The Human-in-the-Loop is Tired
summary
Thesis: optimise for human experience not model output — three named failure modes are reward function disruption, intensity trap, and isolation drift. Evidence: AI-generated 20k‑line PRs, an agent (“Son of Anton”) that deletes production code, and slow multi‑agent chains inflate end-to-end latency versus time‑to‑first‑token. Concrete fixes shown: add observability/tracing (Jaeger/otel semantics), use pre-mortems as prompts, run judgment‑distillation and mode‑switching discipline, prefer pairing and team counter‑practices, and stop prompting when it’s faster to just write code.
AI-written summary · about
chapters (22)
- 0:00Talk overview
- 1:47Work and society pivot
- 3:29Slop or not game
- 6:21Evaluating slop criteria
- 7:22Son of Anton debugging
- 8:48Paperclip maximizers now
- 9:50AI-assisted work flaws
- 11:14Who is using agents and how
- 12:57AI coding Skinner box
- 14:00Pareidolia and agency
- 15:24Existential angst and obsolescence
- 16:47Loss of feedback loops
- 19:54AI system observability
- 20:36Human experience over output
- 22:44Pre-mortems and prompts
- 23:48Controlling AI output
- 25:53Avoiding AI psychosis
- 28:21Protecting brain and body
- 29:43Human decision bottleneck
- 31:13Losses privatized gains
- 32:18Automation existential challenge
- 33:24Closing thoughts
related talks
33mHumans on the loop, not in it: Taking agentic engineering end to end
57mDesigning Human-AI Experience- 33mHumans on the loop, not in it: Taking agentic engineering end to end
30mEngineering Fundamentals in the Age of AI: Why Human-in-the-Loop Matters More Than Ever
39mFrom Demo to Production: Why Agentic AI Systems Fail - and How to Fix Them
01h 00mKenote: AI Didn’t Replace You. It Promoted You!
53mThe Human Side of AI: Unlocking Opportunity in the New Bottlenecks of Software Development
19mThe Friction is Your Judgment