“A lot of AI researchers think this is a really good book.” - Adam Marblestone, AI researcher and CEO of Convergent Research speaking on the Dwarkesh podcast.

I first read about this book via author and columnist Ian Leslie’s substack newsletter. He’d written a review and then later had the author Max Bennett to interview on his podcast, which I also listened to. Since then I’ve heard it repeatedly referenced in discussions of how LLMs differ from human brains. I am fascinated by discussions of how our brains work and was intrigued by a book that offered an overview of all the latest advancements in neuroscience, especially as it promised to also look at how this could influence, and be influenced by AI research. It was published in 2023. I bought my hardback copy from Amazon in August this year almost immediately after reading Ian Leslie’s review. The purchase-to-read time of 5 months is remarkably low for me, which is a mark of how keen I was to read it. Max Bennett is an interesting character. He’s not a neuroscientist or evolutionary psychologist, but a US startup founder. He sold his company BlueCore, an AI system for product recommendations, for a sizable sum and now devotes much of his time to his passion project. While not an academic, he’s no lightweight dilettante, having published several peer reviewed papers on the subject.

The book takes us on a tour of brain evolution from the earliest animals that evolved neurons as simple cellular wires to join sensors to muscles around 600 million years ago, to the human brain, a self aware system of around 80 billion neurons, one of which is currently writing this review. Bennett only covers the evolutionary line to humans and only occasionally mentions possible parallel evolutionary developments in other linages, such as insects or birds, in passing. He describes five “breakthroughs”, when significant new developments in animal brains enabled new kinds of cognition and behaviours.

The first breakthrough came with the first bilaterians. These are the ancestors to all animals with bilateral symmetry; us, insects, worms, and pretty much anything that’s not starfish or sea urchin with radial symmetry. This first breakthrough was steering, which required early nematode like worms to make choices about which way to steer, and whether to move or not. This was the evolution of “good and bad” as Bennett puts it, and the origin of emotions governed by familiar neurotransmitters such as dopamine which steered these ancient worms towards food or away from predators.

The second breakthrough occurred with vertebrates, the earliest fish-like creatures which evolved temporal difference learning structures in the basal ganglia, pattern recognition in the cortex, and the perception of three-dimentional space in the hippocampus. These allowed these creatures to model the world they lived in and learn how to navigate to an internal map. From the second breakthrough onward the complexity of vertebrate brain wiring exceeds current methods of modelling the wiring, so exactly how the world is modelled is unknown.

The third breakthrough arrived with the evolution of mammals. This was the neocortex, a simulation building engine. I found the discussion of this truly fascinating. All mammalian neocortexes are built to the same pattern on a modular two dimensional sheet of cortical-columns. The current hypothesis is that each one is a self contained neural network dedicated to some task. Depending on how you count them the human brain has around one million of these. The evolution of simulation allowed mammals to predict and game different scenarios before acting. Our brains are a constantly running simulation of the world informed by our senses; what one neuroscientist called “anchored hallucination”. It’s why our dreams and imagination can be so real, because we use the same simulation to see the world that we use to imagine it. Again, but more so, we have no idea exactly how these cortical-columns work, how they wire together, or how they evolved. Form behavioural studies, brain imagining and experiments with brain implants in primates it’s clear that it’s some kind of generative model, but other than that it’s a black box.

The fourth breakthrough occurred when primates added a new layer of neocortex to their frontal lobes. This was also a simulation generating system, but now the simulation was of the self. Primates started to predict how they themselves would behave in a certain simulation. They use this to imagine themselves in the future, or analyse what happened in the past. It can also be repurposed to simulate the minds of others to imagine what they might be thinking and how they might act; the birth of politics as Bennett puts it. A third purpose was both learn and teach new skills by simulating observed actions and then applying them to the self-simulation. And to teach you also need to be able to put yourself into the mind of someone who doesn’t yet have your skill. It seems this development created consciousness, or self-awareness; a sentient being.

The fifth and final breakthrough comes when we finally arrive at our human ability to speak and transfer our inner mental state to others. This transfer mechanism allowed the creation of culture, a repository of information that can escape the limitations of a single brain. With culture, and especially with writing human brains become the substrate on which an infinite and exponentially growing hive mind can evolve. While every breakthrough before this is physically manifested in the development of a new piece of brain anatomy, there doesn’t seem to be any specially evolved new part of the brain for speech. The human brain is simply a scaled up chimpanzee brain, and the areas of the neocortex that are usually (but importantly not always) dedicated to understanding and processing speech look much the same as the rest and are present as smaller versions in chimp brains. Since there are no other mammals which speak a little - we speak a lot and chimps not at all - it’s almost impossible to do any more than guess at how it came about. I found Bennett’s musings that the difference between us and chimps is simply an evolved desire to learn how to speak unsatisfactory. Steven Pinker in his book The Language Instinct points out that while the language you speak is learnt, all languages have a shared semantic structure or framework. Since this is the same across human cultures it must be innate and evolved.

The book ends with Bennett speculating about how this understanding of the human brain could better inform AI research. He points out that models like GPT have captured something of the way humans process language. We also try to predict the next word as a mechanism for deducing meaning from language. What they don’t do is our follow up step of using that to build a simulation. If you describe a room to a person they will build a full 3D model of that room in their heads. You “picture” a scene in your mind. GPT style models do not do that. He thinks that this anchoring in a simulation of physical reality should be the next step. He also thinks that the modular nature of the neocortex, the vast array of cortical-columns, might be a useful model to persue. It does seem as though nature has discovered a neat trick for horizontal scaling only limited by the physical constraint of supplying the brain with enough calories. This is very different from how our current monolithic AI models work.

One discussion that fascinates me and that Bennett covers, is how the development of human culture is a paradigm shift as significant as the origin of life itself. It’s a popular conceit to describe ourselves as “just another mammal” and talk of evolution as if it will continue in much the same way as it has over the last 4 billion years, but that clearly is not the case. Something fundamental has changed. For the last 4 billion years all the complexity and variety of the natural world is encoded in the DNA of the millions of species of bacteria, plants, and animals that populate the planet. Change has happened at the rate at which DNA can mutate. DNA has been the repository of all information. Now that is no longer true, a new far faster evolving information space has emerged: our culture. Change on planet earth is no longer constrained by the mutation of DNA, instead it has been replaced by a self-improving exponentially increasing hive-mind that will undoubtedly escape the bounds of Earth and expand among the stars. Although we can debate the timescales it seems inevitable that our human intelligence will be superseded by the artificial kind. What this will mean for humans and each individual human life is impossible to know and impossible to plan for, except to expect ever accelerating change.

The five-breakthroughs, evolutionary approach to understanding and explaining the brain makes the book easy to digest. I happily read 30 or 40 pages at a sitting. Bennett is an excellent writer. The book is pitched at just the right level for an interested layperson to understand, with enough, but not excessive jargon; always a difficult judgement call for an expert in a subject to make. His simulation of the reader’s mind must be very good! Or maybe it’s just that I’m a good fit for his simulation? Either way I found it fascinating and hugely educational. I don’t really have any real criticisms, but there was one surprising error when he misnamed the KT extinction event that wiped out the dinosaurs as the “permian-triassic” extinction. That was the “great dying” that happened hundreds of millions of years earlier. He repeated the same mistake several times. The fact that I, someone with a causal interest in the subject, finds it glaring error makes me wonder how it ever got past his proof readers? Didn’t anyone with expertise in palaeontology read it?

This is one of those books that resets one’s model of the universe. It’s made me think of consciousness and intelligence in a far more nuanced and anchored way. Possibly my book of the year? In any case highly recommended.