason about one by analogizing to the other?" "How much learning model capacity is allocated to each incoming signal line from the retina?" Etc. etc.

In all these cases, there is no "right" answer to the hyperparameter settings. It depends on the domain—how regular vs random are the environmental patterns you're learning? How stable are they over time? How serious are the consequences of false positives vs false negatives in different situations?

There may be an "optimal" set of hyperparameters from the perspective of "highest inclusive genetic fitness in such-and-such specific biological niche". But there is a very wide range of hyperparameters which "work", in the sense that the algorithm does in fact learn things. Different hyperparameter settings would navigate the tradeoffs discussed above—one setting is better at remembering details, another is better at generalizing, another avoids overconfidence in novel situations, another minimizes energy consumption, etc. etc.

I actually already talked about learning-and-inference algorithm hyperparameters in the "intense world" discussion of the previous section—but without using those words. Specifically I said that the brainstem has innate capability to do startle reactions, but the amygdala, medial prefrontal cortex, and cerebellum all (in different ways) learn to anticipate the brainstem. And this algorithm has a tradeoff between false positives and false negatives. And this tradeoff depends on—you guessed it—hyperparameters. "Intense world" would come from the algorithms being tuned to have almost no false negatives, at the expense of lots of false positives.

In this section, I'm talking about learning-and-inference algorithm hyperparameters more generally, focusing more on memory and cognition.

I can think of two reasons that an autistic brain might wind up with systematically different learning-and-inference algorithm hyperparameters (for memory and cognition) than a neurotypical brain:

Maybe, like in the valproic acid rats above, there is an underlying biophysical cause of "hyperreactive and hyperplastic" neurons. When this root cause impacts the neurons in certain parts of the brainstem / amygdala / cerebellum / whatever, it manifests as the "intense world" stuff above. And when this same root cause impacts the neurons in, say, the fusiform face area, it manifests as "different hyperparameters" in regards to learning and recognizing faces. And so on with other learning areas and modalities.
The brain has a capability of doing real-time hyperparameter variation, in a situation-dependent way, using various neurotransmitters like dopamine, serotonin, norepinephrine, and so on. And there seems to be at least two signals of this type (namely acetylcholine—see here—and serotonin) that more-or-less links arousal level to hyperparameters, I think. So if the "intense world" stuff of the previous section leads to high arousal all the time, that would in turn affect the learning algorithm hyperparameters. (I got that idea from Steve Grossberg's book.)

Finally, here are a couple typical aspects of autism that seem to be in the "different hyperparameters for memory and cognition" category:

Autism and memory: Given the same information, I think there are systematic (though not universal) differences between what the memories formed by neurotypical people and people with autism. This can lead the latter to have both "deficits" and "savant-like" memory abilities—even in the same person.
Weak central coherence theory: According to this review, it seems likely (though not definite) that the valproic acid rats have a higher ratio of local-connections-vs-long-range connections in the neocortex. (This comes from an increase in local connections, not a decrease in long-range connections.) This seems at least vaguely to go along with the idea that their neocortex would learn and remembering more narrow and specific aspects of things, and fewer cross-domain, analogizing, big-picture aspects of things, e.g. aspects that incorporate multiple senses and abstract context.

Bonus: "Dim world theory of psychopathy"??

To be clear, this part is my own wild speculation. Call it the "dim world theory of psychopathy". Psychopathy, in this theory, is when the brain is hyporeactive and hypoplastic, as opposed to autism where it's hyperreactive and hyperplastic.

When people with autism activate their empathy brain circuits, it reacts so overwhelmingly that they learn early in childhood to avoid activating those circuits in the first place, using techniques like avoiding eye contact, avoiding the use of empathetic simulation as a method of understanding people, and so on.

Conversely, when people with psychopathy activate their empathy brain circuits, there's barely a whisper of activity. Why do psychopathic kids stereotypically torture animals? Because for them to feel anything at all, it takes a situation that would be emotionally overwhelming for neurotypical people.

Situation	Autistic	Neurotypical	Psychopathic
I'm sitting in a quiet empty room, moving my toes	Stimulating 😀	Boring 🙄	Super boring 😡
I'm talking to someone, while making eye contact and empathetically simulating them.	Overwhelming 😞	Stimulating 😀	Boring 🙄
I'm lying to someone's face as part of a deliberate scheme to make them miserable	Super overwhelming 😱	Overwhelming 😞	Stimulating 😀

I did a little search for studies about sensory sensitivity etc. in psychopaths, and didn't have much luck finding anything very informative. I didn't try very hard, maybe it's out there. (If anyone follows up, let me know what you find. (Wanna coauthor a follow-up post with me? :) ) For now, take this as casual speculation by a non-expert, as if we were chatting over drinks.

(Update 2021: here's a possible lead—this wikipedia article talks a bit about "low arousal theory" and psychopathy. The references they provide for that are a bit underwhelming though.)

(Update 2024: (1) I currently think that this section is applicable to a large subset of ASPD people, maybe the majority, including what I personally think of as the central stereotypical cases … but that there are also some people who get an ASPD diagnosis for a quite different underlying reason, namely being quick to anger. (2) I came across another potential lead for looking into “low arousal theory” here: “The best physiological indicator of which young people will become violent criminals as adults is a low resting heart rate, says Adrian Raine of the University of Pennsylvania. … Indeed, when Daniel Waschbusch, a clinical psychologist at Penn State Hershey Medical Center, gave the most severely callous and unemotional children he worked with a stimulative medication, their behavior improved”.)

Hang on, isn't schizophrenia the opposite of autism?

(Update 2024: This whole subsection is bad and should be ignored; my up-to-date opinions about schizophrenia are at Schizophrenia as a deficiency in long-range cortex-to-cortex communication (Feb 2023) and Model of psychosis, take 2 (Aug 2023). I think that in the many “diametric” cases where schizophrenia-associated gene variants are protective against autism and vice-versa, those genes are generally reducing or increasing (respectively) the quantity and strength of cortex-to-cortex communication. Other genes may be more relevant to one or the other condition, or may increase or decrease the risk of both together, in cases where the gene variant has different effects on long-range cortex-to-cortex communication (more relevant to schizophrenia) versus local connections (more relevant to autism), or if they lead to a more or less “intense world” via their effects on some other system besides the cortex (e.g. the amygdala or brainstem), etc.)

I guess there's a "diametrical" theory where autism and schizophrenia are opposites. I dunno, I know very little about schizophrenia. Here's a random thought though.

People with autism and people with psychopathy are both basically not feeling much effect of their innate social instincts most of the time as they go about their lives. It's for totally opposite reasons! And with radically different consequences! But still, this is a narrow respect in which autism and psychopathy are not opposites. So then we can talk about a different "axis" along which schizophrenia might be the opposite of autism. Maybe people with schizophrenia have their innate social instincts very powerfully active all the time, for whatever reason. Maybe those circuits just randomly trigger all the time for no reason (as opposed to the neurotypical case when they trigger at specific times like eye contact, body contact, empathetic simulation, etc.).

By the way, I don't expect that this is a complete description of schizophrenia; I would assume that (like autism) schizophrenia involves some neuron-related thing that impacts different brain systems in different ways.

My model seems to predict that it should be possible to be both autistic and schizophrenic, whereas it shouldn't be possible to be psychopathic and autistic. As far as I can tell, both those are basically right. Simon Baron-Cohen does say here that he treated a person with both autism and borderline personality disorder (BPD) (if memory serves). BPD is a bit like psychopathy, in that both involve a lack of empathy. Narcissism is in this category too. But BPD is clearly different from psychopathy. By the way, don't ask me for a theory of BPD or narcissism—I have no idea.