Parasma (YC): Human Brain Cells Achieve Next-Token Prediction
ycombinator · x · 2026-08-14
Y Combinator-backed Parasma has launched, aiming to use human brain cells as an energy-efficient alternative to GPUs. Having previously trained neurons to play Doom, the company is now demonstrating their potential in language modeling.
Key technical highlights include:
- Token Prediction: Context words are converted into electrical stimulation for the cells, with responses decoded back into tokens.
- High Accuracy: Cells correctly predicted 90% of next-token decisions when full sentence context was applied.
- Non-linear Capability: In harder non-linear tasks, the biological cells achieved 78.1% accuracy, exceeding the 75% ceiling of a silicon linear decoder.
This suggests custom biological computing architectures could support complex language-based tasks.
More from Venture
- Foxconn July Revenue Jumps 54% YoY Driven by Strong AI Server Demand — Beth_Kindig · 2026-08-14
- Inception Offers YC Startups 250B Free Tokens to Push Diffusion LLMs — volokuleshov · 2026-08-14
- Applied Materials Raises Revenue Forecast: AI Chip Demand Far Exceeds Capacity — firstadopter · 2026-08-14
- TechBio 50 List Released: Top VCs in the AI x Bio Space — hardimanjames · 2026-08-14
- Databricks Raises $5B at a $190B Valuation Amid Massive Investor Demand — TechCrunch AI · 2026-08-14
- Everyone in AI is a Mini Nostradamus: Tech Growth vs. Stock Bubbles — Infamous-Office9698 · 2026-08-14