Shawn Wang: belief

Transcript context

…So very beginning, it was really just an internal research project, because our reinforcement learning team wanted to compare some results that they had with published papers, and they did not manage to. And so they opened a small leaderboard where they had manually evaluated a bunch of models. And the community really took over. People were super motivated. And after one month and a half, that's when it was given over to Nathan and I, so that we could make it into an actual engineering product, which could run an actual production rather than a research project. So at the moment, we have evaluated on the version one of the OpenLLM leaderboard, 7,400 models, which have been community submitted for most of them. I think around 800 discussion threads of users interacting with us, either for support or for suggestions. We have had several million visitors since the creation of the leaderboard. And yeah, the scale of it is quite huge. We actually have, from time to time, startups sending us thank you messages saying, oh, our model ranked so high on the leaderboard that we actually got a funding round thanks to you. And so they are very happy about this, and we get thank you messages. So it's quite used, especially by the community. A lot of the community is using it to test their methods and test how well their ideas perform against the SOTA models. It's instructive to rewind back maybe one, maybe two years ago when this kind of leaderboard practice wasn't normal. It's not normal to have an independently validated leaderboard. Everyone just kind of runs their own evals and publishes their evals against their models on their own paper, and it's not reproducible. So I think it's really about reproducible science. I think the only other, before this, it was kind of like ML perf, that's the other big leaderboard that I can think about where, obviously, maybe AlexNet, specific competitions, specific benchmarks, but not something that aggregates across all the other benchmarks. Maybe HuggingFace was involved in BigBench in the earlier days. Maybe some of the other people in the team.…

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