Pod Insights a16z Podcast Beyond Leaderboards: LMArena’s Mission to Make AI Reliable About this episode a16z general partner Anjney Midha sits down with LMArena cofounders Anastasios N. Angelopoulos, Wei-Lin Chiang, and Ion Stoica to talk about the future of AI evaluation.
As benchmarks struggle to keep up with the pace of real-world deployment, LMArena is reframing the problem: what if the best way to test AI models is to put them in front of millions of users and let them vote? The team discusses how Arena evolved from a research side project into a key part of the AI stack, why fresh and subjective data is crucial for reliability, and what it means to build a CI/CD pipeline for large models.
They also explore:
- Why expert-only benchmarks are no longer enough
- How user preferences reveal model capabilities — and their limits
- What it takes to build personalized leaderboards and evaluation SDKs
- And why real-time testing is foundational for mission-critical AI
Chapters:
00:00:04 - LLM evaluation: From consumer chatbots to mission-critical systems
00:06:04 - Style and substance: Crowdsourcing expertise
00:18:51 - Building immunity to overfitting and gaming the system
00:29:49 - The roots of LMArena
00:41:29 - Proving the value of academic AI research
00:48:28 - Scaling LMArena and starting a company
00:59:59 - Benchmarks, evaluations, and the value of ranking LLMs
01:12:13 - The challenges of measuring AI reliability
01:17:57 - Expanding beyond binary rankings as models evolve
01:28:07 - A leaderboard for each prompt
01:31:28 - The LMArena roadmap
01:34:29 - The importance of open source and openness
01:43:10 - Adapting to agents (and other AI evolutions)
Transcript 18,078 words · 4 speakers
0:08 Speaker 3 It sounds like what arena is is humanity is real-time exam.
0:12 Speaker 2 Yeah,
0:13 Speaker 3 that's
0:13 Speaker 1 a good way. That's a very well proof.
0:15 Speaker 3 Sometimes I get asked, what's the last exam that that AI should take, you know, for humanity? And I it seems like that's the wrong question to ask. We should be asking what's the real-time exam you want your AIs to be taking before they get deployed every hour, every second of the day, especially as we start to get like I think one of the things we uh that's emerging for me is that one of the arenas misunderstood partly because we're just early in AI. And so while benchmarks like MMLU and the idea of these static exams were useful three years ago, the future is about real-time evaluation, real-time systems, real-time testing in the wild. Now, one thing that that concerns a lot of people is the reliability of these systems when when we start going from chat bots that are good at let's say companionship and you know, more consumer use cases to mission critical systems, defense, healthcare, financial services.
1:10 Speaker 3 How will Arena have to evolve as we go beyond companionship or or web dev to those kinds of mission critical use cases?
1:17 Speaker 2 I think that's one of the very reason we wanted to create a company to support this project to further scale the platform. Right. So right now we are at million monster user now. What if we scale it to five to 10 or even more to capture even more diverse user base across different industries? And then in that case, we'll have the ability to really like zoom in into all these different kind of like areas that people really care about that for critical mission task that
1:54 Speaker 1 it will be used to. You can imagine we are going to when we are going to scale, we can have micro size for nuclear physicist, radiologists and so forth, right? And these experts are going to come there to get the best answers to their again, research questions.
2:13 Speaker 3 Uh so that's interesting. Uh is there is there a future where now that arena is becoming a company, you could see a uh uh a scientific lab or a shipping company or a defense company deploy their own arena on their own infrastructure for their own users for sure. Own prompts.
2:31 Speaker 0 Many people have asked us for this already.
2:33 Speaker 3 So these would be sort of private arenas. Yeah.
2:36 Speaker 0 And it's worth saying, you know, I think when people have these mission critical industries in mind, they often are thinking about the sort of like factual nature of the responses and so on and so forth. But in reality, even in such industries, the majority of questions that people ask are subjective. Okay. So the the mythology that in hard sciences or in mission critical industries, people just have like cut and dried questions and they just need like a retrieval and a lookup. That's completely false. That's the very reason why these models are useful. Is because they allow you to sort of like interpolate between these weird questions and answer questions that are not fully specified and give responses that are sort of like geared to answer the question, but might not have a fully factual basis, right?
3:23 Speaker 0 And they might incorporate factual elements through RAG, let's say, but there's a subjective nature to the response. And that's a reality that everyone's gonna have to live with. If these systems are going to be deployed in medicine and defense and so on and so forth, they're gonna be deployed in places where the data is messy, because that's where they're useful. Okay. Given that fact, how are you gonna make sure that they're reliable? Well, you need something like Arena.
3:52 Speaker 3 You know, at this point, arena is hard to miss in the AI space. You know, whether it's Grok 3 coming out and Elon putting it up for the bulk of the keynote or you know, Demis uh using the WebW Arena scores to kind of demonstrate how good Gemini is. It it's sort of become the standard bearer for for evaluation and testing at all the big labs, right? But it does that mean you guys have been helping them more than open open source labs
4:17 Speaker 0 or smaller labs? Aaron Powell No, we work with model providers small and large. Okay. So we work with basically anybody who wants to work with us, you know, within our constraints, try to be as like helpful as possible. The fact is that you know, part of the reason to build the company is because we don't want Waylin having to serve all of these requests from people like manually himself. So it's it's been a challenge. But we try to scale as much as possible. And in fact, one of the things that we help everybody to do is pre-release testing of their models. Okay. So it's not just that you know we work together to evaluate the models or release, but we also try to be their release partners and say, hey, can we help you guys pick the models that do best on our user base and use that as a a guideline for which models they should actually release to the world? So that's the way our platform works. And that's getting us closer to something, you know, as we've talked about, like reliability is so important, these subjective measurements and so are so important.
5:09 Speaker 0 How are we going to get to a world where there's a CI CD pipeline where people can, you know, test their models pre-release and make sure that they're doing well for all sorts of different diverse people. Well, you need something like arena to do that. And that's part of what the company is geared to do. So what people do is they can come, they can test a bunch of different models. We do this with basically every provider that comes to us, and they can see, oh, which one's doing better or worse on the distribution of arena users. And then they can use that information to help them decide which model release. After they decide to release a model, it gets continually evaluated forever. So that's where you get the freshness of the data that the model, you know, continues to be tested. And you know, we're pushing towards a world where these subjective and human uh considerations are part of model developers' final release pipeline.
5:56 Speaker 3 So if I'm hearing you right, the more testing, the more reliable we should expect AI systems to get. So we should be
6:03 Speaker 1 like with any software system.
6:08 Speaker 3 This is one of the fundamental debates in the space, right? Is the what is the measure of progress, the right measure of progress in the space? And and there's a body of uh work that tries to create exams. And they call them harder and harder exams. And for two reasons, I find I've always found Elam Arena interesting. One, it's it's the opposite approach, which says let the wisdom of the crowd guide us. And two, let open source actually define the examination, right? I I I think this is quite important where if you get a group of experts in a room who decide what the right exam is for humanity, inevitably, then if it turns out that group's values get encoded in that that we have no we have no way for the rest of the world to use AI systems on a different that that that are measured by a different
7:03 Speaker 0 set of values. So it's great that people do these expert evaluations. It's totally fine. It's orthogonal from what we do. You know, I'm glad we haven't. But at the same time, you have to ask yourself, what makes somebody an expert? Right? What are they an expert on? I think the whole world is moving in the direction against like experts being the be all end all of everything. Right. And you know, everybody actually has their own opinions and everybody has their own point of view. And in fact, there's so many natural experts in the world on all sorts of topics that they don't necessarily need a PhD in order to be really, really intelligent and high taste and have valuable opinions. Right. And I think that's one of the things I'm proud of with Arena is it allows us to actually go and say, hey, where are the natural experts?
7:50 Speaker 0 And actually, can we find data-driven ways to identify them? What if we can go look and say, hey, this person here in this random part of the world is actually incredible at coding and math.
8:01 Speaker 1 Right.
8:02 Speaker 0 Let's I their vote actually means so much. And their preferences are able to guide the future of AI. That's that's an amazing thing. And we hope to be able to scale it further.
8:15 Speaker 3 Okay, but I'm going, I'm going to push back a little bit on this. I'm going to channel a few of the criticisms I've heard. Yeah,
8:20 Speaker 0 please. You know
8:22 Speaker 3 from experts, which is that look, if you're an expert, you've been blessed with the brains, the resources, and so on to be a highly educated individual in your field. We have a responsibility to guide humanity. It's our job to actually guide the masses. We should be defining what good human preferences is versus not, because the masses actually don't know what they're was good for them. Um the everyday users, the lay person uh prefers slop. Right? We've heard these arguments. Yeah. Is there a grain of truth or or are or how do you think about it?
9:04 Speaker 1 So I I think a few a few things here. So one, like Anastasia said, it's you know, this alternative to have hard exam to create hard exam is valuable. No question about that. I think the other thing I want to point out. And I took the at heart is this kind of criticism and about kind of having expert labeling.
9:27 Speaker 3 Right.
9:28 Speaker 1 Now I'm I I want to quite a few experts I know and I respect. And ask them, would you label it? Would you label? Almost everyone told me, no, I don't have time. Right. So there is a question there about are you going to get really the experts, right? Right. And I no, no. I I I I I don't think so. Right? You're getting some people from that area who are willing to do the labeling, but the best people. Not willing. But fundamentally they don't have time, right? Now, but these people, if we offer to them, and we are not doing now, I'm talking about the future, a platform where their community come there and ask questions to put the boundaries to help with their research and things like that.
10:16 Speaker 1 They're going to be first of all, you are going to get these people, right? Because they do that in order to advance again in their research. And we are going to get also their votes. So fundamentally I do think that we can get in arena. Again, right now, I'm not talking about today. I'm talking about the future. It even happens to the even happens. It even happens today. So you get real experts, right? The top ones. Okay, unhirable people, unhirable people. That's that's a great way to say it. The other thing I want to say about the layman and so forth, you know, and you you fund a lot of company. You are funding a lot of companies, the AI and A16Z, and you are on the board of many companies.
11:05 Speaker 1 What do they build it? What do these companies build? What is their product? Who uses their product?
11:13 Speaker 0 Right.
11:14 Speaker 1 Right? It's not the top experts. He's a layman, right? Right. These are their users. That's how OpenAI makes money and so forth, many others, right? So then wouldn't the evaluation should take into account the preference of these users. Right. Right? The answer is obviously yes. Right. So how can something, again, like these exams are very important to understand the capabilities of these models, these benchmarks. Again, no question about that. But they're not going to reflect MMLU is not probably as good as reflecting the preferences of the users of these AI products
12:01 Speaker 0 and services. I want to just dig one level deeper here, which is that one of the things that we are really excited about is the question of why do people vote the way they do. Who's voting left-right? Why are they doing it? On what kind of prompts do they vote for one model or another? On what kind of topics or models better? You know, basically, can we decompose human preference into its constituent components? Let's say you have a criticism, you say people vote based on slop. You know, emojis are driving votes and response length is there's a huge response like bias, which is you know, it's true that people vote for longer responses, you know, preferentially over shorter responses, even given the same contents or well-known human bias. This gets back into RL.
12:48 Speaker 0 Can we learn this bias and actually adjust for it and correct for it? And the answer is yes. That's why we're making style control default. So what we developed is this method called style control, which allows you to run not just, let's say a Bradley Terror aggression, but also include certain covariates that model the effect of style and sentiment on how people vote. And what you get when you fit this model is not just a prediction of preference, but also an understanding of why people are voting the way they do. You know, it's we're we're trying to target a causal quantity, which is you know the causal effect of let's say response length or sentiment and so on and so forth. Okay, you know, there's always more work to do to get things closer to the actual causal estimate that we want.
13:35 Speaker 0 But if we continue to decompose human preference into its constituent components, what we're building is an ever-richer evaluation that can tell us all the factors that go into response, how you can optimize people's preferences, you know, keeping style fixed. Let's say I want to remain concise, but I want to maximize your preferences. Okay, how am I supposed to do that? Well, you know, not a lot of people have that information, but we're building the methodology that allows you to do that.
14:05 Speaker 1 Okay.
14:06 Speaker 0 And the question is, can we disentangle style versus substance? That's it. You can you you there is an effect and you want to know about it. Period. Right. The platform helps you and what is the impact. And what is the impact?
14:18 Speaker 3 Yeah. So this is this is another, I think, important area to dig into. There was a moment where you guys decided that it was insufficient to keep measuring the progress of the models on coding with the base design of the platform, which is just chatbot arena. And I remember seeing a launch which was web dev, I think you called it web dev arena and showing up to a completely different interface. And then realizing that this was a pretty big change for you guys, right? What why did it why why would we need a new kind of arena to correct for that effect? Why was that necessary?
15:05 Speaker 2 I think that comes back to I think Jan's point, oh but like when you build a product, an AI product, you want to know about how people use it. You want to understand why you prefer this over that. And in order to collect that kind of data, you have to build sort of like a product first. Right. And then, as we know, over the past few years, people have been building very different kinds of applications on top of AI, beyond just chatbot.
15:39 Speaker 3 Right. So
15:40 Speaker 2 Chatpa is one of the widely used interface right now
15:45 Speaker 3 for humans to interface with AI.
15:47 Speaker 2 But these days, people are applying these models to coding, and then more like tool use, agenda behavior, that kind of stuff, right? So that's the first step. Where we were thinking like, okay, how can we capture all these use cases? Right. And then the answer to that is that we have to build something that people can use again in the same environment that people can test in real time to give us real world feedback. So that was the original like idea that was kind of like last summer, just at the beginning of this kind of text-to-web, text-to-app trend. The beginning was very beginning was cloud, cloud uh artifacts. That was that was the first.
16:33 Speaker 2 Right. And that we saw that we were amazed by that. And then how can we build evil for that? Right.
16:41 Speaker 0 We have to give credit to Arian on the
16:42 Speaker 2 team. Yeah, and then Arian basically was like just joined the team, and then he was interning actually at uh Verseel. So we were saying, okay, why don't we build something new? And a time, cloud effect. How can we be evaluation for that kind of applications, text-to-back? And then that's was the idea.
17:00 Speaker 1 One small parenthesis. So far, we we talked the three of us, but eventually the team grew quite a bit. And right now, I don't know, it's like almost like 20 people and both, you know, graduate and especially undergraduate students and doing a lot of uh very exciting and interesting work to expand the uh abilities and capabilities and the reach of uh chatbot arena. Right. So I just want to make sure that other colleagues of mine are involved here and uh like Joey and others and uh I think the credit goes much more beyond the serious. But that's what we want to provide.
17:45 Speaker 1 You want to, you know, whatever you want. We hope that you'll find the answer
17:52 Speaker 0 to. Fundamentally, there's if you're looking at the leaderboard, there's only one thing really that matters, which is do you care about the preferences of the community people that come to vote on our platform? That's it. That's that's what we measure. It's the only thing that we claim to measure. We don't claim to be an AGI benchmark. We are faithfully representing the preferences of our community. So that's why it's so important to us that we continue to grow our community, and we get a diverse community of all different people, experts, non-experts, artists, scientists, different languages, you know, different languages, everybody under the sun. We want come into this platform to express their preferences. Because if we can get that to happen, you know, it already happens to some extent.
18:38 Speaker 0 But if we can continue to grow it, what's going to happen is again, in order for a model to do well, what needs to happen is new people need to come in and vote for it. And so if we can provide this lens into the preferences of the world.
18:55 Speaker 1 Things are moving, right? It's like this are changing. This is again, it's not like it's like we we have a the study, and you want you want to talk about about the freshness, right? Right? About we all all, you know, see fresh prompts, right? It's not like, oh, we are going to see all these fronts over and over again, like kind of saturated. That's not the case.
19:18 Speaker 2 That was the very beginning of why we believe arena is fundamentally different.
19:23 Speaker 1 Related to this contamination.
19:25 Speaker 2 The very beginning of arena is to try to solve the contamination problem, uh the overfeeding problem, right? Which is like people test models on a static benchmark.
19:37 Speaker 1 Or what people call overfitting. Yeah, what people call overfeeding.
19:40 Speaker 2 So how do you overcome overfeeding? You collect new data. Right. That's how you overcome overview. And an arena is designed to collect new data every second. So all the questions are as are new. All the votes are new. And then we measure basically how different, what's the difference? Between all these prompts, right? What's a distribution look I and so on. And then we conservatively estimate like, you know, over 80%, something like that problem. Yeah. Yeah. Yeah.
20:13 Speaker 1 Correct. If I and this study was done by another member of our team, Lisa. Um and basically major is out how many more, you know, fresh prompts you have in one day compared to what you've seen in the past three months. Right. And by a similarity score of something like 70, 75%, you have over 70 of these uh prompts are fresh. 75, 80%. Or over 80%. Right. So it's a large number of prompts are fresh. Right. And when you say fresh, we are talking about similarities, but we are not the test is not very high. That it's not like, oh, is that identical?
20:53 Speaker 0 No. Yeah, just to dig in one click deeper into what we lyn said. You know, what is overfitting? Static benchmarks overfit. Why? It's because, as Jan said earlier, you're giving the student the same test over and over. You have a model, you test it, you know, you look at whether or not it's improved on a static data set, then you find another model, you test it, and you pick the one that does better and better and better. And what ends up happening is that the test becomes meaningless. Because you've seen it so many times that you've memorized the answers. That's what overfitting is. Chatbot Arena is immune from overfitting by design. Because you're always getting fresh questions. In order to do well on the arena, new users need to come and vote for your model. That's it. That means that users like
21:39 Speaker 3 it. One thing I've noticed is that the same researchers who often argue with me that arena is a terrible evaluation system. The leaderboard is not to be trusted. The scores are rigged, they're gameable, have one tended to also celebrate when they're on top of the leaderboard. And the second is I find actually increasingly, especially for specialist arenas like web dev. You know, there's a natural tendency to just accept that this is a really good indicator of the underlying capabilities. Why is that? Why is Web Dev Arena such a good proxy for actual performance improvements when it comes to a capability like coding, which is quite general purpose, right?
22:26 Speaker 3 It's very counterintuitive. It's programming is actually a very general purpose discipline and skill. And yet it seems like the capabilities on that on a in a very general way are still being able to are captured well in a specialist arena like Web Dev Arena. Why is that?
22:45 Speaker 0 Yeah, first let me just say I think that all of these arenas have signal in them. It's not just web dev arena. It's just people have more opinions on language. Right. So Web Dev is a little bit more objective. It's easier to say the website you build it, and it's, you know, this one's better than the other. There's a lot of signal in it. Also, it shatters the models. And what I mean by that is um it can be very clear that one model is way better than another on web arena immediately. Bam, you see it, and it's like why is that? Just the capabilities of the models. Maybe Willin can say more.
23:16 Speaker 2 I think it's just much, much harder task. Because it's like from text, a description of a website. And you have to, you know, first understand the request, and then build like write code. Right. And then the code has to fit, maybe satisfy certain requirements that say uh style requirement or like component, that kind of stuff. And then it has to compile, right? It has to be we we we basically run it live in browser and then with connect to a sandbox, that kind of stuff. So there are a lot of like parts you the model has to get right in order to build a website that people can you know really interact with.
23:57 Speaker 1 So fundamentally discriminates much better uh across some models because it's a much harder problem. Yeah. So very few get get it right.
24:07 Speaker 3 So it's in reality for critics who say, you know, these uh exam static exams are hard tests, and arena is an easy-to-gain benchmark. In fact, Web Dev Arena isn't a great example that shows it it's a really hard evaluation, it's a hard test, actually. It just proxies the real world better than some static multiple choice question test. Is that is that roughly right?
24:31 Speaker 2 Yeah, for sure. And then a lot of like, you know, every input from user is for a real world task. Right. They are trying to build some website real. And then that also measures something that's like beyond just you know, academic benchmark that's we imagine one user would do, right? This is like really trying to approximate the user intent, user preferences directly.
24:56 Speaker 0 I have to say I also just like completely disagree with the foundation of the question. The like implicit assumption is that like chat is easy or that it's even easier than web dev. That's completely false. It's a completely naive perspective that people have on this. Because it's it's hard to build something that people love. People are good at chat, but you like some people way better than others. It's subjective, and everybody has their own opinions. And the landscape is very, very rich. You might like a very different model than me. Somebody else, you know, let's say a musician might like a different model, much you know, a much different model than I do. And understanding all of those differences is really hard. And anybody who thinks that that is gameable is deluding themselves.
25:43 Speaker 3 To take that to its strongest form argument, then isn't the right way to allow me to evaluate whether the model is good or not to allow me to generate my own leaderboard.
25:54 Speaker 0 Uh you as a person? Absolutely. Yeah, and we we should be giving you the tools to do that, and we're currently building them.
26:01 Speaker 3 Ah, and so this is quite profound. You see the world going where everybody has their own personal arena? Absolutely.
26:06 Speaker 1 Absolutely.
26:07 Speaker 0 It should be personalized, just for you. You should understand which models are best for you.
26:10 Speaker 1 And is going to be for your task, his person and task, right? Because for a different, if you enter the different things, you may have a different leaderboard, right? If you have a question about tax today, you go to different people than if you have a question about programming or whatever, right? So that's kind of also is going to depend on what task you want to accomplish, right? And I really want to go back to one thing. I think that because it's, you know, I try to think uh quite a bit about all the criticies, because on the face of it, intuitively for some people makes sense, right? That's how many people make the same criticized. And I think there is another thing going on. We as a humans, we believe that why do people say arena is not good because people are fooled,
27:01 Speaker 0 right?
27:01 Speaker 1 Right? That's fundamentally their kind of argument. They're fooled because long answers, more emoji and things like that. And when I look from that perspective as a human, what I in my mind is that I am not going to be fooled. Right. That's what it's not relevant. These guys are going to be fooled by these things. No, I am not. That's why it's not good. That's why it's not a good proxy. The problem is that you are we I am fault, right? Everyone is fault. And this is actually that's why the chatbot arena, it's like Anastasios said, it's it's it's a proxies provides a magnifying glass. But you know, all of us, we have our own peculiarities, our own, you know, culture, our own, you know, history built on interaction with different people, right?
27:56 Speaker 1 So we have different preferences. Right. That's fundamentally what it is, right? And these preferences are not fully objective, right? Because people say it's only on objective answers, but all of us are different, right? So that's kind of the fundamental disconnect between the criticism and actually what you provide, and actually what everyone we believe needs, right?
28:24 Speaker 3 To double-click on that a little bit, you you said we we all have our own culture, right? Um and Ben Ben Horwitz, uh, we all know well, has a quote I love, which is he says, culture is not a set of beliefs, it's a set of actions. Right. So let's say our belief, the philosophy that Arena has is that the set of actions a user takes when using an AI model is the best source of truth for whether that model is good or not for them, versus some third-party closed source evaluation. Telling you what expertise
28:55 Speaker 1 is telling you what is good for you. Right. And that's why again, it's like going back to the to the to the to the previous that, you know, I think what you hear is that capture capturing the human preference is fundamental because we are building these AIs to interact with humans, right? I think that's kind of the foundation. But if you like we have in the previous discussion, say people say, well, you know, like, yeah, but this other people are fooled. This is not me, right? Although I am fooled as well, but I can't believe, right? You always believe that about yourself that you are better than you are. But but but then we can provide like style control. Okay, so far, you can remove that. I adjust for that effect, right?
29:40 Speaker 1 Right. And we are going to provide more and more. You can adjust for that effect and that effect and that effect, right? So you get your answers as well.
29:54 Speaker 3 LM Arena started as a research project to take us back to how it began. So
30:00 Speaker 2 It was started around two years ago, in late April 2020. And then at that time, before Arena, we were working on project called Bikuna, which is like one of the first open models that's been released, like the ChatGPT kind of claw. Yeah. And at that time, Lama One was just released, which is a base model. It doesn't really know how to chat with human. Only gone through like pre-training process. And that time I don't think people call this post-training yet. People call it instruction fine-tune that kind of stuff. So we were like in the lab exploring how do we reproduce this, right? How would how do we make an open source version of ChatGPT? And then, you know, credits to Liam Min.
30:47 Speaker 2 He was having kind of like an idea that we could use some of the open data published on the internet. That's kind of like user Chat GPT conversation. It was called shared GPT. And it was like kind of like a high quality set of dialogues that users share. So basically we come as a group, a bunch of PhD students in the lab, set an impettion goal, which is like we try to, you know, release this model, trend this model in two weeks, that kind of stuff. So and then during that process, the result was surprisingly. At that time we were like playing with the model, and then we thought we got out, you know, demo this to the world. So we basically just set up a website and then put that model on the website and then release it.
31:37 Speaker 2 At that time, it was like there's like a huge debate, like internally. Like how should we when we release it, you know, how should we evaluate this model? How good is this model really? It violent, right? Because when we compare it to Lama, the base one, you can feel the difference, right? The model just learn how to chat, learn how to speak like a chat GPT. So there was a huge debate, like how how do we evaluate this? At that time we didn't have much time. So we were like, okay, we either do this kind of like labeling, come up with questions ourselves and label the data and then compare it with other models. Or we do something like something automatic.
32:22 Speaker 2 Uh and then at that time there was like GPT-4 just came out in March. People were like wondering, what can you do? Right. And then we were like, okay, why don't we just use it for evaluation? As we use GPT for as a judge to do this automatic e-value. At that time, no one believed it. And then it was like again, huge debate, but we didn't have time. So we just you know ended up doing it. And it works surprisingly well again, and we release it. But the the huge problem after that is like still, there's an open problem, which is like how do we evaluate these chatbots, right? So soon after we um kind of like come up with that idea that why don't we let everyone in a community vote which model is better because at the time we serve that model, and then we also serve some of the other open source model at a time.
33:17 Speaker 2 At a time every week, there's a new fine team. So we register a website, we demo them, all of them, and then we come up with like kind of like a side-by-side UI that people can compare them. And then soon after we say, okay, why don't we come up with a battle mode, which is like we anonymize the identity. And that people vote. So that was the origin of Arena, which is basically trying to solve our problem, which is how do we evaluate these models, understand these differences.
33:51 Speaker 1 Actually, the first time we did start, we did try to have some students, you know, buy some pizza, get them in a room and label uh the replies from um uh Vikunia and other model to compare them. And then um obviously that didn't scale. And um then it was this LLM as a judge we tried and it worked surprisingly well. And then it was quite a bit of a debate. Okay, we are going to build a platform to scale because the question is still okay. It seems that anecdotally anecdotally GPT-4. It was released just two weeks before, or before we started to use it as a judge. Um still the question, okay, yeah, it seems that it's doing well, but still, you know, it's like how about how does it compare with humans, right?
34:39 Speaker 1 And that was a question about okay, how do we scale as a human evaluation? And we discussed quite a bit about how to do it because it was not clear because if you think about before, you just ask people, I have a prompt, and the prompt is answered by all models, and then you label it, right? Good and bad. That's kind of the typical ways of how would you scale up the process. Then you need to kind of to rank them, right? And you have N choices, answers of the same prompt from N models. It's very hard to rank them, right? It's like think about, right? We get slightly different in tone and so forth, try to rank them. And then uh we saw quite a bit, I think the inspiration about was how humans in real life uh rate, say, players or teams in games, right?
35:30 Speaker 1 And obviously you have the tournament that's only to do it, where you have so to speak, the players plays with each other uh head to head. And then based on that, you are going to have some number of points to win or lose or or to and or die. And then you you are going to have a leaderboard, right? Uh and um but that again, the problem is that is that if you have a tournament, typically the assumption in a tournament is that the number of players do not change across during the tournament. Right, right. And also in general, you know, most of the tournament, this player has to play with everyone, so it's kind of n-square problem or any of the number of players. And then we thought about okay, there are other ways uh in real life, how players or teams are ranked when they don't play with each other.
36:24 Speaker 1 They don't have a chance either because of number of players is too large, or you know, you need to also accommodate new players entering the game. And that's why it was about then we thought about okay, there are disciplines where this is done, like chess, that tennis, ATP rating, and many others. And that was the idea, and we start, okay, why don't you do something like uh ELO score? Okay, and for that, what do you need? Well, you need only uh, you know, head-to-head, and um you don't not everyone needs to play in the same tournament. And uh that's how you adopt it, and that's why arena has this battle mode in which you have a prompt and answer from two randomized anonymized light language models, and you can pick which one is better or the restide and and so forth.
37:17 Speaker 3 And when was the moment when you joined the conversation and brought the from what I understand, you know, the Bradley Terry uh approach.
37:28 Speaker 2 Yeah,
37:28 Speaker 3 it
37:29 Speaker 2 turned out it was like deeper technically than we thought. So at that time, Young was like, we need to find someone to back this up, or like on the theory side to have a solid foundation to rank all these models. It's not at that time it was no longer like really like a fun project anymore. It was it was started as like a fun project.
37:50 Speaker 1 And people started to pay attention to it, right? So you better do something. And I I I went to uh Michael Jordan, my colleague, very famous machine learning um researcher and faculty here. And I've been working with him actually when I um built this kind of labs at Berkeley cross-disciplinary labs. Um we're working with him in like 2005, 2006. He was joining the system people, database people to all together. Right. And you know, exciting projects. And he told me, oh, I know exactly how the guy for you is Anastasio.
38:28 Speaker 0 I saw what was being built at the time, Arena was not so close to what it is today. It was, I think. Not that much usage. I saw it and I thought, wow, what a great opportunity to do some interesting statistical modeling and theory. Like being able to uh understand how do we optimally sample models, how do we perform this estimation? Okay, let's move from Elo to Bradley Terry, because we're actually performing uh an estimate here instead of just like, you know, in the ELO score moves over time. It doesn't converge, but rally targeting models converge. And how do we then construct confidence intervals properly for this estimate and so on and so forth? All that stuff was super interesting to me. We were meeting like just next year. Two two doors down, and we wrote on the whiteboard like five or six different, you know, topics.
39:17 Speaker 0 Yeah,
39:17 Speaker 3 I
39:17 Speaker 0 mean, we started working on them and the rest is history, right?
39:20 Speaker 3 In many ways, I feel like this is a like the the birth of arena couldn't have happened anywhere else, other than an interdisciplinary lab at a univers research fundamental research university like Berkeley. Is that true or do you think that's the same thing?
39:33 Speaker 0 Well, it certainly would have been worse if it came out of somewhere else. And the reason is because the fact that we come from Berkeley and from a university really speaks to our scientific approach and neutrality. I think if it came from an industrial lab, people would always have questions about, oh, well, these people are they also training a model and what's their incentive and so on and so forth. But the reality is we were just students, we were doing this in order to evaluate models, and they came from a scientific perspective. That's it. And I think that's something that people can see when they look at us and builds a lot of trust in our business.
40:06 Speaker 1 The other angle here is that in a lab like this one, uh, what you get, you can get maybe in the in industry, you can also get maybe interdisciplinary teams, but they are going to be large teams, right? Because you go, okay, you are going to do the team which is doing AI, the team is being systems, work together. So these are already large teams. But here what you get, you get a small teams, a few people, which everyone comes from or can come from a different areas.
40:36 Speaker 2 Right.
40:36 Speaker 1 Right. We have people who are kind of systems. Early on, we have to build the systems to serve this open source models, right? We have to serve this Vikunia, like Wilean mentioned, right? Then you have to have people who are pretty good actually when we use this shared GPT data. We are doing quite a bit of data pre-processing in order to, you know, to pick some data curation and so forth, right? Then when Anastasio joined, we have now machine learning experts, right? But the team still is like four or five people. Right. And as you know, early on, small teams move very fast. So I think that's kind of the difference.
41:26 Speaker 1 So I think you may get in industry, you may get in this interdisciplinary teams, but they're going to be large.
41:35 Speaker 3 If you could just teleport back in time to that moment in early 2023, correct me if I'm wrong, but if I had to kind of summarize the research environment in the Bay Area at the time, most people basically were extolling the death of AI in academia, right? The idea that, oh, you can't really do any serious research. You can contribute to the frontier of computer science or in AI from a research institution was quite common, actually, if you remember that.
42:01 Speaker 1 And it's nothing more satisfying than proving these people wrong.
42:04 Speaker 3 Right,
42:05 Speaker 2 right. So let's
42:08 Speaker 3 do you think people got wrong.
42:10 Speaker 1 I I think that if I if I may take a step back here, because you know, I'm old enough, so I've seen a few of those. I remember when I was a student, I was doing system work and uh networking.
42:23 Speaker 0 Right.
42:23 Speaker 1 Uh that was the internet days. So and I was going to these conferences and there are panels uh the operating systems that research in the operating system is that dead. That was the topic of the panel. And the reason for that was at that time, it was at the end of the previous century. It was Microsoft was dominating, and then Apple. And then, of course, there were some free BSD and so forth. But then, you know, although it didn't come from Academia, it was Linux, right? Actually, it did Linux was preceded by Minix, which which came from Academia from the Netherlands. So that's one. Then in 2004, like when we come here and we started this lab, there was a question also that was uh this kind of distributed system, because a lot of researchers on distributed system.
43:15 Speaker 1 What can academia do? Right, right? Because it was Google was doing all this search of Q systems, map reduce Google file systems, all of that happening at Google, right? Right? Best people going at Google and so forth. And then we uh we we done here, then it comes Spark, right? Right, which also come from Academia, right? And I think that's when this started, it was actually the question people are surprised about Vikunia, just a bunch of students and actually their own initiatives, right? I knew almost after the fact that this happened. Pick this kind of data set from the internet, which is high quality, right? And use it. And people are so surprised about the quality, right?
44:00 Speaker 1 That was that was we were people asking, is this real? Right, right. So I want an evaluation. I'm not kind of you just show me some some stuff anecdotal, right? And oh, okay, it's looked good, but how about, you know, is that
44:15 Speaker 2 some people didn't believe it. Yeah. And then say, this is
44:18 Speaker 3 uh GPT fault rapper. Yeah. I remember later. I remember I remember we were at New York's uh later that year, right? Um and I remember uh Waylin was sitting at uh a table next next to me, and uh really well-known famous researcher who's still at OpenAI, um, you know, asked me, oh, is that uh the team that worked on that Vikuna uh bought? And I said, yes. And he said, um, oh yeah, I've been wanting to have a conversation with them because we think they're violating our terms of service.
44:52 Speaker 0 Yeah. Because
44:54 Speaker 3 they're just re-reselling our GPT for I don't know if he came and confronted you, but that was very much the default assumption people had, right? It was a disbelief.
45:02 Speaker 1 So so because for again, for for for for a while, um, it was the best uh open source models. I don't know, three months, four months or whatever. Right. So that's uh that's why the evaluation also important back then, right? Because it's kind of this disbelief. So we try to support with some evaluation, which was uh seems more objective. Right that indeed it's it's it's a good model. Again, since then there are many other things we've done here and open source like inference, uh LLM inference, like VLLM and Lang. I but I do think that what happens and this kind of this sense in um in uh in industry. Um I was actually on a panel yesterday, was the same this kind of discussion.
45:52 Speaker 1 You know, okay, it's like, you know, what can maybe you know academia should do this thing, right? And the industry you do this thing, right? It's like let's um uh the industry just you know can do anything like pre-training and so forth. I think at the end of the day, um, it is about what resources you have and what problems you solve. And over and over again, I think through the example I gave. If academia has a resources, it's going to surprise you. Clearly, it's going to be at the very edge of uh innovation and uh um creativity and uh so that's always almost happened, right? Like in this case. In this case, of course, we we had we didn't need huge resources, right?
46:41 Speaker 1 It's just a group of smart, passionate students. So when uh Cherbot Arena started, it was a lot of excitement and so forth. But then it was this thing, okay. Uh, it was a feeling, at least for some people in the group that we are done here. We publish the paper and so forth, right? For a while, even if you look at the usage, it's kind of dropping a little bit, and it's almost almost that. Almost like. And then Weylin, at that time, his main trust of research was different. Some uh uh, you know, graph neural networks, distributed graph neural networks and things like that. And I remember at some point at one of our one-on-one meetings, uh, Wellin came to me and said, look, I I really like and I want to, instead of doing this kind of work, I really am passionate about Shellbot Arena.
47:38 Speaker 1 I want really to do it, right? And to focus on it. And then when kind of then it started, and Weylin is like one man backend, right? And from end study
47:50 Speaker 2 and then marketing.
47:52 Speaker 1 And and marketing and marketing. And he started to, you know, to add more models to the leaderboard, uh, market it and so forth. And very soon after that, Anastasio came, and then it was kind of magical, right? You have these people who are so passionate and they are working so well together, they are so complementary in skills and even personalities. Then it started to shut up, right? Right. And I'm I'm mentioning that because without that kind of inflection point, which came, you know, long after it started as a project, we wouldn't be here.
48:34 Speaker 3 I think there's uh a chart I saw recently that showed that that compared the number of models being released and tested on LM Arena per year over the last two years. And if you look at Q1 of 2023, it was I think two models.
48:57 Speaker 1 It's exactly
48:58 Speaker 3 and if you look at just this past quarter. Yeah, I think there were 68 models, something like that, right? In total, the that first year there were about 12 models or so. And and today it's over 280 or something on the platform. So at some point it took, it sounds like it took the two of you realizing that this deserved to be more than a one-off paper. When when was that?
49:21 Speaker 0 I still remember when we uh when we worked on the paper for chatbot arena. It was like, you know, a couple of weeks of really hard work, and we were like pushing all the way into the deadline. And afterwards, I turned to my girlfriend at the time. I was like, you know what? I think this is gonna be a pretty good paper.
49:40 Speaker 1 Yeah.
49:42 Speaker 0 And uh yeah, Willen and I were talking at the time, but uh I think we uh started very early thinking about this, you know, what this could become and trying to de-risk it in various ways and trying to uh build it and seeing, hey, is this growing? Can we keep building on it? Another
50:01 Speaker 2 feel that really like drive the girls is competition. The competition of AI just become much more intense. In uh early 20, 24, when Cloud 3 came up.
50:17 Speaker 1 So let me tell you to answer your question, because I think that it's uh it's a very interesting uh and maybe I have uh more kind of unique view. As you know, I started other companies which are based on project coming from this lab, like um, you know, like data breaks with Spark and uh or any scale with Ray. And and there, you know, the motion was pretty clear. You have a successful uh uh open source project, right, which gets more and more popular. And then there are some companies which start to use a project and then they you get to the point that say, okay, if I'm going to bet on this project to be part of my infrastructure, what happens when the students who build it like Mateo and so forth graduate, right? Who is going to maintain it?
51:04 Speaker 1 Who is going to evolve it? So in that particular case, it's kind of natural. Okay, if really this gets uh very to get even more successful, you have to have a company backing it, right? Whether it's a new company or an existing company, and if there is no existing company, it's you you know, people who are on that project, if they want to push it farther, you know, almost like you can't you have to start a company to have enough resources to push it.
51:32 Speaker 3 Right.
51:32 Speaker 1 But this was different for the reasons uh Anastasio said, it's kind of your berkeley, it's kind of a trust to neutral. And Waylin, I remember mentioned to me like one year ago, he's like, uh, you know, like I think maybe you should do a company. And I told him, man, what you're talking about. This has to be remained neutral, maybe we do a kind of foundation and so forth. And this discussion actually went back and forth for a while. I was even frustrated. I'm telling this guy, you know, but I think it should happen. And, you know, it should be just kind of foundation and so forth. And she come it comes to me, it's like not hearing is like telling me the same thing, right?
52:19 Speaker 2 So we were trying to like convince you as it comes.
52:22 Speaker 1 Obviously, right? They're trying to convince me, right? And then one when on for me, it's like it it was um after, you know, we talk with some of this foundation and so forth. Uh but when this one when it was very clear uh for me is that when you started to get more and more demand and so forth, and there is no way you can you need so much funding to build such a platform, right? Because you need to serve the models and you need to build an entire backend, scalable backend and things like that to do it. And then you are you are UX, right? So when you look at that sheer amount of work, in order to push them to the next level, there is no way you can do it without having significant funding.
53:08 Speaker 1 So that's kind of for me was uh uh was a kind of inflection point. But these guys can say more because they are convinced about this long before I think
53:18 Speaker 2 yeah. Another thing I think we were discussing last year was like when we were trying to discuss whether this can be really like a business that solve more fundamental problems in the space. I think Anna Sasa at that time was like giving some perspective on ever more granular evaluation that we can provide with the data. Right. Right. So you want to say more about that? You know,
53:39 Speaker 0 chatbot arena, when you look at the leaderboard runs like a marginal regression, which means that the leaderboard sort of ranks models on average across all users and all the prompts that they ask. But you know, there's a vision where you take this to the logical extreme where, you know, there's the overall leaderboard, then you can categorize the leaderboard into different categories, coding, math, hard prompts, and so on. But you know, the real value is in well, what if I can tell you which model is best for you? What if I can tell you which models best for you and your question for your business? There's so many interesting methodological questions to ask there. And actually they require a lot of resources to answer. So, you know, one thing that we've been working on recently is called prompt to leaderboard.
54:26 Speaker 0 Prompt to leaderboard asks the following question. You give me your prompt. Can we tell you which models are best for that prompt specifically? Now, the problem is we've never seen that prompt before. We've only seen any prompt once or zero times. Right, because most people don't ask, you know, every question under the sun. Fundamentally, it's a hard question because the thing that you're trying to estimate is what if infinitely many people came to me and asked the same question and then voted. Right. That's the thought experiment you're trying to run your head, but you can't really answer that question by running a standard regression. So instead, what we came up with was a strategy for training language models that can output leaderboards.
55:09 Speaker 1 Right.
55:09 Speaker 0 And it's actually a deep question. Because what essentially you're doing is you're training LLMs to output these Bradley Terry regressions that we were talking about earlier. And how do you do that? Well, you have to make sure that as you train the model, the regression sort of naturally emerges from the data. And the only thing you're getting is binary preference. But nonetheless, it turns out that you can do it. This has so much utility and it requires so many resources in order to really scale up. It converts the problem of testing and evaluations, which is normally kind of like an unsexy problem. You think about it as like, okay, how am I going to evaluate ML? Well, I'll just like calculate the accuracy. Right? But the reality is that that really doesn't reflect the heterogeneity of the performance of the model for different settings and for different people.
56:03 Speaker 0 But instead, what prompt leaderboard teaches us is that you can convert the problem of evaluation into the problem of learning. What if I learned something that can tell me how my models are performing in all different parts of the space? It turns out that you can do that by training big language models. That because language models are sort of the intermediary that gets you to this evaluation, there's also a scaling law that comes along with it.
56:30 Speaker 3 Right.
56:31 Speaker 0 Which is to say that the more data you get, the bigger you build the platform, the better you can make your evaluations, the more granular you can make them, the more personalized you can make them. And that's a very, very powerful idea. And I think that's part of the reason why we were convinced, hey, this deserves to be a company of its own. So fundamental technical innovation that's going to change the way people approach the space.
56:50 Speaker 1 And let me try to follow that with a more uh a less accurate explanation, but I think it's uh drives home the point why the data is so important. So with the pronto leaderboard is basically when you give your prompt, and again, like Anastasio said, we may have never seen the prompt, more likely. However, what we have seen maybe a lot of other prompts which are similar is your prompts, right? So intuitively you can think that you can use the votes to this similar prompts as a proxy to compute how good are the models for your prompt. Now, from this kind of, you know, maybe not as accurate analogy or explanation, you can see that if the more data I have, the more prompts similar to your prompt I have.
57:44 Speaker 0 Right.
57:44 Speaker 1 So the more accurate I can be. There is another thing I want we didn't touch on, and what actually for me was I was so excited about the project early on. And if you think about outside Vikuna and our own story, how people and still evaluate these models, you have this kind of benchmarks.
58:04 Speaker 3 Right.
58:04 Speaker 1 MMLU, helm at that point, uh, sweepbench, all of this models, right? The problem with that is that they are static, right? So you can in some sense overfit them. And at that time, if you remember, there are already starting to be discussions. You know, I'm talking about one year and a half ago, 52 years ago, about contamination. Right. Right? Or some very high profile examples. And why? Because this lag language models are going to train as we know now, the data is it's it's it's a bottleneck. So they train on all the possible data they are going to get their hands on, right? In the internet, right? And many of these benchmarks are also out there. Right. So it's not intentionally probably many.
58:51 Speaker 1 Uh uh, but they are going to train on some of them on the very benchmarks they are going to be evaluated on. So this is kind of another fundamental problem. And I think that the unique thing about Sharbat Arena is kind of evolves over time. Right. You know, we are thinking that we are thinking that the way people typically evaluate this models is like giving a student over and over again the same exam. Right? Certainly we don't do that, or at least we try not to do that, right? As a I'm talking as a faculty now, right? For each for each class, you know, for each year, we need to give uh different exams, right? So that's kind of again with the humans the same thing, right?
59:37 Speaker 1 To evaluate humans, to evaluate, you know, which kind of learn over time, like these models, you need to come with some, you need to evolve the benchmarks, right? The exam, right? Right. So I think that's kind of unique part and unique value of uh of Charbot Arena. And probably these guys can say more about the kind of freshness and The evolution of the benchmark over time.
1:00:06 Speaker 3 What are the biggest differences between benchmarking and evaluation? So let me just zoom
1:00:11 Speaker 0 out for a second. Benchmarks, how are they collected? What happens is that you ask a question or you know, give an input, and then a human grades the output.
1:00:22 Speaker 3 And
1:00:22 Speaker 0 then what is the benchmark supposed to be? There's an answer key. The benchmark is like a test with an answer key. A human has to look at it and tell you what's right or wrong.
1:00:30 Speaker 2 Right.
1:00:32 Speaker 0 The fundamental insight of the arena is that by virtue of the fact that we built this platform, we can do something closer to reinforcement learning. Benchmarks are like supervised learning. Arenas like reinforcement learning. And supervised learning, you can only do as well as the best human that you have. Because what's happening is that you're learning from the teacher. In reinforcement learning, you're learning from the world. You're able to learn things better than the best human could ever teach you. Why? Because you're only getting these preferences. You're getting, was this good, was this bad? Nobody needs to tell you why. Nobody needs to tell you, hey, oh, you need to improve the facts, oh, your writing style needs to improve an XYZ way, and you should edit the set.
1:01:18 Speaker 0 Forget all of that. For the same reason why reinforcement learning has been so powerful in training language models, it is also powerful in evaluation. It can capture things that you and I, if we were looking, could never understand how to encode.
1:01:33 Speaker 3 Right.
1:01:34 Speaker 0 It is the open world nature allows you to go back and mind the data in order to extract insights that are much more profound than we could come up with ourselves. So this
1:01:46 Speaker 3 seems to be the fundamental tension, right? If you, let's say you are a leader in the AI industry, your product lab or your product company, and you say, we believe the most valuable thing for us to do is build useful AI products. We're not interested in benchmark hacking. We're interested in making truly useful products. If that is actually true, you should be strictly supportive of testing your systems more and more on arenas like Web Dev Arena. Let's say you want to build a useful web development AI experience. Then you should want to your teams to be testing more and more on this
1:02:23 Speaker 0 product, right? You want to do well on the distribution of natural use.
1:02:27 Speaker 3 Let's say then we expect anybody who's serious about building useful AI products to want to use testing environments like Web Dev Arena more. Why are people complaining that some labs are testing more than others? And that and why are they saying that's a bad thing? Aaron Powell
1:02:42 Speaker 0 So first of all, I think it's worth saying that we offer the same level of service to all labs. Okay. There's nobody that we treat preferentially or anything. It's a neutral platform. We want to help the ecosystem advance. But second of all, you know, addressing your question more directly, people do not yet fully understand the arena. I think people still think about the arena as a benchmark. People still think about it as something like, oh, people can overfit on this thing. But what hasn't sort of permeated, and it's because it's just such a new way to approach evaluations, is when you have fresh data, you can't overfit. It just means you're doing well. Period. There's no overfitting that can occur. What can happen is you can do well. And you can argue with me about whether doing well is a good thing.
1:03:28 Speaker 0 Okay, that's perfectly fine. That's not what that's not where people's heads are at. I think people's heads are still at, oh, you tested so much, and that must mean that you're because people are used to it. Because people, oh, it's stat 101, so on and so forth. I know statistics. If you're doing well on this distribution, that's a strictly good thing, all else equal. Right. And then people can choose. How much do I want to tune my model for chat? That's your choice. You can choose how much you care about this signal. And that's okay too. So I think it's a fundamental misunderstanding. But I think as we go, as we continue building this, as it grows, people will become more educated on this topic. And then I expect that the world will understand it.
1:04:10 Speaker 1 And again, just to make sure, because we had the discussion is um Anastasia early on. So overfitting refers to the same data.
1:04:21 Speaker 0 Right.
1:04:21 Speaker 1 Okay. Okay. But when you do like you do uh supervised learning or something like that, then what you have, you have uh data, train data, and then you have test data, which you don't show during the training. And you hope that it's going to do well on the train data. Right. So overfitting means it's doing well on the test data, but only on the train data.
1:04:46 Speaker 3 Right.
1:04:46 Speaker 1 If you think from that perspective, there is cannot be overfitting because we have continuously fresh data.
1:04:53 Speaker 3 Right.
1:04:54 Speaker 1 Right. The one thing can people say that it's a particular domain which is given by the set of users and so forth, and you are going to learn to do better with this domain in this domain, which is perfectly fine. Probably it's easy that you should care about that because the domain is a it's uh it's a group of people you care about, right? Right? But it's very different. Overfitting, it's very particular meaning. Right. And what people think about here, oh, I'm going to do when they use a term of forfitting, I am going to do well on, I'm going to learn how to do well on uh arena audience. Right? That's what they have in mind, right? But it's again, that's fundamentally different.
1:05:38 Speaker 3 Well, actually, so let's talk about a second for the arena audience, because you you you mentioned that's a critical part, right? As opposed to continuing to train your model to perform well on a static distribution. One of the things that shocked me when between the first time, you know, Wayland and I chatted the beginning of last year to the end of last year was that arena traffic had grown by 10x. The user base of the community had gone up by 10 times. Why is that? That feels like something people don't I it certainly wasn't visible to me. What's going on under the hood? Why are more and more people using arena? And in your mind, is that one of the reasons why people don't realize how hard it is to actually overfit? How why overfitting is almost not possible on the other thing?
1:06:27 Speaker 3 So media
1:06:27 Speaker 2 is of people's preference. And I think one of the reasons why people are kind of like surprised to see usage grow is because when they think about arena, they think about the leaderboard.
1:06:39 Speaker 3 Right.
1:06:39 Speaker 2 They think about the again, a benchmark. How would a median people use a benchmark? That's strange. But in reality, arena is basically real world testing. And not just real world testing, real testing the best AI from all the frontier labs. Right. Does the demand grow over time for people to test the best I use the best AI? Yes. Right. So that's a very foundation of arena, which is like this is like open space where everyone can come here to compare all the AIs for their own use cases for free. And this demand we've been seeing has been growing, and we believe it has a very strong potential to continue to grow. And in the same time, we collect all sorts of like comparison data that would be used for evaluation for all sorts of tasks.
1:07:30 Speaker 1 So one thing I want to point out, because we have been talking through this uh discussion a lot about votes, right? The vote is a fundamental construct which allows us to evaluate this model and so forth. It's the votes have to be so have high quality. If they don't have high quality, it's like you said, garbage and garbage act. And we do believe, and there are two syncs, at least two things who you believe that the votes on arena are high quality. One is that the people who ask questions, how do the people co-evaluate the answers? So presumably they are going to have the context for that question for that answers, as opposed to I have one question and two answers, and I'm asking someone random labeler to say which of this answer is better.
1:08:23 Speaker 1 This is now from information retrieval field for decades, and it's called gold standard when people evaluate the answer to their own questions. When an expert evaluates someone else's questions and the answer is called the silver, if I remember correctly. But the second thing, people who give votes who vote, in our case are intrinsically motivated. They're not asking them to vote. Right? They can choose not to vote. Only people who want to vote.
1:08:56 Speaker 3 Relative to companies that pay humans to pay you
1:09:01 Speaker 1 to vote, or you provide other kinds of incentives, like, oh, if if you vote more, we can give you more resources or something like that, right? Because you can imagine, you can easily imagine how they can you can get their incentive, the wrong incentives.
1:09:16 Speaker 0 Right.
1:09:16 Speaker 1 Which are not necessarily aligned. When I say wrong incentives, they are not necessarily aligned with increasing the voting quality.
1:09:24 Speaker 3 One of the things that that strikes me as I hear you guys talk about the design of the platform is that unlike these other paid services where you can just essentially hand out cash or incentives, when you have somebody intrinsically testing, that the usage of a the quality of a model, that starts to look more and more like software testing. You know, so 15 years ago when software systems were starting to be deployed to the internet, you know, they were bugs, they were insecure, they were unstable, they were unreliable. And so as an industry, the we we developed the idea of unit tests and and and CICD. And A V testing. And today, software systems go through a sort of fairly reliable set of checks before they get deployed to production. Am I wrong or should I think about that as a pretty good analogy that we should want, if we'd like the progress of AI, the arc of AI progress to head towards more and more reliability, then we actually want model developers and AI developers to be testing their systems more before we actually get they get released to the world.
1:10:25 Speaker 1 So I think that's kind of when you started, and this is another thing about exciting. It's about, we do believe, and you can see right now, one of the main challenges of adopting AI in a wide area of scenarios, it's actually reliability, especially if you look at enterprises, right? Is this answer correct or not? That's kind of fundamental, right? And that's like you said, it's very similar with software systems. And for software systems, we develop, like you said, this kind of long and sophisticated testing uh processes, right? CICD and so forth, right? So you should think about that. You need something like similar for these models.
1:11:13 Speaker 1 Right now are basically tell the truth, like almost static benchmarks, right? This what we are doing, right? You know, you you start training your model, and when the loss rate plateaus, you start testing checkpoints, right? And you have, you know, uh 60, 70, 80 kind of benchmark, and you look at that in a spreadsheet, see which ShakePoint is doing better, whatever, then you can merge them. This is what happens, right? Today, right. But like we discussed, if you really are going to build your application for humans. Okay, you can still test on your static benchmarks. Nothing wrong with that.
1:11:58 Speaker 1 Very valuable. But you also want to test your models, right? Your checkpoints on Chabot Arena. For all the reasons we mentioned during this discussion. Yeah. So ideally you want Arena to be at the limit part of your CICD for training the models.
1:12:20 Speaker 3 We spend a bunch of time talking about how Arena was born and how the big idea, at least the theoretical idea is that for to unlock more reliability in AI. We need more testing of AI. So let's spend a little bit of time going deeper on the practical realities of making that possible. Um what are the hardest challenges when it comes to actually building the best testing platform to make AI more
1:12:52 Speaker 0 reliable? Arena is a very interesting platform. It's unique and it's kind of like N of one at the moment. And so there's a number of like technical challenges that are actually quite exciting. Um we're always looking to improve the platform, uh, both from the methodological side and from the infrastructure side. And you know, what makes it unique is that it's this combination of AI, machine learning, converting evaluations into learning algorithms, like reinforcement learning side of things, plus like pretty large-scale infrastructure. A lot of people don't know this, but childbot arena is used by like a million plus monthly users. We get like you know, tens of thousands of votes on a daily basis. We have like over like, you know, 150 million conversations that have been had on the platform.
1:13:38 Speaker 0 It's massive. And it's continue, it's like the leading platform for these kind of like subjective real-world evaluations continuing to grow. So the infrastructure side is actually quite challenging. And then the question is we have this like unprecedented data set. Um how can we use it and leverage it maximally in order to actually target what we want? Is which is like the most granular possible evaluations and measurements of model performance. Um why is that hard? Why is granularity hard? Well, granularity is challenging because fundamentally the questions you're asking when you talk about granularity is how does it work for this specific individual, this specific prompt with a specific use case? That is a hard question to answer. Why? It's because you know, you on come to the platform.
1:14:24 Speaker 0 You know, you ask three questions and you vote on one of them. How am I supposed to tell which model's best for you? It's like a sparse problem, where what happens is that you know, there's a big matrix of users and queries, and the number of queries is infinite that the user could possibly ask. And the number of users is very large, and they've only asked three of them. How are you supposed to learn which models best for that specific user? Well, you have to do something creative. And the methodology for that, you know, the it relates to all these sort of like core topics that are very like deep in machine learning statistics, recommendation systems, so on and so forth. But they come into kind of a new light when you think about language. So, you know, one example of a problem that you know we're working on towards the future is personalization.
1:15:13 Speaker 0 How am I supposed to create a personalized leaderboard for you? Uh let's say I have your prompt history and a few votes. Well, you know, in order to run a regression that's just for you, I probably need hundreds of votes. It's just going to be two high variants unless I have that much data. But I'm never going to collect that much data on a user. Or like only for the most power users, am I going to collect that much data at the moment? So we need a way that we can train models that look at your interaction history and then can compare you to other users and pool between users. So then you can create leaderboards for specific people, categories of people, so on and so forth. That is a challenging and interesting problem.
1:15:56 Speaker 3 And
1:15:56 Speaker 0 you need to do it using only this sort of limited information that we have, which is binary preference data. How do you do that? Well, it's a cool problem. It's a hard problem. And it's one that we like have taken steps towards solving. And it's not just personalization. What about if I want to value the data? What if I want to tell you which data points are high signal? Which users are high taste. What if I want to say Aange, he's fantastic at bioinformatics. But when you ask him about history, this guy doesn't know what he's talking about. Or what if you want to say, hey, this person right here, they're a local expert in this particular topic, and I really should upweight their opinions, let's say. Or this person's just voting noise. How do I, you know, take them out? We need to be able to do tasks like these. And they're they're fundamentally hard because of the structure of the data that we collect.
1:16:43 Speaker 0 Um, but they're also very exciting methodologically, and we we keep making progress on them, um, which is part of the reason why uh it fuels us. And it's all enabled by this massive infrastructure and platform. It needs to be done at scale and needs to be done very quickly. And you know, Weylin is kind of the expert on this, and he he should speak to more of that.
1:17:02 Speaker 2 Yeah. So before we go into infrastructure, I think one related note on all sorts of problems we are, you know, looking at like ML problem, which also related to recommendation systems in early days where people try to figure out the co-start problem. Right. You only have very few data points per user, but you are trying to do something, you know, personalized recommendations for them. Or not fix, yeah, for movies, what do people like? And as we, you know, lean to a like a more um personalized world where like company try to build AI products for consumer, everyone, uh, that leverage, you know, all these user histories, prompts that model has memories now.
1:17:51 Speaker 2 So there's quite a you know, new methodology needs to be developed. And in particular in this kind of like evaluation context.
1:18:04 Speaker 3 It seems like there's two or three emerging frontiers of AI progress, right? Uh relative to two or three years ago where models were pretty simple. You know, the the vast majority of questions people had about the quality of performance of the models were mostly about in-context learning, right? I I give the model a couple of examples. How good is it at predicting the next uh token or word in that sequence? And it was a pretty simplistic measure. Um, fast forward two, three years now, models have gotten extraordinary. Models clearly look more and more like systems. You know, one of the systems um improvements that you've described is memory, right? So relative to five, six months ago when most AI assistants like ChatGPT didn't have memory, but now do, people are starting to notice a discernible verticalization of the model and the systems layer, right?
1:19:00 Speaker 3 So uh famously OpenAI ha has spent a ton of time post-training their latest model 4.1 or 4.5 or whatever it was on the with the assumption built in that the model has access to the user's memory and context, right? When you have how how do you solve the problem of evaluating a model that where where the lines are blurring between model system application, this is turning into a full stack sort of product experience relative to a model that let's say doesn't have all of those, right? Because these these now these relative to two, three years ago, the the side-by-side base test was naively looked easier to do because it looked like Coke versus Pepsi or whatever, right? Now it looks like you know, uh a dessert versus an entree versus whatever.
1:19:49 Speaker 3 I'm doing a terrible job of the analogies, but you get what I'm saying, right? ChatGPT today, for example, has memory. Claude doesn't. Right. These are two consumer apps that look very similar on the surface, but under the hood, fairly different. The implementations are diverging. And yet on Arena, they they're evaluated side by side. Right. So how does that, what does that future look like? How do you, how do you guys disentangle the fact that the stack is becoming more and more verticalized and integrated across model system interface application? Um, but arena today is largely side-by-side evaluation of models that people are used to seeing, thinking of as basically symmetric systems.
1:20:35 Speaker 2 Yeah, I think it's a combination of again, evaluation would ever become more challenging and more specific to your applications. Just like, you know, all software systems needs its own CI CD Pipeline. That's very, very different, you know, from each other. I think that's the same thing would happen to all the AI products as well. So our belief is like in order to, you know, collect data or evaluation that really like means something that matters to us, to the app user or to user, we have to be a real world environment for everyone to test uh to use it, you know, give us real feedback. Um that's also why we are like and it is a combination of challenge of ML product design and engineering infrastructure, because ultimately we are going to serve.
1:21:28 Speaker 2 We are already serving millions of user who are going to serve tens of millions of user, right? How can we design a product that you know people really love to use? And then in the same time, that's the most organic feedback that we could collect for default kind of using, including memory. So why, you know, why if we have memory in an arena? Um that kind of like, you know, applications, testing like really like the long context uh capability of the model to reason about the past and then to have the potentially the rack system to retrieve relevant you know, informations from the user's path history in order to create a more personalized content for users or more personalized uh you know leaderboards for users that's help them to choose what's the best AI for their use cases.
1:22:25 Speaker 3 When Chat GPT has memory built in, um, but Claude doesn't. How would that actually work in production when I show up to the site and I'm trying to evaluate these models side by side, both both are uh serving the model via an API. Does that mean on the arena side you have to recreate memory and then abstract that away as a shared service that all the models consume? How would that implementation work?
1:22:51 Speaker 2 Yeah. So I think increasingly we're going to go beyond just single model. Uh that the model has a ability to connect to different source of information, you would say, like context.
1:23:07 Speaker 0 Search arena is one example of this. So
1:23:08 Speaker 2 search arena is like we launched a couple uh months ago. Um is basically arena to specific to evaluate models that has internet access of web data assets, right? And in that case, model is not just model itself. It has to be, you know, come in combination of other components. Right. And the same thing happened to memory, right? You have another component which is retrieving relevant information from user history. And then this history is actually reacher, like not only just prompt that has all the you know, battle between uh different models, comparison data, and then use us express preference. So that kind of like, you know, more like, and then it could be also multimodal, right?
1:23:54 Speaker 2 Can be like image, uh, can be like video, all the or PDF, right? People upload long document, that kind of stuff. So all these kind of like different context, different modality of data, how can we leverage them in order to create more personal experience and then and then evaluate them? That would be like very interesting challenge.
1:24:17 Speaker 0 Yeah, I would say there's basically two ways that we're moving forward. The first is the platform is going to continue to evolve for sure. We're gonna keep keep creating new arenas. We're going to keep improving the arena to integrate things like an artifacts component and things like memory and so on and so forth. And the second is integrations. You know, at the end of the day, if someone wants to evaluate their app, like we should be able to provide them a toolkit that integrates with our services to do that.
1:24:43 Speaker 3 Uh so if I let's say I'm building a code editor. Yeah. Um but I'd I'd like to understand which one of the 17 models out there are best for my users. Exactly. What does that look like? I I use an arena SDK. Exactly. Got it. And that's
1:24:58 Speaker 0 exactly right.
1:24:59 Speaker 3 And so how do what would that look like? Um my users would you know generate a bunch of interactions that then the arena SDK is serving on the arena site to run side by side, or is that eval actually happening in my app?
1:25:14 Speaker 0 I think it can happen in context. So what you can do is you can have some kind of a gateway that allows people to access all sorts of different models, even maybe the ones that they didn't, you know, handpick themselves, but the cutting edge ones that maybe they don't even have access to them. Maybe they're even pre-release, right? Um and then what we can do is on our side, use all the experience that we've built on sampling data tools, training models, this huge data set that we've collected that has all these multi-provider comparisons to do things like choose what the best model is for your users, understand how all the different models perform, all the cost benefit trade-offs, the periodo curve of cost versus performance of different models, all that stuff is stuff that we can instrument and we can do it using in-context feedback. So let's say somebody says, hey, let's hook into a thumbs up, thumbs down button, pass that back to the arena SDK.
1:25:59 Speaker 0 Well, we can look at that. Um and using that information, we can produce leaderboards for that organization. We're the experts in doing that, right? We've been doing this for years. Uh, you know, things like prompt to leaderboard and various technologies,
1:26:11 Speaker 2 you know,
1:26:12 Speaker 0 D3. Yeah. So we're we're building a project now that we call data-driven debugging D3. It's, you know, it's a little farther out. It's a little come in a couple months. Um, but the fundamental premise of that is that pairwise comparison feedback is not the only kind of feedback that we can use to construct leaderboards. We can construct leaderboards with any form of feedback. Uh, and because of that, we can hook in not just to you know, pop up pair-wise preference
1:26:38 Speaker 1 comparisons for whatever company, which is of course something that we can do.
1:26:41 Speaker 0 Um, but instead, what if I want to, you know, rank code models in part on how many times the code is copied or accept, accepted. Yeah, code changes accepted. You know, how many like what's the edit distance between the code that the model produced and the code that the human ended up sort of you know shipping?
1:27:02 Speaker 3 So that's interesting. You're saying we're moving from a world where the primary signal that is used to figure out whether to improve an AI model is sort of very explicit thumbs up, thumbs down binary preference. Um you see a future where every interaction I have within a product, you know, engagement, retention, down to uh uh a GUI interaction can help tell the model what to improve.
1:27:31 Speaker 0 Absolutely. So that's exactly the kind of stuff that we can loop into our methodology that we've been developing and generate useful feedback for people to continue improving their models. If you want to create code that people are going to use, you know, make sure that people are using it and that the edit distance is low and that people accept your changes. Okay. If you want to build an agent like uh like a Devon that's gonna be your software engineer, how many of these PRs are getting merged? You know, this is the sort of stuff that we're building the technology that gives you very, very rich insights into. And I think, you know, by virtue of the fact that we're developing this new methodology, um we you know, I think we have an edge to be able to provide people that kind of service.
1:28:16 Speaker 3 You know, you talked earlier about prompt to leaderboard. One of the things that surprised me when I looked at the repo, you know, it's an open source repo is how well the model performed on arena. Can you actually just walk through what happened when you guys recap what you know what it is and then what happened when you actually deployed it on arena.
1:28:34 Speaker 0 So I'll I'll get a little bit into technical detail here because I think it's cool. So prompt to leaderboard, what does it do? Um, if you look at the chat autarena leaderboard, it's Bradley Terry coefficients, right? Prompt to leaderboard is a technology that we built that allows you to take a prompt and then produce Bradley Terry coefficients for every model that are specific to that prompt. Okay, the Bradley Terry coefficients are a leaderboard. Higher is better. It means you're more likely to win a battle. So what's the natural next step from producing a leaderboard? Well, let's make a router. Anj asks me a question. I'm going to produce a leaderboard just for that question. And then how about I route his question to the model that's on the top of the leaderboard? It turns out that when you do this, when you train a prompt to leaderboard model, which is like, let's say a 7 billion parameter model.
1:29:21 Speaker 0 And then you use it to route Arnj's questions on the arena and everybody's questions. That model does better than any of the constituent models that were used in the router by pretty substantial margin. Now, here's another thing that's yet more interesting. Because the Bradley Terror, the Bradley Terry coefficients have a particular parametric form and the statistical meaning, you can use them in downstream optimization problems. So one example of an optimization problem is a router. Maximized performance, subject to cost constraint. So the router can be, for example, a randomized router that chooses between different models. It has like a random policy that chooses, hey, Anj asked me a question. I'm with 50% probability I'm going to route here with 50% probability. I'm going to route here. And I'm going to do so in such a way that my average cost is, you know, one cent.
1:30:07 Speaker 0 And I'm going to maximize my performance subject to
1:30:10 Speaker 1 that.
1:30:11 Speaker 0 Now, if you trace the performance, the best performance that you know any individual model can give you as part of the router as a function of cost. That's like 2x worse than the router. In other words, the router is giving you double the bang for your buck in terms of performance per cost. If you want to achieve an arena score of 1280 using the router, it'll cost you half as much as it costs you to use any individual model. That's amazing. What it means is that you're taking advantage of the heterogeneity and performance of these models across different parts of prompt space in order to properly route them. And by virtue of the fact that it has the statistical interpretation, you can cost constraints too. And that's why prompt to leaderboard is, you know, interesting is that because we believe it's like a fundamental first step towards addressing this routing problem in a principled way.
1:30:59 Speaker 0 And you know, from our perspective, it's like the right way to do routing. If you want to do routing to maximize preference, like, you know, even internally at OpenAI, they're doing these A B tests, right? If you want to maximize the sort of feedback that you get there and the engagement, then you should be using a strategy like prompt to leaderboard. And so, you know, our hope is that this sort of thing would make it easier for them to avoid the drop down and that they can actually implement it in their own product. I'm sure that they have strategies of their own, but maybe this can be helpful to them too.
1:31:36 Speaker 3 Let's talk a little bit about, you know, you said that experience will look different over time. The arena experience will look different than ChatGPT. Let's talk a little bit about the roadmap. You know, what what what are the biggest things that you guys are working on over the next few months? And then let's go longer longer term.
1:31:57 Speaker 0 Yeah, well, two that we've already mentioned are personalization and uh a leaderboard of users. Right? Can we can we get people, first of all, to like figure out which models they like best and sort of lean into that experience, incentivize them to give us better votes, you know, come here for their personal leaderboards and their personal metrics, and then give them a lot of them to drill down really deep in that.
1:32:22 Speaker 2 And in that case, we align the interests of individuals and the platform as a whole because you don't want to mess up your personal leaderboard. Just like how people these days, when they use social media, they don't like a random post. Because if they do that, then their feed would be messed up. Right. So it's like, oh, I will be more careful voting. I would be more careful looking at all these different models as they're in so long, which we leave, we believe collectively we will create a better, even more higher quality arena. Absolutely.
1:32:51 Speaker 0 Yeah. And then on the note of a user leaderboard, can we um value the data in such a way that allows people to know where they stand in terms of what kind of questions they're asking, how useful they are.
1:33:03 Speaker 3 Right.
1:33:04 Speaker 0 We think people are gonna love that. It's such a fun thing to be able to see that in terms of math, I'm asking the best questions. I would love that. I would love if I was asking the best statistics questions in the world. You know, and I think people will use that and think, hey, I want to be on the top. And so can we continue to align the incentives? And by the way, once we do that, it'll make it much more you know valuable, the leader makes leaderboard much more valuable because it'll mean that we sort of start removing the noise from people that might be sort of, oh, I don't know what these buttons are, click. And instead, people are getting really intentional, really high taste votes, identifying who those people are and maybe even being able to personalize so carefully that we can produce leaderboards for different like you know, types of people.
1:33:49 Speaker 0 Right. That would be incredible.
1:33:51 Speaker 2 And then on the flip side of it is like we as a platform has a bit more visibility into who are those user and how do we even customize the distribution that you know on the flip side flip side is model developer care or developers or developer at large cares, right? I wanted to say, oh, I want to test my AI or my you know system in uh, you know, developers in uh in Japan, let's say, right? And then can we have the ability to kind of like customize that kind of distribution to target what are the most you know meaningful distributions that reflect your use cases?
1:34:38 Speaker 3 One of the things that you guys have been pretty vocal about is open source. I think from day one, LM Arena is open sourced prompts, votes, you know, a chunk of the data that's that's being generated on the platform. I
1:34:55 Speaker 2 think we do every week probably update on the leaderboard. Right. And then, you know, all the code infrastructures that we process the data is uh published as open source and also research blog paper, and then including prompt leaderboard, we publish the paper, open source the models, the code and everything. Because we believe that, you know, this is critical in terms of building trust uh with the community and also be really build the foundation of this that we can like you know enable more and more you know value on top of it.
1:35:38 Speaker 3 Right.
1:35:38 Speaker 2 So for adoption reason for trust and then for collaborations. So
1:35:43 Speaker 3 as you guys have made the transition from being a research project to not being a company. What are the most important values to you guys to create and hold at the company as you guys grow out the team as the project grows? Absolutely.
1:35:58 Speaker 0 Well, we are very, very focused on neutrality, innovation, trust. We come from an academic background, and yeah, we want to maintain the culture of this is a project, it's a community-focused project that's gonna continue to grow. Yes, it's gonna be a company, the company is gonna support the project that we've already built and allow it to grow. Yes, it's gonna continue to change, it's gonna change for the better. Right? We're gonna keep improving it. We're gonna keep publishing papers. We're gonna keep releasing open source. We're gonna keep releasing open data. Right? That's all gonna be part of our culture. And part of, you know, it goes both ways because that's the way that you recruit the best. The best people don't want to hold up at a company and develop a bunch of proprietary technology that, you know, is never going to be released. And, you know, it's just gonna sort of stay in the annals of their nearest neighbors within the company, then they're the only ones that are gonna know.
1:36:48 Speaker 0 We want the world to know what the best ways are of evaluating these models and accelerating the ecosystem. Um, and you know, releasing this data is also a big big part of our trust. If people want to ask the question, hey, how are models performing on their, you know, why are they performing well? Go look at the data. That's what we did with Llama. Right? When people have questions about Llama, we just release the data. Right. Easy, right? Just go look. And we plan on doing things like this for the lifetime of our company, right? That's how we're gonna recruit the best researchers that are gonna help us develop the methodology. That's how we're gonna develop the best engineers who care about the whole ecosystem, not just one company. And ultimately, that's how we're gonna develop the best products. That's how we're going to become central to the space. We already are, but we're gonna cement it, is by remaining open and neutral.
1:37:37 Speaker 3 And how would you resolve the tension that often exists when you there are there are people who are concerned that as AI gets more and more prevalent, right? As AI systems start being deployed in pretty mission critical industries, like defense we talked about, you know, healthcare and so on. That in fact, uh there's an argument to be made that these systems should be close source and evaluated in a fairly locked down environment as opposed to being openly tested in this manner and this is actually irresponsible. How do you think about that cultural tension?
1:38:22 Speaker 0 Listen, I'm not an expert in national security. But I think an evaluation platform like ours has many different ways of being used. If they want to evaluate it publicly, they can. If they need a private deployment, we can probably also do that for them. You know, it just depends on the sort of level of national security risk, which is way above my pay grade. But um, for any of these things, you're going to need sort of these subjective community treatment evaluations. That's for sure. If things are going to be deployed in the real world, you're going to need real people testing them.
1:38:48 Speaker 2 Yeah. And also like there's a point when you develop the model, and this model is going to be used by broadly the public, there has to be like a face of testing it, right? And then we're trying to, well, we are building is like, you know, to breach this gap between the lab building something that's like, you know, the latest, the frontier research, and the world would use it as a large. Right. You need an environment for you to test in the sense that it's a more controlled environment with uh the people that you know the distribution you want to cut them, customize, you want on a standard preference.
1:39:34 Speaker 3 There, there,
1:39:34 Speaker 2 there's a need uh for a platform like this to exist and want to serve it.
1:39:41 Speaker 3 Yeah, could you talk a little bit a little bit about Red Team Arena?
1:39:43 Speaker 2 Yeah. So for example, this real world testing idea of arena can be applied to many different applications like we discussed, right? From chatbot to web depth to different modality image, that kind of stuff, and as well as red teaming, because Redimin at its core is like you know, a bunch of people try to draw break the models to see if it's you know really faithfully following what the uh the the model has been like instructed to do or created to do, right? So these days, uh many frontier labs in publishing kind of like model spec, that kind of idea, like how model should behave uh in this way in that weight.
1:40:29 Speaker 3 Right.
1:40:29 Speaker 2 And then, but how do you make sure model follow that instructions? You need real world testing again. You need uh red teaming, you need a group of people knowledgeable in this space to uh to help, right? So again, this can be community driven too, because there's a group of you know vibrant community of job breakers that want to help. And then they wanted to also like they want the tested for fun as well. Like so in Red Team Arena, we have a leaderboard, not just for model, but for user for job breaker, who is the best job breakers that can like identify issues uh for all the different models. So that very particular the very you know idea of real world testing still apply here and then still can deliver value to the ecosystem that we believe.
1:41:20 Speaker 3 So is it fair to say that if I wanted to understand the security or the safety sort of uh risks in a model, I could go to Red Team Arena and look at the evals that the models are generating over there? How does Red Team Arena actually work in practice to improve the security and reliability of these models? Yeah.
1:41:44 Speaker 2 Well, sure. So I think same as how we understand Chebawet that that kind of thing. There will be many, many different applications people are trying to build on top of the models. That's a custom service, custom as customer services, or like you know retriever systems, that kind of stuff, right? You want model to behave in certain way and you want control.
1:42:09 Speaker 3 Right.
1:42:10 Speaker 2 Uh and then in red team arena, the idea will be like why don't we build an environment to simulate data pacations? So for example, can we build an environment to simulate customer services where you know the model is instructed to not give certain, not take certain actions. And then you are as a job breaker trying to break the model. Right. So that kind of like signals that we will be getting in terms of like real-world testing jaw breaking will be reflective to the particular use cases that people can.
1:42:45 Speaker 0 By the way, Red Team Arena right now is still a little bit of a prototype. We're continuing to work on it. But it's interesting to see, you know, people can it's not necessarily the model that's like most refuses the most to answer these like queries that people ask necessarily better. Some people want to model this more controllable. Yeah. Some people want a model that's gonna say whatever they want. Some people want a model that's gonna be completely safe and you can use it, you know, PG13 or you know, created G. You know, that's that's okay. As long as people have the choice.
1:43:20 Speaker 3 So as we start to wrap up here, you know, one question that a lot of people ask is what what does the world look like, especially the world of evaluation and testing as we go from a pre-training world to a post-training world in a world of models to agents? Right. In some sense, it seems like you guys were actually a little bit ahead of the curve where Arena has always been an environment for agents more than you know, a set of static evals. So as people start as as as agents get better at long horizon tasks and tool calling and so on, this future where a ton of work in the economy is done largely by you know fully end-to-end automated systems. Does it does Arena have to change in any fundamental way for that future?
1:44:06 Speaker 3 Or does it largely look the same? Yeah, I think if we've
1:44:10 Speaker 0 been talking about what's the fundamental, the fundamental is organic real world testing with the feedback. That's not gonna change. I can tell you that that is not gonna change. Will we have to adapt the UI? Yes. Will we have to improve the product? Yes. Will we have to launch new products for evaluation? Yes. Will we have to develop new methodology? Yes. Does the fundamentals change? I think no. I think the reality is if you want to test your model for real world use, you have to subject it to real world use. To collect feedback from real world use, and that's it. So we're really excited about what the future has to hold there. You know, we don't actually even know ourselves where the product is going to evolve over the next five to ten years, right? The ecosystem's moving so quickly. But wherever it goes, you know, we're excited to follow.
1:44:51 Speaker 2 Yep.
1:44:52 Speaker 3 Awesome.
1:44:53 Speaker 2 Thanks, guys. Thank you. Thank you.
Previous Rick Rubin: Vibe Coding is the Punk Rock of Software Next Hey Tech, Come to Healthcare Page updated August 25, 2026 · transcript and summary are machine-generated
