Alessio Fanelli: evaluation

Transcript context

…Yeah, and I think this is the mark of a useful framework, that when a new thing comes along, you can break it down in terms of the four words and sort of slot it in or analyze it in those four frameworks, and have nothing left. So it's a MECE categorization. MECE is Mutually Exclusive and Collectively Exhaustive. And that's a really, really nice way to think about taxonomies and to create mental frameworks. So, what is Gemini 1. 5 Pro? It is the newest model that came out one week after Gemini 1. 0. Which is very interesting. They have not really commented on why. They released this the headline feature is that it has a 1 million token context window that is multi modal which means that you can put all sorts of video and audio And PDFs natively in there alongside of text and, you know, it's, it's at least 10 times longer than anything that OpenAI offers which is interesting. So it's great for prototyping and it has interesting discussions on whether it kills RAG. Yeah, no, I mean, we always talk about, you know, Long context is good, but you're getting charged per token. So, yeah, people love for you to use more tokens in the context. And RAG is better economics. But I think it all comes down to like how the price curves change, right? I think if anything, RAG's complexity goes up and up the more you use it, you know, because you have more data sources, more things you want to put in there. The token costs should go down over time, you know, if the model stays fixed. If people are happy with the model today. In two years, three years, it's just gonna cost a lot less, you know? So now it's like, why would I use RAG and like go through all of that? It's interesting. I think RAG is better cutting edge economics for LLMs. I think large context will be better long tail economics when you factor in the build cost of like managing a RAG pipeline. But yeah, the recall was like the most interesting thing because we've seen the, you know, You know, in the haystack things in the past, but apparently they have 100 percent recall on anything across the context window. At least they say nobody has used it. No, people have. Yeah so as far as, so, so what this needle in a haystack thing for people who aren't following as closely as us is that someone, I forget his name now someone created this needle in a haystack problem where you feed in a whole bunch of generated junk not junk, but just like, Generate a data and ask it to specifically retrieve something in that data, like one line in like a hundred thousand lines where it like has a specific fact and if it, if you get it, you're, you're good. And then he moves the needle around, like, you know, does it, does, does your ability to retrieve that vary if I put it at the start versus put it in the middle, put it at the end? And then you generate this like really nice chart. That, that kind of shows like it's recallability of a model. And he did that for GPT and, and Anthropic and showed that Anthropic did really, really poorly. And then Anthropic came back and said it was a skill issue, just add this like four, four magic words, and then, then it's magically all fixed. And obviously everybody laughed at that. But what Gemini came out with was, was that, yeah, we, we reproduced their, you know, haystack issue you know, test for Gemini, and it's good across all, all languages. All the one million token window, which is very interesting because usually for typical context extension methods like rope or yarn or, you know, anything like that, or alibi, it's lossy like by design it's lossy, usually for conversations that's fine because we are lossy when we talk to people but for superhuman intelligence, perfect memory across Very, very long context. It's very, very interesting for picking things up. And so the people who have been given the beta test for Gemini have been testing this. So what you do is you upload, let's say, all of Harry Potter and you change one fact in one sentence, somewhere in there, and you ask it to pick it up, and it does. So this is legit. We don't super know how, because this is, like, because it doesn't, yes, it's slow to inference, but it's not slow enough that it's, like, running. Five different systems in the background without telling you. Right. So it's something, it's something interesting that they haven't fully disclosed yet. The open source community has centered on this ring attention paper, which is created by your friend Matei Zaharia, and a couple other people. And it's a form of distributing the compute. I don't super understand, like, why, you know, doing, calculating, like, the fee for networking and attention. In block wise fashion and distributing it makes it so good at recall. I don't think they have any answer to that. The only thing that Ring of Tension is really focused on is basically infinite context.…

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