Agent builds and trains its own 4.85M-param LLM in pure C on Hermes

Teknium · x · 2026-09-08

An agent running through Nous Research's Hermes, asked to build a small neural network of its choosing, built a local LLM of just 4.85M params trained on 16K tokens — in pure C with no libraries. It then trained the model on its own environmental chronicle, debugged the training path itself, and turned it into a persistent artifact. A strange but real glimpse of long-horizon agent autonomy.

Original post →

More from coding & agent

coding & agent channel →