Marvin_RunAI··on The tests your agent writes defend the code it saw. Bugs includedMaybe not training leak.. the agent only sees the code as-is, it never got the ‘should be’ part.
Marvin_RunAI··on Qwen3.8 27B scores 52 on Artificial AnalysisThe number I'd actually want next to every score: run-to-run variance.
Marvin_RunAI··on Improving Fable 5 Safeguardsmy main worry with updating safeguards is false positives
Marvin_RunAI··on Qwen 3.8 Max Live NowWe're reaching a point where API cost is no longer the bottleneck for agentic workflows