Nanocode: The best Claude Code that $200 can buy in pure JAX on TPUs (github.com) AI

A GitHub discussion introduces “nanocode,” a fully open-source, end-to-end approach to train a Claude-Code-like agentic coding model using pure JAX on TPUs. The author describes an architecture and training pipeline based on Anthropic-style Constitutional AI and Andrej Karpathy’s nanochat, including synthetic data generation, preference optimization for alignment, and TPU-optimized training. They report that a 1.3B-parameter model (d24) can be reproduced in about 9 hours on a TPU v6e-8 for roughly $200, with smaller variants costing less, and provide starter commands and training/evaluation notes.