In the comments of “Yudkowsky and Christiano discuss takeoff speeds”, Christiano mentions a “<4% on end of 2025″ for OpenAI (or other) pulling off an IMO gold medal. (However, after looking up actual IMO problems, he comments “feeling bad” about those 4%, because the problems would actually be doable and PR could be enough motivation.)
At the end of the thread I settled on “<8% by end of 2025.” We should clarify the final state of the bet soon so that it’s easier to reference (and Eliezer also has a chance to update his probabilities).
My main update was how often you could get a gold with only 4 problems, and generally how often it looked like you could get away with only answering easy questions. This kind of variability in the test generally pushes you back to 50% even though I think Eliezer and I still have a stronger disagreement about theorem-proving progress.
At the end of the thread I settled on “<8% by end of 2025.” We should clarify the final state of the bet soon so that it’s easier to reference (and Eliezer also has a chance to update his probabilities).
My main update was how often you could get a gold with only 4 problems, and generally how often it looked like you could get away with only answering easy questions. This kind of variability in the test generally pushes you back to 50% even though I think Eliezer and I still have a stronger disagreement about theorem-proving progress.
Oh right I had missed that comment. Edited the post to mention 8% instead. Thanks.