TL;DR: The early development of the brain, from molecular biology via dynamical systems to the theory of computation. Some speculation about how everything fits together.


There is a story about John Nash that I like a lot. John Nash is a famous mathematician who got a (fake) Nobel prize for his work in game theory as well as a Hollywood adaptation starring Russell Crowe. But (if you trust his biography) his contribution to mathematics is at least as great. He had a very curious way of working:

Nash's main mode of picking up information he deemed necessary consisted of quizzing various faculty members and fellow students. [...] Some of his best ideas came "from things learned only halfway, sometimes even wrongly, and trying to reconstruct them - even if he could not do so completely. (A Beautiful Mind, Sylvia Nasar, pg 68)

He started out with something very wrong, then went with his first (wrong) attempt at a solution to one of his colleagues, who helped make it slightly less wrong. Nash then repeated this procedure until, in the end, he had a working proof.

Apart from some hard questions about ownership (can Nash really claim that he produced the proof?) and the obvious issue that Nash turned very schizophrenic in his later years, I admire this way of working (at least for doing very hard research) and aspire to do the same. I really really want to get answers - and I don't care if I have to recruit help and produce a lot of nonsense on the way. This is, of course, just a poor fig leave of a defense for all the nonsense I am producing on this platform. Mea culpa.1

Several perspectives on development

After building a digital person, writing love letters to albatrosses, philosophizing about scale in academia, and asking large language models pointed questions, it is time to write about something where I actually have institutional credentials, i.e. neuroscience and the development of the brain. Why is this an interesting topic?2 While traditionally brain development is a very biology-heavy topic, there are some beautiful results and ideas that generalize far beyond biology and become relevant for people working on artificial intelligence or computations in general. And those results are not widely known, especially not outside a small community in (computational) neuroscience. This essay is written with this in mind and I have tried to reduce jargon to a minimum.

The biology info in this post is focused on small mammals (mouse pups), which might feel like a limitation3. But then again, mice can do a lot more than you’d think. Exhibit 1 - skateboard mouse!

This is a lot better than I do on a skateboard. I guess it helps to have all four limbs so close to the floor. source

A “pretty pictures” perspective

I would not call myself a biologist by any stretch of the imagination - but I can get very excited about videos. Seeing biology in action is very cool.

Embryonic Day 1: Here is a recording of a mouse embryo at the "two-cell stage".

The cell surface marker in green was only imaged for the final division from 32 to ~64 cells. It doesn’t just pop into existence. Strnad et al Nature Methods 2016

Embryonic Day 6-8: From there, things get out of hand quickly. Cells just keep on dividing.

The green-red color indicates where new cells are born. McDole et al Cell 2018

It's not super easy to see, but inside that ball of cells is a sheet of cells that folds into a tube (appropriately named the neural tube) that will eventually develop into the central nervous system (i.e. brain and spinal cord).

This clip is supposed to show the folding of the neural tube, but it's still kind of hard to see.

Embryonic Day 11: And then we're basically there. Just add a few tiny paws, mouth, eyes, tail, all the good stuff.

Note the number at the bottom that increases - that's the age of the embryo in days. Wong et al Development 2015

Embryonic day 20/Postnatal day 1: The mouse pup is born! But at that point, the pup is still “extremely-not-ready” for the world. "Day 1" here is the first day after birth.

Check out the full video on YouTube, it's worth it!

Postnatal day 10: While there is a lot happening "on the inside", not so much is showing on the outside. We still only have a pile of adorable mouse pups.

(The big hairy ones are actually female adult mothers, called "dams").

Postnatal day 25: But fast-forward 15 more days and it's like day and night - the mice (not pups anymore) are basically ready to go! Around postnatal day 14, they open their eyes, grow hair, and start running around.

There is still a lot more that happens after this stage (the brain keeps changing throughout the entire lifetime), but these first 2-3 weeks set up a lot of the central nervous system.

It is great how visual this part of biology is. I hope you found those videos as enjoyable as I did. It's very important to get a feel for what you are talking about, otherwise, you might miss what's in plain sight. In the end, all models and theories cash out to a handful of cells or a pile of mouse pups. Everything is very messy.

A molecular biology perspective

Okay, that’s what development looks like. How does it work? Let's zoom in a bit further. Those first cells in the clip at the top contain the entire genetic code. Can we understand that? How does it instruct the cells to divide like that and to form that ball and the pile of pups and do all the other things? Those are reasonable questions, right?

Well, take this as an answer:


The caption of this image contains the encouraging phrase "Interactions among these pathways are multiple and complex." This is also just the first example taken from Google, there are many many more diagrams like this and they all draw a lot of arrows in a lot of different directions. source

That's a lot to take in and if you make the mistake of asking an expert about it, you will learn 1000 interesting facts about "Bmi inhibition of p19" before lunchtime. Luckily I am not an expert on molecular biology, so I will give you just the most high-level explanation of how development happens on the molecular level:

The behavior of a cell is determined by the proteins it expresses. The "central dogma" of molecular biology says that "DNA makes RNA, and RNA makes protein", so proteins come from the genes. Since we want a lot of different cells that do different things, we need cells with different proteins. "Signaling pathways" (like Sonic Hedgehog, Wnt, ErbB, IGF-IR, and Notch) can tell the cell which genes to express and, therefore, which protein to express. How does the brain know which signals to send to which cell? It just makes sure that different signalling molecules end up in different parts of the developing brain.

All of this is, like, really complicated, but we4 are making good headway in improving our understanding. And perhaps there is hope that things will get easier to understand when we know more. Perhaps everything will just beautifully fall into place. But also perhaps not:

There is no rule that says nature can not be more complex than our brains can handle. And after billions of years of evolution, why should it be simple?

Perhaps there are no shortcuts in explaining how the brain develops. Perhaps there are no simple sub-modules that can be carved out, studied in isolation, and reduced. That would be kind of annoying.

A computational perspective

But explanations in terms of genes feel very unsatisfactory anyway. If I ask "How does a computer work?" it is not technically wrong to give me an explanation of the chip-making procedure.5 But what I actually want to hear is an explanation of boolean circuits and perhaps a bit on the theory of computation. It's not actually very useful to understand exactly how a microprocessor works at the transistor level - but the theory of computation has legs in terms of what it can explain. So are there statements about brain development that are flavored a bit more like the theory of computation?

Despite the (somewhat fatalistic) arguments about the possible futility of understanding biology, there are also good reasons to believe that we can get nice & simple answers. The implicit assumption that runs through the "gene-centered" perspective outlined in the previous section is that "the genes execute a (complicated) program that eventually produces a functional body+brain. To understand that process, we have to completely reverse-engineer the genetic program.” But as we have seen in the videos, a newly born mouse experiences a lot of development after birth, at which point the animal is already exposed to its environment.

How exactly does that make it better? Doesn't that just make it... even more complicated? Indeed, the environment can be very hard to predict and quite perturbative: predators, conspecifics, seasons, losing an eye or a limb, the list goes on and on. It would be very hard to come up with individually tailored solutions to compensate for all of the possible perturbations. And weirdly enough that's a silver lining! In these situations, where the system is constantly perturbed and has to solve a lot of different problems, general solutions emerge (often). These general solutions tend to be not too complicated and they tend to work well even outside the training domain.6 And they are one thing above all: robust to perturbations.7

A dynamical system perspective

How do you get robustness? There is a standard answer from control theory that generalizes: Tightly coupled feedback loops. And indeed there is a very natural candidate for implementing such a loop: the reciprocal interaction of developing brain circuits and the activity they produce.

The little red triangles and the blue dots are neurons, of course.

The circuit is the network of neurons that emerges over development. And these neurons spontaneously produce action potentials and excite/inhibit each other8. The developing brain is highly active9!

Here is what the activity looks like in the developing retina, with cells that express a specific marker that produces light when the neurons are active:

The red color indicates "activity", i.e. those are cells in the retina that are firing. In this clip, you cannot actually see individual cell bodies, only the simultaneous activity of many hundreds/thousands of neurons. source

These patterns of activity are called "retinal waves" and they propagate from the retina to a central relay station and then finally into the visual cortex10:

The red color again indicates activity. Here the experimentalists managed to record activity from a central relay (SC = superior colliculus) and the visual cortex (V1 = primary visual cortex) simultaneously, so you can actually see how some of the activity in the bottom right translates into activity in the top left. Note that while most of the activity in the top left appears to be driven by the bottom right, there is also some “additional” activity (the very large event at the end) that doesn’t appear to come from the SC. source

But spontaneous activity is only half of the story:

The other half comes from the fact that the circuit organizes itself depending on the activity it experiences. In machine learning, this is the "gradient descent" that wiggles the parameters of the network to improve performance. In neuroscience, it might also be gradient descent, but it can also mean the formation or removal of connections.

The yellow dot marks where they are shooting a laser at the neuron to stimulate it. A few seconds later the neuron reaches out to form a new connection.

Exploring these rules for reshaping the brain is perhaps the most exciting thing we do at the lab where I do my Ph.D. The most famous principle is probably the Hebbian postulate (“Cells that wire together, wire together”), but there is a lot more nuance to be explored. Changes in the circuit depend on the exact timing of neural activity (on the scale of milliseconds), the type of cell experiencing the change, the location of the connection on the neuron, and is modulated depending on the context. But the big picture that emerges is that these changes coordinate to increase the computational capabilities, stability, and flexibility of the circuit.

Not only is the interaction between activity and plasticity goal-directed in the sense of making the circuit more capable. It is also very robust: f.e. if the animal loses an eye and thus receives less input to the brain, the brain will increase the amount of input from other sources until it again reaches the "correct" level. But what is the "correct" level? What is the goal of development?

Development might be establishing a prior for predictive processing

A particularly powerful theory of how the mind works is the “predictive processing” theory.11 In this theory, the brain is constantly generating and updating a mental model of the environment. The model is used to generate predictions of sensory input that are compared to actual sensory input. This comparison results in prediction errors that are used to update and revise the mental model. Hence, the “goal” of the brain is to arrive at a mental model that is really good at predicting sensory input. And, in this view, the “goal of development” would be to prepare the brain as well as possible for predicting future sensory input. Development establishes “prior knowledge” that is useful for understanding the world.

There is a bit of evidence for something like this. Here are some of my favorites:

This is about as far as the field has come, i.e. this is the “state of the art”. This is a bit of a shame because it is only just starting to get interesting. I have so many more questions:

  • What can we say about this prior?
  • How much/what kind of information is contained in spontaneous activity?
  • Is this prior “just” for sensory and motor cortices, or might it also extend to higher cortical function?
  • Where does the prior come from (probably evolution, but when and can we identify changes over time)?
  • Under which circumstances is it possible to “bootstrap” a prior in this fashion (changes in circuit -> changes in activity -> changes in changes in circuit -> …)?
  • How might a developmental prior relate to large pretrained/foundational models in ML?
  • Can we apply some of the techniques/vocabulary from the theory of computation to this domain?

If I ever get to have my own research lab, these are questions I would love to work on.

In conclusion…

Condensed into a neat list of three points, what is my take on the development of the brain?

  1. Brain development is a protracted process that continues well after birth.
  2. After birth, the animal is exposed to a highly volatile environment that necessitates robust and general solutions.
  3. The reciprocal interaction between circuit organization and activation might robustly “bootstrap” prior information into the brain.

As mentioned above, point 3 is conjecture. But I’m very excited about that conjecture and if anyone has some leads or ideas (or questions), feel free to send me an email or to book some time for a chat with me :) Looking forward to talking!


1 I guess that was just a long-winded way of saying: “epistemic status: I’m trying my best.”

2 There is a whole zoo of arguments for the possibility (and timing) of human-level artificial intelligence. Most of them have an "existence" clause like this one

[...] I focus on one particularly salient vision of a transformative model: a model that can perform a large majority of economically valuable jobs more cheaply than human workers can. I see the human brain as an “existence proof” found in nature for this type of transformative model, so I use evidence from biology to help estimate the computation needed to train it. Ajeya Cotra in the draft report on timelines to transformative AI

If we want to build (artificial) intelligence, it is reasonable to look at the (arguably) only available example of a thing that has intelligence - the brain - to get inspiration and to estimate timelines. Two interesting upper bounds jump out immediately:

  • the amount of compute required for training. A rough Fermi-estimate:
    “Suppose it takes on average about 1 billion seconds (~32 years) for an intelligent human to go from an infant to their peak level of productivity. If a human brain performs ~1e15 FLOP/s), that would be (1e15 FLOP/s) * (1e9 seconds) = 1e24 FLOP, only about 1 OOM larger than the amount of computation used to train AlphaStar.” Ajeya Cotra in draft report on timelines to transformative AI
  • the size of the source code. The human genome can be encoded in ~750 megabytes of data. Since this includes a lot of junk and encodes the entire body and not just the brain, the size of the actual "source code" could be orders of magnitude smaller.

Moore's law lets us estimate how we will progress in terms of compute. But progress in terms of "source code" is much harder to estimate. The Microsoft git repository is apparently 300 gigabytes large, but of course, that's an unfair comparison. It's not about size, it's about having the right code. Looking at the development of the brain can help us to put more constraints on what an example of the right code looks like.

3 I know, I know justsaysinmice. I wish it was otherwise and there is a real danger of the streetlight effect.

A policeman sees a drunk man searching for something under a streetlight and asks what the drunk has lost. He says he lost his keys and they both look under the streetlight together. After a few minutes the policeman asks if he is sure he lost them here, and the drunk replies, no, and that he lost them in the park. The policeman asks why he is searching here, and the drunk replies, "this is where the light is".

But we know that some results obtained in animal models translate to humans and that this translation rate varies a lot depending on the phenomenon under investigation. And there are a few examples where theory developed for one or two animals generalizes to a much wider range of animals. So perhaps it's not so bad.

I would also, in the spirit of conceptual engineering, propose an important distinction: In my experience there are two types of researchers working on animal models:

  • some researchers study animal models because they are often the only thing available. Once something better becomes available, they will drop the animal model like a Totino's pizza roll that was in the microwave for too long.
  • other researchers really, genuinely care about that particular animal model. If it turns out that 0% of the results in that model generalize to anything else, that would only affect them in so far as they would have to come up with a different justification when writing grants.

Both types of researchers are doing valuable things, but I think it's important not to mentally lump them together since that can produce a lot of confusion in practice (i.e. when picking a lab for the Ph.D. or when motivating your research question or when explaining the impact of your research).

4 With “we” I mean “someone”.

5 Actual chip making, like molecular biology, is hard enough to effectively qualify as magic.

6 But often suboptimal in terms of efficiency on any given problem.

7 Rather than a fragile procedure with a lot of "if-then-else" clauses, I like to think of the development of the brain as a robust-agent-agnostic-process (RAAP) dead-set on producing a brain, no matter what.

8 One of the central truths that I learned in my studies is that "Neurons want to fire." A neuron that cannot produce action potentials will literally die. Very similar to that Keanu Reeves movie, actually.

9 This spontaneous activity was not visible in the clips in the section above (since there we didn't visualize the activity of the cells).

10 While these examples are specific to the visual cortex, spontaneous activity exists in the entire developing brain: the somatosensory cortex, the auditory cortex, the olfactory cortex, the thalamus, the hippocampus... just everywhere! And everywhere it has slightly different properties, shaped by the different neuron types in each area.

11 While I have peppered the previous sections with a lot of references, we are now getting into conjecture territory. “Predictive processing” is a broad field and there are a lot of different opinions represented. In particular, it is very unclear if/how the prediction error is exactly represented.

New to LessWrong?

New Comment
7 comments, sorted by Click to highlight new comments since: Today at 7:35 AM

I think the way I would say it, which might or might not be exactly the same as what you're saying, is:

Brains involve learning algorithms, trained on predictive loss (among other things). Learning algorithms need input data. In the case of rat vision, you don't get much "real" input data until quite late in the game (14 days after birth). If that's when the vision learning algorithm starts from scratch, then it would take even longer beyond those 14 days before the learning algorithm has produced a decent predictive model, and there's risk that the predictive model will still kinda suck when the rat is going out into the world and relying on it for decision-making.

To help mitigate that problem, the learning process kicks off earlier. Before there's any real visual data yet, there are retinal waves. We can think of these as creating synthetic input data that can start (pre)training the learning algorithms. Obviously, the synthetic data is not the same as actual real-world naturalistic vision data; therefore the model will obviously need to be further trained on the real data once we have it. But still, if the synthetic data is similar enough to the real data (in the ways that matter), this retinal wave pre-training could be better than nothing.

Anyway, if that's what you're saying, I view it as a plausible hypothesis. (But it's also possible retinal waves are doing something entirely different, I wouldn't know either way.)


Hey Steven! 

Yep, that's a pretty accurate summary. My intuition is actually that the synthetic training data might even be better than actual sensory input for pretraining because millions of years of evolution have optimized it exactly for that purpose. Weak evidence for that intuition is that the synthetic data comes in distinct stages that go from "very coarse-grained" to "highly detailed" (see f.e. here).

And you are also correct that retinal waves are not universally accepted to be useful - there was a long debate where some people claim that they are just a random "byproduct" of development. The new Ge et al. paper that came out a few months ago is a strong indicator for the functional importance of retinal waves though to the point where I'm pretty convinced they are not just a byproduct.

btw, I really enjoyed your post on the lifetime anchor. My take on that is that it doesn't make a lot of sense to estimate the lifetime anchor and the evolution anchor separately. Evolution can do the pretraining (through providing tailor-made synthetic data) and then the environment does the fine-tuning. That would also explain why Cotra's lifetime estimate appears so low compared to the amount of compute used on current ML models: current ML models have to start from scratch, while the brain can start with a nicely pretrained model.


Interestingly, I think you put a lot more importance on this "synthetic data" than I do. I want to call the "synthetic data" thing "a neat trick that maybe speeds up development by a few days or something". If retinal waves are more important than that, I'd be inclined to think that they have some other role instead of (or in addition to) "synthetic data".

It seems to me that retinal waves just don't carry all that many bits of information, not enough for it to be really spiritually similar to the ML notion of "pretraining a model", or to explain why current ML models need so much more data. I mean, compare the information content of retinal waves versus the information content of an adult human brain, or even adult rat brain. It has to be a teeny tiny fraction of a percent.

If I was talking about the role of evolution, I would talk instead about how evolution designed the learning algorithm, and inference algorithm, and neural architecture, and hyperparameter settings (some of which are dynamic depending on both where-we-are-in-development and moment-by-moment arousal level etc.), and especially the reward function. And then I would throw "oh and evolution also designed synthetic data for pretraining" into a footnote or something. :-P

I'm not sure exactly why we disagree on this, or how to resolve it. You're obviously more knowledgeable about the details here, I dunno.


Yeah, it's a tricky situation for me. The thesis that spontaneous activity is important is very central to my research, so I have a lot of incentives to believe in it. And I'm also exposed to a lot of evidence in its favor. We should probably swap roles (I should argue against and you for the importance) to debias. In case you're ever interested in trying that out (or in having an adversarial collaboration about this topic) let me know :)

But to sketch out my beliefs a bit further:

  • I believe that spontaneous activity is quite rich in information. Direct evidence for that comes from this study from 2011 where they find that the statistics of spontaneous activity and stimulus-evoked activity are quite similar and get more similar over development. Indirect evidence comes from modeling studies from our lab that show that cortical maps and the fine-scale organization of synapses can be set up through spontaneous activity/retinal waves alone. Other labs have shown that retinal waves can set up long-range connectivity within the visual cortex and that they can produce Gabor receptive fields and with even more complex invariant properties. And beyond the visual cortex, I'm currently working on a project where we set up the circuitry for multisensory integration with only spontaneous activity.
  • I believe that the cortex essentially just does some form of gradient descent/backpropagation in canonical neural circuits that updates internal models. (The subcortex might be different.) I define "gradient descent" generously as "any procedure that uses or approximates the gradient of a loss function as the central component to reduce loss". All the complications stem from the fact that a biological neural net is not great at accurately propagating the error signal backward, so evolution came up with a ton of tricks & hacks to make it work anyhow (see this paper from UCL & Deepmind for some ideas on how exactly). I have two main reasons to believe this:
    1. Gradient descent is pretty easy to implement with neurons and simultaneously general that just on a complexity prior it's a strong candidate for any solution that a meta-optimizer like evolution might come up with. Anything more complicated would not be working as robustly across all relevant domains.
    2. In conjunction with what I believe about spontaneous activity inducing very strong & informative priors, I don't think there is any need for anything more complicated than gradient descent. At least I don't intuitively see the necessity of more optimized learning algorithms (except to maybe squeeze out a few more percentage points of performance).

I notice that there are a lot fewer green links in the second point, which also nicely indicates my relative level of certainty about that compared to the first point.

Thanks! Oh it's fine, we can just have a normal discussion. :) Just let me know if I'm insulting your work or stressing you out. :-)

I believe that spontaneous activity is quite rich in information.

Sure. But real-world sensory data is quite rich in information too. I guess my question is: What's the evidence that the spontaneous activity / "synthetic data" (e.g. retinal waves) is doing things that stimulated activity / "actual data" (e.g. naturalistic visual scenes) can't do by itself? E.g. "the statistics of spontaneous activity and stimulus-evoked activity are quite similar and get more similar over development" seems to be evidence against the importance of the data being synthetic, because it suggests that actual data would also work equally well. So that would be the "shave a few days off development" story.

In conjunction with what I believe about spontaneous activity inducing very strong & informative priors …

The brain (well, cortex & cerebellum, not so much the brainstem or hypothalamus) does "online learning". So my "prior" keeps getting better and better. So right now I'm ~1,200,000,000 seconds old, and if I see some visual stimulus right now, the "prior" that I use to process that visual stimulus is informed by everything that I've learned in the previous 1,199,999,999 seconds of life, oh plus the previous 21,000,000 seconds in the womb (including retinal waves), plus whatever "prior" you think was hardcoded by the genome (e.g. cortico-cortico connections between certain pairs of regions are more likely to form than other pairs of regions, just because they're close together and/or heavily seeded at birth with random connections).

Anyway, the point is, I'm not sure if you're personally doing this, but I do sometimes see a tendency to conflate "prior" with "genetically-hardcoded information", especially within the predictive processing literature, and I'm trying to push back on that. I agree with the generic idea that "priors are very important" but that doesn't immediately imply that the things your cortex learns in 10 days (or whatever) of retinal waves are fundamentally different from and more important than the things your cortex learns in the subsequent 10 days of open-eye naturalistic visual stimulation. I think it's just always true that the first 10 days of data are the prior for the 11th day, and the first 11 days of data are the prior for the 12th day, and the first 12 days of data … etc. etc. And in any particular case, that prior data may be composed of exogenous data vs synthetic data vs some combination of both, but whatever, it's all going into the same prior either way.


What's the evidence that the spontaneous activity / "synthetic data" (e.g. retinal waves) is doing things that stimulated activity / "actual data" (e.g. naturalistic visual scenes) can't do by itself?

I don't think direct evidence for this exists. Tbf, this would be a very difficult experiment to run (you'd have to replace retinal waves with real data and the retina really wants to generate retinal waves).

But the principled argument that sways me the most is that "real" input is external - its statistics don't really care about the developmental state of the animal. Spontaneous activity on the other hand changes with development and can (presumably) provide the most "useful" type of input for refining the circuit (as in something like progressive learning). This last step is conjecture and could be investigated with computational models (train the first layer with very coarse retinal waves, the second layer with more refined retinal waves, etc. and see how well the final model performs compared with one trained on an equal number of natural images). I might run that experiment at some point in the future. Any predictions?

a tendency to conflate "prior" with "genetically-hardcoded information", especially within the predictive processing literature, and I'm trying to push back on that

Hmm, so I agree with the general point that you're making that "priors are not set in stone" and the whole point is to update on them with sensory data and everything. But I think it's not fair to treat all seconds of life as equally influential/important for learning. There is a lot of literature demonstrating that the cortex is less plastic during adulthood compared to development. There is also the big difference that during development the location & shape of dendrites and axons change depending on activity, while in adulthood things are a lot more rigid. Any input provided early on will have a disproportionate impact. The classic theory that there are critical periods of plasticity during development is probably too strong (given the right conditions/pharmacological interventions also the adult brain can be very plastic again), but still - there is something special about development.

I'm not sure if that's the point that people in predictive coding are making or if they are just ignorant that lifelong plasticity is a thing.

Thanks!! :-)

it's not fair to treat all seconds of life as equally influential/important for learning

I agree and didn't mean to imply otherwise.

In terms of what we're discussing here, I think it's worth noting that there's a big overlap between "sensitive windows in such-and-such part of the cortex" and "the time period when the data is external not synthetic".

Any predictions?

I dunno….

O'Reilly (1,2) simulated visual cortex development, and found that their learning algorithm flailed around and didn't learn anything, unless they set it up to learn the where pathway first (with the what pathway disconnected), and only connect up the what pathway after the where pathway training has converged to a good model. (And they say there's biological evidence for this.) (They didn't have any retinal waves, just "real" data.)

As that example illustrates, there's always a risk that a randomly-initialized model won't converge to a good model upon training, thanks to a bad draw of the random seed. I imagine that there are various "tricks" that reduce the odds of this problem occurring—i.e. to make the loss landscape less bumpy, or something vaguely analogous to that. O'Reilly's "carefully choreographed (and region-dependent) learning rates" is one such trick. I'm very open-minded to the possibility that "carefully choreographed synthetic data" is another such trick.

Anyway, I don't particularly object to the idea "synthetic data is useful, and plausibly if you take an existing organism and remove its synthetic data it would get messed up". I was objecting instead to the idea "synthetic data is a major difference between the performance of brains and deep RL, and thus maybe with the right synthetic data pre-training, deep RL would perform as well as brains". I think the overwhelming majority of training on human brains involves real data—newborns don't have object permanence or language or conceptual reasoning or anything like that, and presumably they build all those things out of a diet of actual not synthetic data. And even if you think that the learning algorithm of brains and deep RL is both gradient descent, the inference algorithm is clearly different (e.g. brains use analysis-by-synthesis), and the architectures are clearly different (e.g. brains are full of pairs of neurons where each projects to the other, whereas deep neural nets almost never have that). These are two fundamental differences that persist for the entire lifetime / duration of training, unlike synthetic data which only appears near the start. Also, the ML community has explored things like deep neural net weight initialization and curriculum learning plenty, I would just be very surprised if massive transformative performance improvements (like a big fraction of the difference between where we are and AGI) could come out of those kinds of investigation, as opposed to coming out of different architectures and learning algorithms and training data.