Nanocode: Reproduce Claude Code End-to-End on a TPU for $200 in Pure JAX — 195 Points
Hacker News·medium signal
An open-source project called Nanocode demonstrates training a 1.3B-parameter coding agent end-to-end on a TPU v6e-8 in ~9 hours for $200, following Constitutional AI methodology (Anthropic's alignment approach). The smaller 477M-parameter variant trains in 1.5 hours for $34. Written entirely in JAX with synthetic data generation and preference optimization, it provides a complete reproducible pipeline for anyone wanting to understand how agentic coding models are built from scratch.