We train human brain cells for compute. In March, we trained them to play Doom; now we got them to do next-token prediction.

The Problem

Modern models require enormous amounts of energy and training data, while still struggling with continual learning and adapting efficiently to new tasks. Biological brains solve these problems very differently. They learn continuously, adapt from limited experience, and operate on roughly the power of a light bulb.

Solution

Parasma is developing the algorithms and infrastructure needed to turn human brain cells into useful compute. The human brain is the most capable and energy-efficient learning system we know. Our goal is to reliably train biological substrates and use them for general computation, robotics, and eventually direct biological training. In March, we trained human brain cells to play Doom, which got hundreds of millions of views across social media. Since then, we’ve been working toward more general tasks. Most recently, we got the cells to perform token prediction: given a sequence of tokens converted to electrical stimuli, the neurons learn to predict what comes next with no GPU.

Long-term, biological compute will offer orders-of-magnitude improvements in energy and sample efficiency, while enabling native continual learning.

Team

I previously worked on neuromorphic computing and reinforcement learning during my master’s.

Before that, I built a successful digital-item marketplace with > $1.7M of listed goods and played games professionally.

Check out our press release here:
https://parasma.com/news/human-brain-cells-do-next-token-prediction