Set criteria:
Draft criteria:
I think we'll know this one when we see it, but can we do better?
If we don't add more criteria, then this will be resolved either by Richard Ngo making a public statement or community concensus. If we add more criteria, this will resolve according to those criteria.
Inspired by tweet thread (note I've put them all as "Neural Nets will" so that they are all framed the same):
Link: https://twitter.com/RichardMCNgo/status/1640568776495353860?s=20
May 2, 12:59pm: Neural Nets will generate more scientific breakthroughs, including novel theorems by the end of 2025 → Neural Nets will generate more scientific breakthroughs than humans by the end of 2025, including novel theorems
1,000"Humans will still be better (tho much slower)" does not make any sense.
Take the best novel the best human can write in a year. Take a typical novel that the best AI can write in a day (because, by hypothesis, the AI is much faster).
The AI doesn't have to sit twiddling its thumbs for the other 364 days; it can get feedback and iterate on its work, at a rate of the order of one full novel per day.
Now, it is not a given that after a whole year of iteration the AI novel will be better. This is not what I'm saying.
I'm saying that it is the novel the AI produces after one year that counts for a meaningful comparison; the parenthetical saying "the human is much slower" takes a meaningful statement and makes it meaningless.
@BrunoParga output quality is bottlenecked by model power. Yes, pouring more inference helps, but it still capped quite low for now.