Engineering Enablement by Abi Noda Unpacking METR’s findings: Does AI slow developers down?
24 snips
Aug 1, 2025 
Quentin Anthony, Head of Model Training at Zyphra and EleutherAI contributor, dives into the surprising findings of METR’s study on AI coding tools. He discusses how developers often feel productive using AI but may actually slow down. Quentin emphasizes the importance of self-reflection in coding practices to avoid bloated code and suggests strategies for effective AI integration. He also explores the concept of 'context rot' and advocates for a tailored approach to using AI tools based on specific tasks to enhance efficiency.
AI Snips
Chapters
Transcript
Episode notes
Unexpected AI Slowdown Experience
- Quentin Anthony initially expected AI to speed him up significantly on tasks based on his previous experience.
- He was surprised when the METR study revealed an average slowdown, despite subjective enjoyment of AI use.
Practice Self-Reflection with AI
- Developers need to self-reflect on whether AI tools are actually speeding them up or producing good code.
- Teams should call out bloated or low-quality AI-generated PRs to avoid degrading codebase health.
Know When to Stop Chasing AI
- Stop using an AI model if you find yourself fighting more with it than solving the problem.
- Critically review AI-generated code and look for repeated requests or poor outputs as signs to disengage.