A fascinating and very approachable treasure trove introduction to the evolution of the brain as a recursive prediction machine via a framework of five breakthroughs, each building on top of each other: valence associations and steering, reinforcement learning, simulation, mentalization and language.

Max S. Bennet’s “A Brief History of Intelligence” is a well-written mix of overview and detail about one of the most fascinating and relevant questions of our times; the nature of intelligence. The purely descriptive narrative starts all the way from the evolutionary pressures and enablers for the development of the very first brains that integrate sense stimuli into steering decisions for bilaterian bodies and ends with various considerations about artificial intelligence as the potential next, sixth breakthrough. This is the area where I found the book to be the most limited to very surface level stuff, but that might simply be the combination of my background and it not really being the main point.

The book’s strengths lie overwhelmingly in the evolutionary neuroscience parts of which it gives a comprehensive tour with just enough detail to provide good starting points for further self-directed research and a sense of emerging, nuanced understanding with a fair balance of arguments and perspectives. While being a couple of years old makes the parts on LLMs specifically in the end quite outdated already, the book does, however, draw various very useful, enlightening and much more timeless analogies to AI and machine learning throughout, demonstrating how parallel and synergistic the development of the disciplines has been and just how closely inspired many machine learning models are by the different breakthroughs.

It is obviously valuable for all those interested in the evolutionary and mechanical correspondence between AI and the brain, but even more interesting philosophically, as it paints a rather mechanistic, materialist picture that doesn’t necessarily leave a lot of room for dualism and other similar such ideas. The book is full of what could be considered fascinating trivia for most, such as the bipedal body plan being central for steering, neuronal adaptation largely being driven by maximum neuron firing rates being around 500 while needing to represent much larger and smoother continuums of sense data, categorization of emotion into coordinate axes of arousal and valence, curiosity being fundamental for all learning, and the typical localizations of various functions and capabilities, such as language understanding being located in the Wernicke’s area and production originating from Broca’s area, but the real kicker is the description of the brain as a recursive prediction machine, that could explain the unreasonable effectiveness of something like the transformer architecture scaling so well across all modalities.

Bennet presents a picture of the brain, where many of the breakthroughs can be explained by the more and more abstract inputs and outputs to and from neocortical columns, which can be likened to a generic unit of computation in the brain. Breakthrough 2 (reinforcement learning) is about learning to predict future rewards for actions taken in the present, breakthrough 3 (simulating) is about learning to predict all sensory data in response to those actions, i.e., modeling the world and breakthrough 4 (mentalization) is about learning to predict the simulations of your own and those of others, i.e., modeling the agency of entities including yourself. This paints the development of intelligence - and perhaps even consciousness - as a rather continuous process, which humblingly reminds just how transiently we exist on the same spectrum with all other life. This is a major simplification, but a really beautiful and eye-opening model that presents very interesting questions and challenges to various worldviews.

The chapters on breakthrough 5 describes how we are separated from other primates that have gone through the four previous breakthroughs by our use of language, which is acknowledgedly the most controversial take in the book, but well justified by the degree to which it is different. Here, however, Bennet argues that while all the previous breakthroughs have been relatively clear, structural changes, our language capabilities are mainly due to a hardwired script for learning them, as the main difference with primates is mostly just size. This was absolutely fascinating to me. It seems to be the case that humans have an instinct to conversation - babies taking turns making noises with their parents even when they make no sense and paying attention to the same things - that comes first, before the vocabulary and grammar is then just slotted into pre-existing capability via the hardwired script and it is the existence of these that really differentiate us by enabling never-before-seen global co-operation.

This was a great book that taught me a lot and gave me a ton to chew on. Thank you Aaron for both the recommendation as well as for lending the book itself! I will have to return to this still and so, if you ever feel compelled to buy me a present for one reason or another, this would be a great pick! You might need at least high school level understanding of certain subjects to breeze through the book, but I highly recommend it to everybody the least bit interested in neuroscience, AI, evolution, intelligence or philosophy of consciousness, or you feel like you need a bit of a grounding from anthropocentric human exceptionalism.