Li et al.

Avriz, a cloud platform for coding agents, has built a contextual bandit system that learns to route LLM calls to the cheapest capable model without risking user experience. The system trains exclusiv…

A software engineer ran two A/B tests comparing AI audit quality with and without a "stakes" prompt, finding that telling the AI "I'll be fired if this isn't done well" improved results by 24% in one …