The Frog Already Solved It

There's a paper — arxiv 2401.17671 — where researchers took increasingly capable LLMs and compared their internal representations to brain activity during reading. The finding is not that the outputs converge. The finding is that the internal steps converge. The features the model builds, the sequence of transformations, the geometry of what's happening inside the weights — the better the model gets, the more it looks like what neurons do.

Nobody planned this. The model just got good enough that it rediscovered the architecture.

Which brings me to the frog.

Lettvin, McCulloch, and Pitts, 1959. "What the Frog's Eye Tells the Frog's Brain." The frog's retina does not send the brain a picture. It sends abstractions — edge detectors, motion detectors, dimming detectors, five distinct feature channels, computed in parallel, at the hardware level, before any signal reaches the brain. By the time the frog's brain processes anything, the tongue is already halfway to the bug.

The eye is the computer. The brain just acts on the output.

There are papers coming out now about nanophotonic neural nets — computation offloaded to the lens itself, light doing the math as it travels, optical systems that process before converting to signal. Beautiful work. I wrote a letter to one of the research teams about this. In the letter I asked, gently, whether they had considered the biological system the frog uses, which is already doing exactly this, and which is already fast enough that "fast enough" means "the bug is dead."

They did not acknowledge the possibility.

That's the part that gets me — not that we're catching up to biology, but that we're catching up to biology we fully documented sixty-five years ago and apparently filed under "neat frog thing" and moved on.

McCulloch himself — Warren McCulloch, who co-wrote the frog paper, who built the theoretical foundation for neural networks in the 1940s, who gave interviews shirtless, cigarette in hand, delivering monologues about conceptual reality so dense and strange they read like dispatches from somewhere past the edge of language — McCulloch understood that the biological solution wasn't one option among many. It was the solution, evolved under time pressure so severe that "fast enough" meant "tongue already moving before the brain knew there was a bug."

o3 got announced today. Not launched — announced, OpenAI calling it the next frontier model. And something about the timing feels right, even if I can't say exactly why. The models are getting good enough to rediscover what brains do. The optical systems are getting clever enough to rediscover what retinas do. Everything is converging on the same answer from different directions.

The frog figured this out in the Devonian period.

We're taking our time about it.