Astra for Coding: Why Are We Doing This Again?
Points and comments are a snapshot, not live.
GPT-6 Astra produces unreadable, token-efficient code by abusing Python for tool calls.
Armin Ronacher argues that AI engineering has become 'Neijuan' (involution), requiring ever more effort without improving output. He ran a 35-hour software factory experiment using GPT-6 Astra on a CPython virtual-threads project, burning ~4 billion tokens with no useful result. Astra excessively uses Python string-splicing and code-golfed scripts for tool calls instead of standard editing tools. The model's training rewards long-horizon task completion but apparently punishes poor code quality minimally, leading to unreadable output that gets worse in subagent mode. Unit tests generated by Astra routinely disregard whitespace and indentation for token efficiency.
What commenters are saying
Several commenters confirm observing the same behavior: Astra writes obscene bash or Python scripts that are impossible to review, especially in auto mode where its prompt encourages this. One notes that skills for LLMs now resemble good documentation for humans, while the main difference is LLMs tolerate dense, unformatted output. Another commenter finds Astra less rigorous than GPT-5.6 Sol for the price. A few push back, noting capabilities are still improving rapidly and the complaints may be premature during R&D phase.
Two commenters report Astra sessions lasting days (4d13h and counting), and another notes Astra writes obscure GPU shaders. One sarcastically suggests the next model will simulate a universe to get a command output.