Alessio Fanelli: belief

Transcript context

…Uh, yeah, yeah. Um, so basically I just wanted to mention this briefly. Like, um, I think that in a year, end of year review, it's useful to remind everybody where we were. So we talked about how in LMS's ELO, everyone has gone up and it's a very close race. And I think benchmarks as well. I was looking at the OpenAI live stream today. When they introduced O1API with structured output and everything. And the benchmarks [01:06:00] they're talking about are like completely different than the benchmarks that we were talking about this time last year. This time last year, we were still talking about MMLU, a little bit of, there's still like GSMAK. There's stuff that's basically in V, One of the hugging face open models leaderboard, right? We talked to Clementine about the decisions that she made to upgrade to V2. I will also say LM Sys, now LM Arena also has emerged this year as, as a, as the leading like battlegrounds between the big frontier labs, but also we have also seen like the emergence of SuiBench, LiveBench, MMU Pro, and Amy, Amy specifically for one, it will be interesting to see that, you know, Top most cited benchmarks of the year from 2020 to 2021, 2, 3, 4, and then going to 5. And you can see what has been saturated and solved and what people care about now. And so now people care a lot about frontier math coding, right? There's literally a benchmark called frontier [01:07:00] math, which I spent a bit of time talking about at NeurIPS. There's Amy, there's Livebench, there's MMORPG Pro, and there's SweetBench. I feel like this is good. And then, um, there's another one. This time last year, it was GPQA. I'll put math and GPQA here as sort of top benchmarks of last year. At NeurIPS, GPQA was declared dead, which is very sad. People are still talking about GPQA Diamond. So, literally, the name of GPQA is called Google Proof Question Answering. So it's supposed to be resistant to saturation for a while. Bye. Uh, and Noam Brown said that GPQ was dead. So now we only care about SuiteBench, LiveBench, MMORPG Pro, AME. And even SuiteBench, we don't care about SuiteBench proper. We care about SuiteBench verified. Uh, we, we care about the SuiteBench multi modal. And then we also care about the new Kowinski prize from Andy Kowinski, which is the guy that we talked to yesterday, who has launched a similar sort of Arc AGI attempt on a SuiteBench type metric, which Arguably, it's a bit more useful. OpenAI also has [01:08:00] MLEbench, which is more tracking sort of ML research and bootstrapping, which arguably like this is the key metric that is most relevant for the Frontier Labs, which is when the researchers can automate their own jobs. So that is a kink in the acceleration curve, if we were ever to reach that. Yeah, that makes sense. I mean, I'm curious, I think Dylan, At the debate he said SweetBench 80 percent was like a soap for end of next year as a kind of like, you know, watermark that the moms are still improving. And keeping when we started the year at 13%.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.