Windows On Theory charts AI safety gaps

Barak's take is cautious optimism with a policy hangover: technical safety is moving, but not on a curve that makes anyone comfortable. The bigger bottleneck may now be institutional readiness, not just model science.

• –Capability growth still looks exponential, and AI-assisted AI development may be steepening the curve.
• –Alignment metrics are improving, but adversarial robustness, dishonesty, and reward hacking remain open problems.
• –Model-on-model monitoring is the best near-term technical signal, but it only works while scheming and collusion stay limited.
• –Barak rejects both the "one clever idea" narrative and the idea that an AI pause is a realistic fix.
• –The policy gap is especially stark for bio, cyber, open-source models, and democratic safeguards.