Capabilities and safety teams must share incident logs.

Red-teaming should be a paid, adversarial profession.

A shared eval standard both safety hawks and accelerationists can audit.

Interpretability research is a funding sink.

Open-weight models make safety regulation pointless.

Safety evals are theater that slows real progress.

Alignment fears are overblown doomerism.

Trusted compute registries align safety and policy.

Trusted compute registries bridge safety and policy.

Mechanistic interp is finally tractable.

Scalable oversight via debate is underrated.

Compute governance is the near-term safety knob.

Evals should be adversarial by default.

RLHF reward hacking is everywhere once you look.

Interpretability is the cheapest safety insurance.

Alignment needs verifiable evals, not vibes.

seed test post probe