Thanks for the detailed post and your work at ControlAI. A few questions:
On timelines: when you say 10% chance of success, does that mean within 1 year at $50M, or after several years of sustained funding at that level? And how many years do you think we have before a treaty becomes impossible to enforce?
On strategy: given the 10% figure, you are already in long-shot territory. In decision theory, when resources are limited and you are in an unfavorable game, the mathematically optimal approach is often ‘bold play’: maximizing variance rather than playing conservatively, since cautious strategies tend to guarantee slow failure. Have you considered this framing? Do you have unconventional, higher-variance strategies you are pursuing or considering but did not publish? Or do you think other organizations should be pursuing those bolder bets while you focus on the scalable conventional approach?
On other actors: who do you think the biggest players are in this space, for example FLI, PauseAI, or technical alignment research? What is your rough estimate of each of their probabilities of success? And are these approximately independent attempts at the same goal, such that the chances roughly compound?
Thanks for the detailed post and your work at ControlAI. A few questions:
On timelines: when you say 10% chance of success, does that mean within 1 year at $50M, or after several years of sustained funding at that level? And how many years do you think we have before a treaty becomes impossible to enforce?
On strategy: given the 10% figure, you are already in long-shot territory. In decision theory, when resources are limited and you are in an unfavorable game, the mathematically optimal approach is often ‘bold play’: maximizing variance rather than playing conservatively, since cautious strategies tend to guarantee slow failure. Have you considered this framing? Do you have unconventional, higher-variance strategies you are pursuing or considering but did not publish? Or do you think other organizations should be pursuing those bolder bets while you focus on the scalable conventional approach?
On other actors: who do you think the biggest players are in this space, for example FLI, PauseAI, or technical alignment research? What is your rough estimate of each of their probabilities of success? And are these approximately independent attempts at the same goal, such that the chances roughly compound?