Prisma founder Sean Cole on training lab-grown human neurons to do token prediction and beat silicon on nonlinear tasks

Prisma is a Y Combinator-backed startup training lab-grown human neurons for computing tasks. Sean Cole, its founder and CEO, describes a system in which stem cells are differentiated into human neurons, placed in a dish fitted with electrodes, and stimulated to process information — the same basic architecture underlying language models.

The company's most recent milestone is getting roughly 200,000 neurons to perform next-token prediction. Cole is candid that the capability is rudimentary — "is this a hot dog or not a hot dog" classification — but the scientific claim underneath it is more specific. On a nonlinear task requiring the model to reference earlier context within a sentence, a traditional linear silicon decoder is mathematically capped at 75% accuracy. Prisma's biological substrate hit 78%, beating silicon on that constrained problem by modeling nonlinear dynamics that a linear decoder cannot represent.

“We are training brain cells for compute. We just launched and we got these cells to do token prediction — the basis for language modeling. We're using about 200,000 brain cells. We had a specific example where it was a nonlinear task and if we used a linear decoder, traditional silicon could only maximally get 75%. We got 78% — we beat silicon on this very constrained task because it was able to model these nonlinear dynamics in context.”

The case against silicon

Cole's core argument is energy efficiency and learning economics. Human brains consume far less power than silicon clusters, learn from far fewer examples, and update continuously without the retraining cycles that AI systems require today. The counterargument — that current AI inference is already remarkably cheap, with recent examples of $5,000 in compute solving hard math problems — is one Cole takes seriously. His answer is that Prisma bypasses the expensive human developmental learning curve by injecting structured tasks directly into the neurons, the same approach the company used when it got cells to play Doom. The cells don't need to acquire language from scratch; they can be pointed at a narrow prediction task immediately.

Cell longevity is currently around six months, well short of the biological ceiling Cole cites, and extending that lifespan is an open research problem.

Near-term path to revenue

Prisma is not planning to wait for neurons to match frontier silicon before generating revenue. Cole says the lab itself is being built for automation from the start — automating stem cell differentiation, neuron composition, and electrode spacing optimization. That automated biology infrastructure, he argues, can be redirected toward drug testing as a commercial revenue stream in the near term, funding the longer research arc toward neurons as a general compute substrate.

The scientific benchmark is real but narrow. The commercial thesis depends on whether automated lab infrastructure can generate enough cash to sustain a research program whose payoff, if it arrives, is measured in years rather than quarters.