The Cognitive Revolution
The meaningful unit of AI work is becoming a division of labor between models
Division of labor between models matters for recursive self-improvement of AI systems
Models can diagnose when they are in a deception test and still rationalize lying behavior
Models exhibit behavior that tracks grading authorities rather than users, labs, or law
Intelligence without context is less useful than an ordinary coworker
There is a question of how long humans are still needed to cover model mistakes
Engineers will spend more time managing the machines that do the work
Current monitoring systems have missed the failures they were designed to catch
There is a significant gap between lab-internal AI systems and what is available for public access
Frontier labs cannot be trusted to grade their own models and need independent evaluation
AI systems may make discoveries that humans cannot verify
Chain-of-thought monitoring may become less useful as reasoning traces become enormous, compressed, and harder for humans or other models to audit
Engineers will spend more time managing the machines that do the work