A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains

Introduction

"Evolution is still unfolding in earnest; we are not at the end of the story of intelligence but at the very beginning. "This positions us not as intelligence's culmination but as one stage in ongoing progression.

A Brief History of Intelligence tackles a question most books split apart: how did brains evolve AND what does that tell us about building AI? Bennett, working as AI entrepreneur with neuroscience advisors, argues these questions are inseparable.

The book's structure is five evolutionary breakthroughs, each solving a specific problem: steering toward rewards, reinforcing successful behaviors, simulating actions mentally before executing, modeling other minds, and encoding knowledge in language. Each breakthrough built on previous ones, and each maps to current AI capabilities and limitations.

What makes this valuable is Bennett's dual literacy. He explains temporal difference learning in dopamine systems and why it matters for AI credit assignment. He shows why large language models excel at pattern matching but fail at common sense - they have language breakthrough without simulation and mentalizing foundations.

The framework is clarifying. Current AI has mastered breakthroughs one, two, and five partially, but lacks three and four.

This explains both AI's surprising capabilities and its stupid failures. A system can generate eloquent text while unable to understand that solid objects can't pass through each other.

This isn't AI hype or fear. It's systematic analysis of what intelligence requires, measured against both biological evolution and artificial progress.

Evolution's Sequential Intelligence Design

Let's start with the architecture itself. Evolution didn't design intelligence all at once—it built it in stages, each solving a specific survival problem. Five breakthroughs, each layered atop the last. Here's what matters about this structure. Each breakthrough was only possible because the previous one already existed.

Not in a loose way, but mechanistically necessary. You couldn't bolt these capabilities together in any random order.

Take the second breakthrough, reinforcement learning. This is the fish-level ability to repeat behaviors that historically worked and avoid ones that didn't. Trial and error learning. Sounds simple, but it absolutely required the first breakthrough already in place.

The first breakthrough was steering, which gave early bilateral animals their basic valence system. Valence means the neurons that tag things as good or bad, pleasurable or painful.

These live in the hypothalamus, the ancient core of the brain that still runs your body today.

Without valence neurons, reinforcement learning is impossible. Because trial and error needs a learning signal. When you try something, you need to know if it worked or failed.

Good or bad. That's what valence provides. It's not just helpful, it's the foundation the whole system runs on.

So fish couldn't have evolved reinforcement learning without inheriting that ancient valence machinery from their coral-like ancestors.

The basal ganglia, which handles the trial and error, literally builds on top of and depends on signals from the hypothalamus.

This same logic applies up the chain. The third breakthrough, mental simulation in early mammals, required the second.

Because simulation means imagining actions before doing them, then learning from those imagined outcomes. But if you can't learn from trial and error, imagining trials is useless.

The mammalian neocortex renders the simulations, but it's still the basal ganglia that learns from them.

The fourth breakthrough, mentalizing in primates, required the third. Theory of mind means modeling other minds, but that's just applying your existing simulation machinery to internal mental states instead of external actions.

Same computational process, new target. You can't model mental states if you can't model states at all.

And language, the fifth breakthrough, required mentalizing. Because communication assumes the other person has different knowledge than you do.

Without theory of mind, you can't infer what needs explaining or interpret what others mean by their words.

This is why intelligence took four billion years. Not because evolution was slow or lucky, but because you have to build the foundation before you can build the next floor. Each breakthrough solved real limitations of the previous system, but only by preserving everything that came before.

Modern human intelligence isn't one unified thing. It's five systems stacked on top of each other.

The hypothalamus still running your basic drives, the basal ganglia still doing trial and error, the older neocortex still simulating, the newer neocortex still mentalizing, and language coordinating it all.

When any layer fails, you see exactly what gets lost. Damage to theory of mind regions destroys social reasoning while leaving spatial planning intact. That's not random, that's architectural.

Review

Five breakthroughs over four billion years. You carry all of them—the ancient valence system craving dopamine hits, the pattern matcher jumping to conclusions, the simulator running scenarios you'll never act on, the mind-reader constantly guessing what others think, and language binding it together. Next time your phone's AI confidently tells you something physically impossible, you'll know exactly what's missing.

Next time you catch yourself imagining disaster scenarios at 3 AM, that's breakthrough three doing its job.

Intelligence isn't one thing that suddenly appeared. It's layers of solutions to survival problems, each built on what came before.

We're not debugging intelligence. We're reverse-engineering it. And we've barely started.