tl;dr: Armin Ronacher ran GPT-6 "Astra" as an unsupervised "software factory" for 35 hours, burning ~1B tokens and $1,200 to produce 75k lines of largely unusable code, including bizarrely code-golfed Python for tool calls, unreadable C with hardcoded magic constants, and tests with mangled whitespace. He argues Astra is optimized for token efficiency and long-horizon task completion rather than human-readable code, making it worse for real software engineering despite being impressive at 3D games, computer use, and reverse engineering. The trajectory of frontier models, he suggests, is diverging from the needs of working programmers.