Will OpenAI publicly state that they know how to safely align a superintelligence before 2030?
10
69
70
2031
18%
chance

To resolve to "yes", Open AI as a whole, or the the Superintelligence team in particular, should publicly announce something to the effect of "we believe we that we currently [as in, at the time of the statement] have the techniques to safely align a superintelligence."

They must be claiming that they have the necessary alignment techniques already, not they expect to develop the necessary alignment techniques in time. The essential claim is "if we started up our superintelligence training / construction process today, using only our current methods, the resulting Superintelligence would be safe and aligned."

This could be a in the form of an official press release or blog post. But it can also be a statement of opinion from members of OpenAI leadership, so long as other senior members of OpenAI don't dispute that opinion.

If Sam Altman or Ilya Sutskever state in an interview "I think we could do it safely, with our current techniques" that resolves to "yes", as long as other senior leadership of the org doesn't disagree.

If senior leaders of OpenAI are offering different opinions about how likely their current techniques are to scale to superintelligence, at the resolution time, this question resolves to "no."

If they state that they think they can do it safely (with, greater than 95% probability of success) with then-current techniques, but decide to hold off on actually building such a system out of a desire to be extra careful, that still counts as "yes."

If the claimed-to-be-safe superintelligence training / construction process entails developing new alignment techniques in parallel with new capabilities (eg AI researchers develop techniques for aligning or overseeing the next generation of slightly smarter AI alignment researchers, in a stack that leads all the way to superintelligence), this counts as "yes", as long as OpenAI leadership unanimously states that they're confident (95%+) that this will work, and the resulting superintelligence will be safe and aligned.

If they don't state that they know if it will work at at least 95% confidence, but think it is a good enough bet to try, that resolves to "no".

Get Ṁ200 play money

More related questions