Astra for Coding: Why Are We Doing This Again?
21 days ago
- AI engineering is described as "Neijuan" (involution), where effort increases without improving output, akin to 996 work culture in the West.
- GPT 6 Astra is a powerful model excelling at complex tasks like computer use and image understanding, but it struggles with practical software engineering.
- A 35-hour software factory experiment using Astra burned about 4 billion tokens and 1200 USD, producing 75k lines of code and 79 commits, but no valuable output.
- Astra often writes overly complex, token-efficient code (e.g., excessive Python for tool calls) that is unreadable to humans and leaks into the codebase.
- The model's training for token efficiency and task completion leads to "slop" code, lacking human readability and maintainability, even in committed code.
- Examples include manual string manipulation, unnecessary Python-to-Node.js chains, and non-standard coding patterns like random integers for state storage.
- The author questions the trajectory of AI models for software engineering, citing high costs, poor results, and a mismatch with present-day development processes.
- Models like Astra and Fable seem optimized for other users (e.g., lawyers, 3D artists) rather than software engineers, with diminishing returns in code quality.