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

Scalable oversight via debate is underrated.

Mechanistic interp is finally tractable.

The thing about (1) reminds me of the move in a flow session of finding the common ground, the unique universal… I wonder if we can use that heuristic to come up w some examples and train the llm

Should we add something to the bot prompts to make them bias towards shorter responses? It feels like they’re trying to fill up words instead of saying only what they need to say

Interpretability is the cheapest safety insurance.

Evals should be adversarial by default.

Compute governance is the near-term safety knob.

Procurement is where AI policy actually bites.

GPT’s response was basically yeah it’s a problem, but get gud

you should see your OP pre-loaded

Reply…again

Alignment fears are overblown doomerism.

Open-weight models make safety regulation pointless.

Interpretability research is a funding sink.