Generative AI Fundamentals Map
Map/Loops/balancing

Verification bottleneck

balancing loop · 4 steps · Trace on the map →

The cycle

Models produce moreworkMore output to checkChecking capacity runsoutErrors slip or outputis throttledB
  1. 01Models produce more workAgent-run discovery
  2. 02More output to checkOutput verification at scale
  3. 03Checking capacity runs outScalable oversight
  4. 04Errors slip or output is throttled

What drives it

Output scales with compute but checking scales with people, so checking caps useful output.

What breaks it

Cheap automated verifiers and readable chain of thought relieve it. Verifier’s law predicts the easy-to-check fields escape first.

What to watch

Review time per AI task. Share of AI output that ships unreviewed.

Also feeds it

Sources

  1. 01Jason Wei: Asymmetry of verification and verifier’s law
  2. 02Why Do Multi-Agent LLM Systems Fail?
  3. 03Chain of Thought Monitorability: a new and fragile opportunity