Eugene Cheah: evaluation

Transcript context

…Next I think we have Eugene Chia from RWKV previous guest. I'm going to present about the RWKV/Eagle project. So, Eon Transformers. There's been a lot of excitement lately. And, and, like one AI year ago apparently when we launched our 7B AI model, there was a lot of excitement in the buzz, because for the first time, an attention free model beat other transformer models at one trillion tokens at a 7B class. And if everyone's been playing open source AI, you know 7B class is one of the best. Most important class 'cause it's the ones that works on most devices, laptops, and everyone's been playing around a bit. And the excitement is compounded by the fact that we even showed that even with 300 million tokens and a few that we perform similarly, transformers, that means people are projecting is what happens if we train another 1 trillion? Will we match or can we go beyond that? And, and it also spurs up questions beyond actually our architecture itself. It's spurs up questions that. Maybe what we need is good data and an efficient architecture, not just RWKB, it could be beyond that. And that's what caught the attention for a lot of folks, even yeah. And why we do very different is that our architecture scales linearly. So, we are in this space together with Mamba and a few other architecture where we are trying to build the next architecture to, that can scale much larger for, for everyone. But, and we share that with Mamba because we believe that attention is not all you need, and it's like, it's been a running bet right now. We are the strongest evidence to date. But sometimes, like, talking about scale, right, sometimes we get lost in numbers. Because, like, I can show this chart. The last time I showed this at a Linear Transformer event, which only 8 people took pictures of it and understood what it means. And they were all from either Google or Facebook. Because, like, what it says here, right, is that We are able to run run on a single GPU with one model, 256 on a single 4090, or a thousand concurrent users. But, to put that into contrast, right, what that transformers typically handle 8 or 16 concurrent requests per GPU. We're talking about 256 or a thousand, many orders of magnitude higher. And all we're sustaining at NeoChat GP speed. And so I sometimes like, like, sometimes when I get lost in these words, these days I'm actually trying to step back into like, Why are we doing this for our group, for our organization? And, and this, and, and some, and for us right, we are actually making the AI model for everyone in the world. And in every country, in every language. So, what does it take to make an AI for the world? Apparently some folks think it's 7 trillion dollars. But, I think 7 trillion is a bit too much. Like, what's going to happen to half of the world that doesn't even have a trillion dollars? Yeah, so I want AI to be accessible at scale. trillion dollars. But, I think 7 trillion is a bit too much. Like, what's going to happen to half of the world that doesn't even have a trillion dollars? Yeah, so I want AI to be accessible at scale. So, apparently ChatGPT produced, or OpenAI produced 100 billion words per day. That's 3. 4 million tokens per second. No one has the exact numbers, but it's typically 50k, H100s and above remote, like these are some old numbers, like the numbers have gone way beyond this, apparently. But, with our architecture, for a 7B model, that's just a thousand GPUs, or ten thousand GPUs for a 70B model. We're talking about one data center to handle all of OpenAI's workload. And if we want AI agents everywhere, cheaper, at a much larger scale, we need to be thinking about that fundamental shift. Because it's not just about who can it's not just about you can afford it in the US, it's about everyone else in the world. And that brings us to the second advantage of our model, which is not even architecture. Because we are accessible by language. We apparently beat Mistro and everyone else in Mountain Lingo, but that's not because our architecture is better, but because we're an open source team that came from all around the world and wanted our model to work for our mom and grandma. That was the real reason, and we We iterated and refined the data accordingly. We created a custom tokenizer that supports all languages, not just English. And sometimes in the race for the English benchmark, because one of the reasons why other models don't perform as well in multilingual, is because the truth is, if you add multilingual, you hurt your English eval. But, who are we building the AI for? Are we building it for our evals? Or are we building it for the people to use? And, and, even in evals, my frustration is, we trained on 100 languages, I only got 23 languages for evals. Like, where's everything else? So, where are we now? Just like I mentioned 1. 1 trillion, that's where we are, we are in between the 1. 5 trillion and the 1 trillion models for for all, all the, all the English models benchmarks. And, yeah, zooming in further, it just shows that we have more room to go. And, for me, like, The emphasis on English is weird because only 70 percent of the world speaks English, but we are here for the 83%. That's for us. If you all want to get the best English model, sure, it may not be true for us, but we are here for everyone else. And, yeah, and a lot, a lot, the launch of that model, I think what was the biggest feedback I had, was not that it was a linear transformer, was that it can run on their own. Laptops. Some people even ran it on a Raspberry Pi, very slowly. And it supported their language, which was more exciting because that's more important for most people. And I think the last one that I've recently like heard that was unique for us and is a lot more important is that ultimately this model is owned by everyone because We put it into the Linux Foundation.…

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