Rob Bensinger lists some things that would make him more optimistic about AI risk:

“- Evidence that interpretability is making some truly deep progress -- e.g., directly producing deep, novel insights into workable CS algorithms that we didn't previously know about.

- Evidence for more radical interpretability progress, e.g., the ability to robustly read off models' thoughts or motivations.

- Evidence of serious progress on hand-crafting at least some critical features of model cognition, as opposed to everything being end-to-end learned.

- Evidence that deep learning is suddenly hitting a wall.

- Evidence of investment in next-generation frontier AI drying up.

- Evidence that AIs converge on robustly good values even when we don't train them to do so and don't train them to imitate human text.

- Evidence that AIs readily converge on corrigibility in one shot.

- Evidence that recent rogue AI incidents were badly mischaracterized, or that RSI and ML-progress claims at Anthropic and OpenAI are radically overhyped, etc.

- A good argument that any of the key premises for worrying about the risk were wrong -- e.g., a good argument that building more capable and useful models won't tend to make them more tenacious and won't cause them to treat safety measures of human overseers as obstacles.

- A good argument that powerful tech like nanotech is totally unfeasible, or would have far more narrow applications than one would have thought.”