Agents in Machine learning

Preprint: Claude Code moved GPU projects to TPUs, but 49% of failed runs raised no error

Researchers had Claude Code migrate ten PyTorch projects; eight met the project goal, with 2.4 human interventions on average. In one, an unnoticed gradient-clipping change cut accuracy from 86.53% to 84.89%.

Why it matters: An agent reporting success doesn't prove the migration worked. The authors advise independent reference runs on the hardware you'll actually use.