Evolution could encode the brain in DNA
Many people have observed that the human brain is extremely complex and that a complete description can’t possibly fit inside the DNA. For example, from Dwarkesh Patel:
The difference is that evolution has to be titrated in the case of humans through three gigabytes of DNA. That’s very unlike the weights of a model. Literally, the weights of the model are a brain, which obviously does not exist in the sperm and the egg. So it has to be grown. Also, the information for every single synapse in the brain simply cannot exist in the three gigabytes that exist in the DNA.
I think this is only true in a very narrow sense and is not a big constraint for evolution.
With compression algorithms, there is always a trade-off between the compression ratio and the length of the algorithm. The more mutual information there is between the compressor and the data, the better the compression.
Per the Kolmogorov-Levin theorem, mutual information in
Relatedly, while the brain is very complex, the world outside is much more so. And evolution doesn’t have to figure out how to build the brain, only how to transform information in the environment into one.
A useful reframing is to imagine evolution programming in an extremely high-level language, with everything in the environment being part of its (unimaginably large) standard library. This might allow even extremely detailed brain design to fit in DNA, even though it doesn’t help much with the compression of, e.g., a random string.
The brain contains roughly 10^14 synapses. Assuming the strength of a synapse contains at least 4 bits of information, and that there’s at least 4 bits of information per synapse in the choice of which nearby cells this cell has synapse connections versus which it don’t, that’s roughly 10^14 bytes of information. The human genome is roughly 10^9 bytes of information [Dwarkesh’s numbers are wrong, 3.1 billion base pairs is less than 1GB, not 3GB: a base-pair is two bits, a byte is eight bits], not all of which is devoted to the brain. In fact, since the difference between the chimp and human genomes is only about 2%, while the human brain is roughly 4 times as large as the chimp one, it would seem that probably, at least for the neocortex rather then the brainstem (since the neocortex is where the differences between chimps and humans are concentrated: our brainstems do fairly similar things), probably pretty small proportion of the genome is devoted to that part of the brain. So at least 99.999% of the information in the human brain comes from its environment or is built up internally during development rather than being directly hard-coded by the genome.
Nevertheless, the genome clearly gives a very developmentally useful jump-start and set of priors (like “here’s a good start on several useful ways to do depth perception”, and “pay attention to anything that looks roughly like a primate face”). But likely it’s focused on what to learn and how to learn it, rather than providing bulk of the actual information learnt: there simply isn’t space for the latter, by many orders of magnitude. And in AI terms, the genome has to be worth less than a well-chosen 1B parameters.
Thanks for the numbers, I completely agree on the information counting. I was trying to make (and failed) a slightly different point. Let me try to put it in more precise language.
This is the main sentiment I was trying to push against. It is not that I think DNA provides the bulk of the information. It is just that it doesn’t start from nothing, create a few priors and then hand things off.
Imagine some basic brain structure + sensory inputs encoded by some simple set of genes. When facing the real world, this brain will reliably be “filled” (i.e., learn lots of information). But, DNA doesn’t have to only encode priors or loss functions for this learning process. It can also encode transformations to perform on the completed thing. For example, you could have things like “copy structure from region X to region Y”, “extract half of circuit X and combine it with circuit Y to create a new circuit Z”.
This still does not mean that the minute details are in the DNA itself. But, I think you can build practically anything you want with tricks like that, because many specific learned circuits will contain general-purpose reusable subnetworks.
My main point is that evolution isn’t limited to designs like “create a starting point with good priors and let go”. It can encode things that start with a completed brain, and then copy/paste/merge/shuffle parts around to create something new. Given that most algorithms you might want to encode likely can be expressed as parts of already existing learned ones, a complicated design can fit in very few bits. Just like you could write a complicated program with very few lines of code in a high-level language.
I have a crackpot theory related to this that I don’t take too seriously, but I think about it sometimes anyway. Much of developmental biology depends on gene regulation networks and chemical gradients. Encoding complicated information into the brain would therefore require a large number of regulatory switches. If each of these switches needs its own protein, then not much complexity can fit in the genome. So the crackpot theory is that some of the non-coding DNA is actually what contains most of the information. RNA can control gene expression, so it’s not totally crazy. Of course, we somehow have to make this consistent with the amount of sequence conservation in non-coding DNA.
EDIT: Just to explain further, the crazy part of the theory is not that the non-coding DNA is doing stuff, it’s that there is a fairly general and modular system for specifying patterns of neuron connections, where adding a new circuit or drastically modifying an existing one can be done only by modifying non-coding DNA, without needing to touch coding DNA at all.
I think “gene expression is significantly affected in systematic ways by non-coding DNA, and RNA is a vector by which this happens” is completely mainstream in genomics, and has been for a pretty long time?
Yeah, there’s a spectrum there from things that are reasonable to things that are unreasonable, and I think I didn’t include the most unreasonable parts. I will edit.
The claim that human DNA is 3GB is false. I assume the assumption was that 3 billion base pairs equates to 3GB.
There are a few issues with this:
1. DNA is a quaternary sequence, not a binary sequence (Adenine, Guanine, Cytosine, Thymine). 3 Billion base pairs are informationally equivalent to 6 billion bits.
2. Human haploid cells have a single set of chromosomes composed of ~3 billion base pairs. However, somatic cell’s are diploid with 2 sets of chromosomes composed of 6 billion base pairs. The bases present at each allele are not identical and allele combinations have significant phenotypic relevance.
3. A GB is 1,000,000,000 bytes, where each byte is 8 bits.
With these adjustments, one concludes each human’s genome is ~1.5GB.
Additionally, our focus can be reduced further by considering that not all DNA directly encodes information relevant to phenotypic expression, including introns, pseudogenes, transposable elements, satellite DNA, etc.. Although subject to ongoing debate, there are primarily two definitions of function: biochemical activity (which the ENCODE project determined to be ~80.4%), and DNA conserved across mammalian species under negative selection (~8.2%, although there are a range of values). These fall further when restricted to the cell types relevant to the structure of the human brain.
The range of genome size across species is much smaller in magnitude than the differences in the number of genes, and these variables don’t appear to positively correlate. Considering many complex traits such as intelligence are highly heritable, one might expect the number of genes would be relevant but it appears not (C-Value Paradox). The solution seems to be that DNA encodes a developmental pathway, which even single nucleotide differences can permanently damage, and cell types which form complex phenotypes through emergent structures, even with their foundations in simple cellular architecture. Although, I will be careful to note that this is essentially a question which would require a full description of human development as an answer, which is clearly more complex than a few sentences.
Yes, that’s just what evolution does, it compresses the enormous amounts of information in the brain into a few gigabytes of data. The decompression algorithm takes a decade or two and involves a lot of looking around and trying things and judging the outcomes of the things that were tried.
My standard analogy for this is the “busy beaver” problem (https://en.wikipedia.org/wiki/Busy_beaver) in computer science: given a Turing machine of a given size/shape, what’s the starting configuration which produces the longest output that reaches a termination state?
It turns out that even small Turing machines that could be described in a handful of bytes can have colossal output runs (which terminate). As examples:
4-state, 2-symbol busy beaver: Result: 0 0 1 0 1 1 1 1 1 1 1 1 1 1 1 1 0 0 (107 steps, thirteen “1”s total)
5-state, 2-symbol busy beaver: Result: 4098 “1”s with 8191 “0″s interspersed in 47,176,870 steps.
Current 6-state, 2-symbol best contender: Result: more than 2↑↑↑5 “1”s in more than 2↑↑↑5 steps, where 2↑↑↑5 = 2↑↑2↑↑2↑↑2↑↑2 and ↑↑ represents tetration.
The TL/DR is that small computational systems can produce gloriously huge structured output; and what is biology but a computational system?
@olehif @Dennis Towne I find it unlikely that a huge structured output like the brain can be used without a period of being taught. Suppose that, instead of the brain, the DNA encoded an LLM’s initial weights. Then I don’t expect the LLM to be usable without a long period of pretraining, RLHF singling out the Assistant Persona and RLVR which made the LLMs far more capable.
Agree! I just think “being taught” might be a wrong frame. As a though excercise, imagine evolution “hardcoded” the concept of “bear” directly in DNA like this:
Make a structure in the brain that pattern matches a few characteristics of a bear
Make this stucture capture and save other signals present at the time trigger first fired
If parents reliably show pictures of bears to kids, then this concept will be created in everyone, even though true description of “bear” might be orders of magnitude more complex.
People would normally call this thing “learning”, but I think you might also call this compressed encoding (or compilation), because you could not have learned anything else.
There’s pretty clear evidence that the human brain has the concepts “snake” and “spider” genetically encoded in it, as pre-grown neural recognition circuits for their shapes and movement styles. Yes, they really are creepy — to humans. Humans are not instinctively afraid of either snakes or spiders, but there’s good experimental evidence that we’re genetically predisposed to specifically recognize them and to quickly and efficiently learn to be afraid of any specific type of snake or spider, if and only if we see adults in our community act afraid of that type, often in as little as a single exposure to this behavior. So we’re instinctively poised to rapidly learn to be afraid of any specific types of snakes and spiders that our community fears. Which is a fascinating example of instinctively predisposed cultural learning.
Stuff like this is the reason the old Nature-vs-Nurture argument was particularly unhelpful: the answer isn’t just “usually both”, it’s often “a subtle intricately-interleaved combination of the two”: genetic and cultural coevolution.