Story of Intelligence

huangpf
Feb 21
5 min read

Updated: Feb 23

Ancient Egyptians thought the heart did the thinking. Aristotle agreed. We've come a long way since then. Or have we...?

Last Friday at the Stage, we shifted gears away from practical applications to history. Clarence Lam, who runs Chiasm, a Swiss-based startup focused on the design and optimization of complex systems, brought us on a brief journey from early life and nervous systems to present-day human belief systems and artificial intelligence, exploring intelligence as a recurring solution for shaping the future across biology, society, and machines.

If you're building with AI, this history will be an interesting read for you.

The Question Nobody Can Agree on

Here's the test he gave to us. Is a jellyfish intelligent? What about a worm with 300 neurons? A tree? A coral?

One group said everything was intelligent. Another said only things that make conscious decisions qualify. Someone else argued that a jellyfish "doesn't put any effort in.", and so isn't intelligent.

The purpose of that question wasn't to get the right answer, because there isn't one. It was to highlight that when it comes to intelligence, we still don't have a shared definition of what it actually means.

Flight Is Simple. Intelligence Is Not.

So why is Intelligence so hard to define? Here's a simple framework:

Flight is a binary problem. Something either flies or it doesn't. You'd never debate whether a rock flies. The Wright brothers solved a complicated problem: lots of moving parts, but predictable. Understand aerodynamics, build the right wing shape, and you're airborne.

Intelligence on the other hand is a complex problem. It's continuous, multi-dimensional, and the goalposts never stop moving. A few years ago, generating an image was considered intelligent. Now it's just another a feature in your phone's photo app.

This distinction between complicated and complex:

Complicated: Linear cause and effect. Predictable. Reproducible. A calculator gives you the same answer every time.
Complex: Nonlinear. The sum is greater than the parts. Sensitive to initial conditions. You can't truly A/B test it.

And emergence, the thing everyone in AI keeps chasing, is very complex

What 3,000 Years of Being Wrong about Intelligence Tells Us

You may be taking your understanding of intelligence for granted, because for the longest time, intelligence has always been misunderstood. Here's a quick history:

Ancient Egypt (3000 BCE): People believed that the heart was the seat of thought, while the brain was considered unimportant. During mummification, the brain was discarded. The heart was preserved.
Hippocrates (5th century BC): Went against the grain. Argued the brain is where thought and sensation happen. Revolutionary for a time when disease was blamed on curses.
Aristotle: Thought leader of his era pushed everyone back to the heart theory. He argued that the brain's job was to cool our blood. (Even thought leaders get it wrong.)
Galen (2nd century): Ran experiments on animals. Established the brain as the organ of the mind. But because Aristotle had already won public opinion, most people kept believing it was the heart.
Descartes (17th century): Only humans are intelligent. Everything else is not.
Thomas Willis (17th century): Willis proposed that intelligence depends on brain tissue organization — not individual components. Think diamond versus graphite. Both are carbon atoms, completely different properties, entirely because of arrangement.

This history is also a reminder that we cannot be over-confident. We could still be far from a correct understanding of what intelligence is, despite our confidence.

The Brain's Hidden Network

This is where it gets really interesting for me.

When we think about intelligence, we always think about neurons. But there's an entire parallel network in your brain called the glia network. They are like the "cleaning crew" of our central nervous system. They modulate which neural pathways get faster, which get pruned, and which get reinforced. Research increasingly shows that this cleaning crew don't just support neurons, they also actively participate in how the brain processes information.

What's most surprising is this relatively unknown network is actually has a one-to-one ratio with neurons. (they are also the ones that become tumours)

Sometimes, it's as important to prune as it is to grow.

The Brain Region That Makes Us Human

Deep in the brain, at the junction of visual processing, language, attention, and motor control, sits the temporoparietal junction (TPJ).

What it does is remarkable: it gives you the ability to infer the thoughts and beliefs of another agent. It's not exactly empathy, but something more fundamental. The capacity to understand that someone else has a completely different mental model than yours.

The Sally-Anne test measures this. It only develops as you get past a certain age. A child will only pass this test around the age of five. In contrast, crows, who have the ability to make tools, recognize faces, and have complex social relationships can't pass this test.

This ability, theory of mind, is what enables morality, deception, game theory, collective belief systems, religion, complex language, and art. It's also an ability that explains why humans built civilizations but crows could not.

For AI builders, this is the frontier. Current language models can simulate perspective-taking, but whether they truly model another agent's beliefs is one of the deepest unsolved questions in the field. And this also raises the question of whether AGI is possible.

In Max Bennett's book A Brief History of Intelligence breaks down the evolution of intelligence into five critical moments:

Steering (Bilaterians) – Reactive control: sensor → action loops. No learning, just immediate movement toward/away from signals. (e.g., A*, obstacle avoidance)
Trial-and-Error (Vertebrates) – Behavior shaped by rewards and punishments over time. Learns what works. (e.g. Reinforcement Learning)
Simulating (Mammals) – Internally predict outcomes before acting. (e.g. Model-based planning and generative simulation.)
Mentalizing (Primates) – Model other agents’ beliefs, intentions, and strategies. Multi-agent reasoning. (e.g. Theory of Mind)
Speaking (Humans) – Symbolic compression of thoughts and simulations, enabling coordination, culture, and shared reasoning (e.g. LLMs parallel this layer).

Each breakthrough built on the last. We've gotten pretty good at replicating breakthroughs 1 through 4 in artificial systems. Breakthrough 5 is where things get complex.

What This Means If You're Building With AI

Here's what you can take away from 3,000 years of intelligence research:

Structure matters more than components. Thomas Willis knew this in the 1600s. The arrangement of your system, more so than individual node, determines what it can do.
Don't ignore the maintenance layer. Glia taught us that the system managing the network is as important as the network itself. What's the "glia" of your AI product?
Intelligence isn't binary. Stop asking "is this AI intelligent?" Start asking "intelligent at what, in what dimensions?"
Complicated solutions won't solve complex problems. If you're treating user behavior, market dynamics, or multi-agent interactions as complicated (predictable, reproducible), you're going to be surprised.
The hardest problem is perspective. Building AI that generates text is complicated. Building AI that understands what you need right now? That's complex.