Evidence receipt / evaluation
Published · transcript-backedShawn Wang: evaluation
9 Mar 2024 Latent Space Top 5 Research Trends + OpenAI Sora, Google Gemini, Groq Math (Jan-Feb 2024 Audio Recap) + Latent Space Anniversary with Lindy.ai, RWKV, Pixee, Julius.ai, Listener Q&A!
“I think that I ki I had to feature this because the, one of the top news of January was that Gemini Pro beat GPT-4 turbo on LM sis for the number two slot to GPT-4.”
Source trail
Everything needed to verify it.
- Speaker
- Shawn Wang
- Attribution
- Verified speaker
- Claim type
- evaluation
- Recorded
- 9 Mar 2024
- Publisher
- Latent Space
Transcript context
…There's a lot of them. And just to wrap these up, online LLMs? Yeah, I think that I ki I had to feature this because the, one of the top news of January was that Gemini Pro beat GPT-4 turbo on LM sis for the number two slot to GPT-4. And everyone was very surprised. Like, how does Gemini do that? Surprise, surprise, they added Google search. Mm-hmm to the results. So it became an online quote unquote online LLM and not an offline LLM. Therefore, it's much better at answering recent questions, which people like. There's an emerging set of table stakes features after you pre train something. So after you pre train something, you should have the chat tuned version of it, or the instruct tuned version of it, however you choose to call it. You should have the JSON and function calling version of it. Structured output, the term that you don't like. You should have the online version of it. These are all like table stakes variants, that you should do when you offer a base LLM, or you train a base LLM. And I think online is just like, There, it's important. I think companies like Perplexity, and even Exa, formerly Metaphor, you know, are rising to offer that search needs. And it's kind of like, they're just necessary parts of a system. When you have RAG for internal knowledge, and then you have, you know, Online search for external knowledge, like things that you don't know yet? Mm-Hmm. . And it seems like it's, it's one of many tools. I feel like I may be underestimating this, but I'm just gonna put it out there that I, I think it has some, some potential. One of the evidence points that it doesn't actually matter that much is that Perplexity has a, has had online LMS for three months now and it performs, doesn't perform great. Mm-Hmm. on, on lms, it's like number 30 or something. So it's like, okay. You know, like. It's, it's, it helps, but it doesn't give you a giant, giant boost. I feel like a lot of stuff I do with LLMs doesn't need to be online. So I'm always wondering, again, going back to like state of the art, right? It's like state of the art for who and for what. It's really, I think online LLMs are going to be, State of the art for, you know, news related activity that you need to do. Like, you're like, you know, social media, right? It's like, you want to have all the latest stuff, but coding, science,…
Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.