If you have interesting questions about AIS, or about life, please bring them to my attention! Maybe I’ll blog about some of these!
- Is adversarial robustness useful?
- Could you possibly gain relevant insights from solving it, even if it’s not directly useful?
- Why don’t constitutional classifiers solve this?
- Are evals useful?
- ie METR (note that METR also does other stuff too).
- Note: over-eager sharing of work is just a way of avoiding responsibility
low priority:
- TG: Would it be better if the USA was the only country with nukes?
- TG: we have survived nukes thus far, just based on the “no one chooses to use them” principle. Why are you so pessimistic that a multi-polar situation where multiple countries have AI couldn’t just be fine based on the “no one chooses to use them to do large amounts of harm” principle?
- AW: What (if anything) is the difference between threats and rewards?
- Should you ever make threats?
- AW: How impervious to blackmail should you be?
- For the record (i.e., for anyone considering trying to blackmail me :]), my current stance is that I will not respond to any blackmail attempts.
- Also blackmail might not really be the right word for the concept I’m thinking of. Maybe “threats” is the right word.
- EY has a post that basically argues that you should accept positive sum acausal trades, and reject negative sum acausal trades. I think this seems good.