Will Schenk

In 1976 Hans Moravec wrote a paper at Stanford that made the argument that Intelligence is not a miracle it’s about have enough power.

He opened with a tour through evolutionary pathways, all leading to intelligence by different paths. Vertebrates and mollusks diverged from a common ancestor, something like a hydra with a primitive nerve net, and both independently produced intelligent species.

Cephalopods evolved imaging eyes, large brains, and problem-solving ability through a completely separate architecture. The octopus brain is a ring around its esophagus, its blood is green, there is no blind spot in its vision, and in the Cousteau film one figures out how to uncork a sealed jar to get at a lobster inside.

Birds got there a third time. Crows and ravens outperform all mammals except primates on learning tests, using brain regions – the Wulst, the hyperstriatum – that don’t even exist in mammalian brains. The cortex is small and irrelevant in birds. They built the same capability out of completely different parts.

Three lineages. Three architectures. Intelligence is not an anomoly. It’s a consequence.

Moravec measured the threshold. A hundred neurons runs a sessile animal. A thousand gets you a worm. A million gets you a bee – fast and interesting but stereotyped. A billion gets you imaging vision. A hundred billion gets you language, planning, culture. The relationship between connection count and behavioral complexity is remarkably consistent across totally unrelated lineages. Intelligence is what nervous systems do when they cross a threshold of connectivity. The form varies. The emergence doesn’t.

Nietzsche, who knew nothing about neurons, named the drive a century earlier. What he called will to power is not a drive to dominate. It’s what living systems do: they discharge their strength, grow, overflow. “A living thing seeks above all to discharge its strength – life itself is will to power.” Intelligence, in this framing, is not a gift or an accident. It’s what accumulated power produces at sufficient scale. Convergent evolution is the empirical proof – cephalopods, corvids, and primates are not coincidence. They are what power does.

Moravec had essentially an engineering argument. If intelligence is a predictable consequence of sufficient connectivity, then building it shouldn’t require genius-level theoretical breakthroughs. It requires power. He estimated the brain at about 40 x 10^12 bits/sec – a million times the PDP-10 that Stanford’s AI lab was running on back in 1976. And then:

Although there are known brute force solutions to most AI problems, current machinery makes their implementation impractical. Instead we are forced to expend our human resources trying to find computationally less intensive answers, even where there is no evidence that they exist. This is honorable scientific endeavor, but, like trying to design optimal airplanes from first principles, a slow way to get the job done.

Today we call it the bitter lesson since the AI field spent the next 70 years doing exactly what he warned against: expending human intelligence trying to find clever shortcuts around the power deficit. The belief that pure symbolic reasoning – the right algorithm, the right representation – could substitute for raw compute is the Platonic move Nietzsche spent his career attacking: spirit over body, form over force.

The old hope that cleverness could do what only power can, the sort of thing you’d expect a weak clever person to go for. It was honorable scientific endeavor. It was also, as Moravec predicted, a slow way to get the job done.

With enough power anything will fly.

Original PDF from Stanford