High Signal Podcasts Evidence ledger
Method
Browse
← Back to evidence

Evidence receipt / evaluation

Published · transcript-backed

Shawn Wang: evaluation

31 Dec 2024 Latent Space Latent.Space 2024 Year in Review

“As someone who also does charts, XAI is continually snubbed because they don't work well with the benchmarking people.”

— Shawn Wang

Source trail

Everything needed to verify it.

Speaker
Shawn Wang
Attribution
Verified speaker
Claim type
evaluation
Recorded
31 Dec 2024
Publisher
Latent Space

Transcript context

…publish the rest. Yeah, we'll publish on this feed the best of 2024 in those domains. And hopefully people can benefit from the work that our speakers have done. But I think it's, uh, these are just good slides. And I've been, I've been looking for a sort of end of year recaps from, from people. The field has progressed a lot. You know, I think the max ELO in 2023 on LMSys used to be 1200 for LMSys ELOs. And now everyone is at least at, uh, 1275 in their ELOs, and this is across Gemini, Chadjibuti, [00:10:00] Grok, O1. ai, which with their E Large model, and Enthopic, of course. It's a very, very competitive race. There are multiple Frontier labs all racing, but there is a clear tier zero Frontier. And then there's like a tier one. It's like, I wish I had everything else. Tier zero is extremely competitive. It's effectively now three horse race between Gemini, uh, Anthropic and OpenAI. I would say that people are still holding out a candle for XAI. XAI, I think, for some reason, because their API was very slow to roll out, is not included in these metrics. So it's actually quite hard to put on there. As someone who also does charts, XAI is continually snubbed because they don't work well with the benchmarking people. Yeah, yeah, yeah. It's a little trivia for why XAI always gets ignored. The other thing is market share. So these are slides from Sarah. We have it up on the screen. It has gone from very heavily open AI. So we have some numbers and estimates. These are from RAMP. Estimates of open AI market share in [00:11:00] December 2023. And this is basically, what is it, GPT being 95 percent of production traffic. And I think if you correlate that with stuff that we asked. Harrison Chase on the LangChain episode, it was true. And then CLAUD 3 launched mid middle of this year. I think CLAUD 3 launched in March, CLAUD 3. 5 Sonnet was in June ish. And you can start seeing the market share shift towards opening, uh, towards that topic, uh, very, very aggressively. The more recent one is Gemini. So if I scroll down a little bit, this is an even more recent dataset. So RAM's dataset ends in September 2 2. 2024. Gemini has basically launched a price war at the low end, uh, with Gemini Flash, uh, being basically free for personal use. Like, I think people don't understand the free tier. It's something like a billion tokens per day. Unless you're trying to abuse it, you cannot really exhaust your free tier on Gemini. They're really trying to get you to use it. They know they're in like third place, um, fourth place, depending how you, how you count. And so they're going after [00:12:00] the Lower tier first, and then, you know, maybe the upper tier later, but yeah, Gemini Flash, according to OpenRouter, is now 50 percent of their OpenRouter requests. Obviously, these are the small requests. These are small, cheap requests that are mathematically going to be more. emini Flash, according to OpenRouter, is now 50 percent of their OpenRouter requests. Obviously, these are the small requests. These are small, cheap requests that are mathematically going to be more. The smart ones obviously are still going to OpenAI. But, you know, it's a very, very big shift in the market. Like basically 2023, 2022, To going into 2024 opening has gone from nine five market share to Yeah. Reasonably somewhere between 50 to 75 market share.…

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

Search evidence