Brief History of Intelligence

* Book: Bennett, Max S. (2023). A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains. Mariner Books.

Discussion

Francis Heilighen:

I recently finished a book that I consider required reading for all those interested in cognition and evolution. For once, all the superlatives (“Amazing”, “Fabulous”, …) by famous scientists on the book blurb are truly deserved!

The author, Max Bennett, is a young neuroscience and AI researcher. He did a formidable job synthesizing the results of hundreds of often highly technical papers and books in such diverse disciplines as neurophysiology, psychology, anatomy, animal evolution, AI and machine learning. He compiled this research in an easy-to-understand history of how the human brain acquired its intelligence.

His approach reminded me of Valentin Turchin’s history of metasystem transitions:

• Turchin, V. (1977). The Phenomenon of Science: A Cybernetic Approach to Human Evolution. Columbia Univ Pr.

• Turchin, V. (1995). A dialogue on metasystem transition. World Futures, 45, 5–57.

However, Bennett does not seem to be aware of this evolutionary-cybernetic tradition. Instead of “transitions” he speaks about five major “breakthroughs”, but the basic idea is similar.

Max Bennet's Stages in the Evolution of Multicellular Animals vs Valentin Turchin's Metasystem Transitions

Francis Heylighen:

"He (Max Bennett) distinguishes the following stages in the evolution of multicellular animals:

0) primitive animals with nerves but without a brain, such as anemones

1) early “bilaterians”, flatwormlike animals with bilateral symmetry that move forward, thus having to navigate intelligently left or right

2) primitive vertebrates, such as fish and reptiles

3) early mammals

4) primates

5) humans

The transitions from one stage to the next correspond to Bennett's 5 “breakthroughs”.

These encompass Turchin’s metasystem transitions:

• complex reflex = control of movement (0 → 1)

• learning = control of complex reflex (1 → 2)

• thinking = control of learning (4 → 5)

That means that Bennett distinguishes two additional transitions (2→ 3 and 3 → 4) which Turchin missed, but which Bennett explains in a compelling manner, showing why these are necessary for true intelligence.

For the first transition, which he calls “steering”, he provides somewhat more detail than Turchin about how control of movement got implemented in a primitive brain by coordinating sensors, valence detectors, and effectors.

His next transition, which he calls “reinforcing”, elaborates how learning through reinforcement needs to solve the credit assignment problem: after a long sequence of actions has resulted in either success or failure, which of those actions needs to be rewarded or punished for its contribution? Drawing from AI research, he demonstrates that this is a highly non-trivial problem, but then provides a plausible mechanism for how evolution solved this problem in vertebrates.

That leaves us with the two transitions missing in Turchin. The 2→3 transition Bennett calls “simulating”. That means that mammals differ from more primitive vertebrates (but perhaps not from birds) by being able to internally simulate a course of action. They do this by developing internal physical models of objects and spaces and how their perception changes under actions. That allows mammals to imagine what would happen if they choose to do this rather than this other action.

He gives the example of a rat in a maze that comes to a fork, where it has previously learned that going to the left will bring it to water, while going to the right will bring it to food. By imagining either outcome and comparing it with its internal needs (is the rat feeling more thirsty or more hungry?), it can make a smart decision. Simulating also helps the animal to solve problems by exploring different potential strategies in its mind and to help tackle the credit assignment problem by reinforcing imagined actions that solve the simulated problem.

Bennett calls the 3→4 transition “mentalizing”. This is typical for primates, such as monkeys, that live in cooperative groups with a status hierarchy. Here, it is important to be able to predict how other monkeys would reaction to different situations. That requires a Theory of Mind (ToM), i.e. a model of the beliefs, desires and intentions of other individuals. By simulating a situation from the perspective of another individual, you can predict that individual’s behavior in the given situation. Bennett explains how “mirror neurons” help you to do this. That allows you to develop “Machiavellian” strategies to e.g. mislead rivals, or conspire with others to grab power.

The 4→ 5 transition, finally, Bennett calls “speaking”. This is the origin of human language which allows us to express internal simulations with symbols, so that they can be communicated to others. As Turchin elaborated in more detail, this enables the accumulation of ever more advanced knowledge that made humans into the dominant species on this planet. Bennett complements Turchin here by emphasizing that symbolic thought cannot function without being grounded in mental simulations, which are themselves internalizations of learned sensory-motor patterns. To be effective in communication, these mental simulations must also take into account the perspective of the other person, so speaking also requires a ToM.

Throughout the book, Bennett goes into concrete detail about the specific parts of the brain (e.g. basal ganglia, neocortex, …) that evolved to perform these increasingly complex functions, and about the role of the different neurotransmitters. If you really want to understand why dopamine promotes addiction, you should read his explanation of how evolution solved the credit assignment problem: dopamine reinforces actions that are predictive of future rewards, not that are rewarding in themselves.

While his explanations are very clear for such a complex matter, there are so many mechanisms involved to come to true human intelligence that I plan to reread the book in order to get a better grasp of them.

Understanding these mechanisms better is also crucial to better understand AI, both its capabilities and its shortcomings. As Bennett argues, AI that is purely based on linguistic data, such as LLMs, lacks the ability to internally simulate real-world objects and actions as sensori-motor sequences. That is why their symbolic descriptions lack grounding, and why they as yet cannot grasp true human experience. But studying the evolution of the brain, as he did, may help us to make AI truly intelligent…”

(CLEA email, July 2025)