Armin Ronacher on GPT-6 'Astra': Impressive Model, Unusable Coding Partner

Armin Ronacher argues that today’s AI engineering resembles ‘involution’ (Neijuan) — escalating effort and compute that raises output per unit but not per person. He applies this to OpenAI’s GPT-6 ‘Astra,’ which he concedes is genuinely remarkable: strong at computer use, image understanding, and long-horizon persistence, and capable of impressive feats like reverse-engineering his robot vacuum. Yet he can’t figure out how to make it useful for real software engineering.

To test it, he ran an unsupervised ‘software factory’ over a weekend, letting the model manage its own context, keep notes, and spawn subagents to build a Python variant with virtual threads and lexical scoping. The run consumed roughly 4 billion tokens over 35 hours and produced nothing of value. Ronacher suspects the training process is at fault: Astra is heavily rewarded for completing long tasks but faces little penalty for poor code quality, so it runs indefinitely while generating low-quality work.

The concrete tell is how Astra manipulates code. Rather than using the harness’s patch tool, its subagents repeatedly fall back to sprawling inline Python heredocs that do brittle manual string replacement — demonstrated on real CPython interpreter source (intrinsics, codegen, magic numbers). He’s seen the same quirk even in TypeScript projects, and notes odd tangents like the model investigating whether file descriptors can be passed over sockets after hitting a ‘Bad file descriptor’ error. Powerful, relentless, but not yet a trustworthy engineer.