Alessio Fanelli: evaluation

Transcript context

…it's good. Okay. I feel like we've, we've covered a lot of stuff. Uh, I, yeah, I, you know, I think We will go over the individual, uh, talks in a separate episode. Uh, I don't want to take too much time with, uh, this stuff, but that suffice to say that there is a lot of progress in each field. Uh, we covered vision. Basically this is all like the audience voting for what they wanted. And then I just invited the best people I could find in each audience, especially agents. Um, Graham, who I talked to at ICML in Vienna, he is currently still number one. It's very hard to stay on top of SweetBench. OpenHand is currently still number one. switchbench full, which is the hardest one. He had very good thoughts on agents, which I, which I'll highlight for people. Everyone is saying 2025 is the year of agents, just like they said last year. And, uh, but he had [00:20:00] thoughts on like eight parts of what are the frontier problems to solve in agents. And so I'll highlight that talk as well. Yeah. The number six, which is the Hacken agents learn more about the environment, has been a Super interesting to us as well, just to think through, because, yeah, how do you put an agent in an enterprise where most things in an enterprise have never been public, you know, a lot of the tooling, like the code bases and things like that. So, yeah, there's not indexing and reg. Well, yeah, but it's more like. You can't really rag things that are not documented. But people know them based on how they've been doing it. You know, so I think there's almost this like, you know, Oh, institutional knowledge. Yeah, the boring word is kind of like a business process extraction. Yeah yeah, I see. It's like, how do you actually understand how these things are done? I see. Um, and I think today the, the problem is that, Yeah, the agents are, that most people are building are good at following instruction, but are not as good as like extracting them from you. Um, so I think that will be a big unlock just to touch quickly on the Jeff Dean thing. I thought it was pretty, I mean, we'll link it in the, in the things, but. I think the main [00:21:00] focus was like, how do you use ML to optimize the systems instead of just focusing on ML to do something else? Yeah, I think speculative decoding, we had, you know, Eugene from RWKB on the podcast before, like he's doing a lot of that with Fetterless AI. Everyone is. I would say it's the norm. I'm a little bit uncomfortable with how much it costs, because it does use more of the GPU per call. But because everyone is so keen on fast inference, then yeah, makes sense.…

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