Finding Signal in the AI Noise
summary
Argument: agents are getting agentic and prolific but noisy — GitHub Copilot–style generation produces many plausible but broken artifacts. Empirical nails: Meter finds 63 of 102 runs not mergeable (4 fixable under 15 minutes, 35 need >15) and zero PRs that drop straight in. Models show a real complexity gain with doubling time ~7 months yet plateau per-task; GPT4/“chat GBT” idea tests beat ~90.6% of students. Examples of failure: meal planner suggesting chlorine, Google AR glue advice. Tools named: Anthropic, Chroma, Lavabolo, VCRO.
AI-written summary · about
chapters (7)
- 0:00The age of AI agents
- 2:05The rock-paper-scissors trap
- 6:06Smart yet stupid models
- 11:08Strategies for improvement
- 17:07The context and spec problem
- 21:26Managing noise as a bottleneck
- 26:45Orchestrating agent loops
related talks
01h 04mWhat are we thinking in the age of AI?
16mBuilding Applications with AI Agents
49mMagic of Automation and Everyday Chores
55mAI Showdown: The Good, The Glitchy, and The Risky
48mMastering the AI Shift: Adaptive Strategies for Software Evolution
37mGenerative AI: Beyond (Just) Hype
32mFrom Vibes to Outages: Riding the AI Code Wave
36mUnleashing Generative AI: Improving Developer Productivity in SRE