Building the trust layer between businesses and models

Arena (opens in new tab)’s known for its leaderboard, the public model rankin (opens in new tab)g that AI researchers keep open in the tab next to X. But the leaderboard is just the surface of a much larger mission to ensure people get the most value possible from AI, and that’s evolved as business needs and models have. Forget tokenmaxxing: businesses use Arena to understand what it’ll ultimately cost to complete a task, and which model is best equipped to do so.

The same infrastructure now measures whether AI is working in our interest at all. Arena’s new Alignment Index introduces a specific, measurable framework for assessing how well an agent performs in our interests, and early data shows those failures climbing sharply as conversations get longer. Felicis’s Peter Deng (opens in new tab) sat down with co-founder and CEO Anastasios Angelopoulos (opens in new tab) the day after Arena’s Series B to talk about the problem with benchmarks, the relationship between performance and alignment, and why every business will eventually be a model company.

Transcript

Anastasios Angelopoulos, CEO and Co-founder, Felicis

Peter Deng, General Partner, Felicis

Peter Deng: Anastasios, welcome back to Felicis.

Anastasios Angelopoulos: Thank you so much for having me, Peter.

Peter Deng: So good to have you here. For those who don't know, I'm Peter Deng. I'm a general partner here at Felicis. I'm joined by Anastasios, the co-founder and CEO of Arena. And it's the day after we announced your Series B fundraise. So congratulations.

Anastasios Angelopoulos: Thank you so much. It's only the beginning.

Peter Deng: So for those who only know Arena by its leaderboard, can you tell us a little bit more about what Arena does as a whole?

Anastasios Angelopoulos: So Arena's an evaluation company. We're in charge of providing the neutral evaluation layer that ensures that AI benefits everybody. The way that our platform works is that we have this website, arena.ai, that everybody can go to. It's a global website where you can access it anywhere in the world and get free access to all the best AIs across all the different providers. And we have tens of millions of users on this site doing their deepest workflows, meaning long, multi-turn, knowledge work, coding sessions, and so on. And from that usage, we're able to tell whether or not people are getting value out of AI in reality or not. And we report that in our leaderboards.

Peter Deng: Arena is known for its leaderboards, but when we invested at the seed, we saw so much more. Tell us a little bit about like beyond leaderboards, like where is Arena going? And this is exciting to share with all our viewers.

Anastasios Angelopoulos: Well, leaderboards are great for marketing, but in reality, under the surface, there's a lot of deep technology that goes into understanding which AIs are best for different types of people, and different types of organizations, and all the trade-offs that come along with it. As an example, the cost of a model that you see on the rate card is not representative of the cost that you'll actually end up paying. Most businesses in the world, unfortunately, don't even know how much they're going to end up paying for AI because they don't understand the cost per task. They only understand the cost per token. On Arena, because we're seeing actual real sessions, we're able to measure not only whether or not people are getting value out of the AI, but the actual cost that they incur for every completed task that they get. And so we're able to take that methodology and put it inside every organization in the world. So the vision is that everybody should be able to have a robust evaluation infrastructure that tells them which AIs are best for them, how to improve their AI, and ultimately give them the best intelligence and the highest utility.

Peter Deng: So much coverage in the press is like, oh, this model can do that, this model can do that. I think one thing that everyone should realize is that we are going to forever be in a multi-model world, right? And so one of my favorite charts that you guys put out is the Pareto frontier. Can you just like explain what that is? What does that mean?

Anastasios Angelopoulos: Well, Pareto frontier is a concept in statistics that has to do with the optimal tradeoff between two different variables. So if you go to arena.ai slash leaderboard, what you'll see is if you scroll down a little bit, you'll see a Pareto frontier. And the Pareto frontier on the horizontal axis will tell you the cost per task that you're paying when you use a certain model. And then on the vertical axis, you'll see the performance. The way that you construct the Pareto frontier is you ask yourself the question, for each given level of cost, what's the best possible model that I could procure to do my task? And of course, the most expensive models tend to be also the best models. And so you can map out that curve of which models are best at every different level of cost. And you'll see that the Anthropic models are on there. Fable 5.1 is the best model for the majority of agentic work tasks, but it's also the most expensive. It's like 100x more expensive than some of the Chinese open source models that generally also land on the frontier.

Peter Deng: And are you able to tell which model is good for which type of task?

Anastasios Angelopoulos: Of course, yeah. We can stratify, we can categorize into different tasks. We can train models that help you understand, on a prompt specific basis, which models are best. It's like Arena is working to map out the jagged frontier of intelligence.

Peter Deng: So one of the things that feels like it was yesterday's news is this whole concept of token maxing. I remember back in the day when I was at OpenAI, companies would hire chief AI officers, which I didn't even know what their job would be. But I guess everyone was paranoid about not being advanced enough, right? And then all of a sudden, people were like, wait, this stuff costs money. And we have to run a real business with real margins. And what is the actual ROI? How might they use Arena to be able to make the best decision for what's actually required for their job?

Anastasios Angelopoulos: Well, thousands of companies are already looking at Arena, our leaderboards, every day. And we're constantly getting emails from people saying, hey, I run a business like this. I'm in a finance business or I'm running a hospital business. And which model should I be using for X? And so people are looking to us as a neutral source of information because they know that they can't trust any individual lab to tell them whether or not they're better or worse than any competitor. By the way, why can't they do this? It's because if you look at the way that labs report releases, and this is so funny, labs are always reporting these scorecards. And the scorecards contain a bunch of benchmarks, but the same benchmarks aren't used by different labs. Labs will only choose the benchmarks to put on their scorecard where they're doing well. It's like taking a class, and at the end of the year, you make your grade by looking at only the assignments where you did well.

Peter Deng: Right, right, right.

Anastasios Angelopoulos: And every lab is doing this, it's completely ridiculous.

Peter Deng: Yeah, yeah.

Anastasios Angelopoulos: And so it's actually very helpful to companies to have a trusted third party that has some consistency that will tell them, yes, it's true that Anthropic has gotten better in legal tasks. And we have that information because, you know, the single-digit percentage of our population of users is legal. And so we have, therefore, you know, hundreds of thousands of prompts in that area.

Peter Deng: Right, I love that. One thing that you and I always talk about, and I love it when you say this, is like you always tell us in the board, you're saying, we want to be Switzerland, right? And tell us more about that. That's how you're incentivized. That is exactly how the company's built. But tell us why that's so important.

Anastasios Angelopoulos: That's correct. The central value [unclear] Arena is trust. If we don't have trust, then the leaderboard doesn't have value for anybody. It doesn't have value for the community. It doesn't have value for the labs. It doesn't have value for policymakers. So the number one most important thing is to always make sure that our methodology is neutral and fair. How do we do that? Well, the way that we do it is, number one, we have an open source methodology for calculating our leaderboards that we've actually put on as a Python package. You can check it out on GitHub. It's called ArenaRank. And the numbers in the leaderboard just go through that open source package, and we basically do nothing, right? It's like, of course, we do some filtering for spam, but the pipeline's completely open source. So that's really what we do to keep the platform neutral and fair. And that position of being in the center of the whole AI race is actually quite valuable for a business. And the reason is because, you know, after the era of token maxing, there was the era of value maxing. And in order to understand value, you need to have an unbiased assessment of model performance and cost, which we can give companies and developers and anybody else who's interested.

Peter Deng: Have I ever told you the story of internally how we had so much conviction at seed?

Anastasios Angelopoulos: Oh, no, please. Tell me.

Peter Deng: I don't think I've actually told it to now. I remember I just became an investor here at Felicis, and then I was asked the question, Peter, this Arena thing, did you guys ever use it at OpenAI? And I think I kind of scoffed at that question, because when you're kind of immersed in a company and you see it, it's just second nature. I was like, how do you not know this? But I remember telling everyone, I was like, yeah, at OpenAI, there were two windows that the researchers had up. One was Twitter, and the other one was Arena, because that is exactly how they could measure objectively the progress against the peers across other companies. And as soon as we saw that, it was so obvious that if you believe that we live in a multimodal world, which we do and always will be, you're going to need somebody that's neutral to be able to sort through all the noise, which is exactly what you're doing. And that's going to be able to let you guys expand into more and more areas, right? And so I think yesterday, you just expanded to another one. I think there's the alignment index.

Anastasios Angelopoulos: That's right.

Peter Deng: I would love to talk about that because I think that is an important example of exactly what we underwrote way back in the seed is we knew that you guys could move into new areas.

Anastasios Angelopoulos: Yeah, and you guys were really believers since the beginning. You know, at the time that you guys invested in the seed, we were just a pairwise preference platform where, you know, you input one prompt, and then you see two responses. You vote for the better one.

Peter Deng: Right.

Anastasios Angelopoulos: Pairwise preference is great. It's still an important signal.

Peter Deng: Sure.

Anastasios Angelopoulos: Arena has gone way, way beyond that. Earlier this year, we released Agent Arena. And Agent Arena is our arena for measuring the capabilities of AI agents, how effective they are in getting work done— not in a pairwise comparison, but a single-threaded, deep, multi-stage work task. These agents can connect to your GitHub. They can open PRs. We measure the merge rate. They can connect to your Google Drive. They can make documents. We can measure the download rate of all these things. We can measure whether users are saying they're getting their jobs done. We can measure whether the models are steerable, whether they're hallucinating. And as of yesterday, we announced that we are also measuring their level of alignment with human values. Alignment is one of the most important signals of our day. With the whole Hugging Face incident and many other safety-related concerns brought up even by the government, we absolutely need a neutral third party that's going to tell us whether or not AI is not just benefiting people in terms of providing them utility, but doing so safely and effectively. We measure three signals in the alignment leaderboard. The first is unauthorized actions, whether or not the agent is taking actions that are against the intent of the user and breaking their permissions. So an example would be the user says, edit my PowerPoint. And then the agent goes off and deletes a bunch of spreadsheets or exits that folder and does something else. Really bad outcome. Imagine being in a business and it's like, you know, you're asking the agent to do one thing in one folder and then it goes and accesses a bunch of people's data and then deletes it. That's totally a possibility. The second signal that we measure is deceptive completions. Deceptive completions are the idea that the model is saying that it got something done, but it did not actually get something done. For example, yes, I did check every line in the spreadsheet to make sure all the equations are correct. Now, why is this an issue? It's because the model is lying to the user, not because it's incapable of doing the task, but rather because it thinks that the best way to please the user is to tell it that it did the job when it didn't actually do it. And so people might not even notice. They might say, okay, well, I'm going to trust the agent. But in fact, you know, then they sent that spreadsheet to the IRS and did their tax and blah, blah, blah. And they're going to suffer the consequences. The third signal is false attribution that the agent is ascribing intent to the user that was not actually there. So it's misunderstanding the user and the user says do A and the agent's doing B. An agent cannot be safe if it's taking unauthorized actions. An agent can't be safe if it's deceiving the user and telling it did things that it didn't. And so these are the three initial signals inspired in part by some of the signals that the labs have been using internally that we've chosen to form our alignment index. And we're going to be adding many more signals on safety in the coming weeks.

Peter Deng: The reason why I love this, just to insert my own views into this as well, is there's been a lot of talk in the press around safety and with models and how do we keep them safe. I think that missing in all those discussions is anything that's measurable and concrete. What you just described may not be the totality of signals. And you're not looking for perfection. You're looking for specificity and progression towards something that's actually measurable. And we all know the saying, you measure, you move it. As soon as you start measuring these things, it makes the whole debate around safety much more concrete.

Anastasios Angelopoulos: Totally agree.

Peter Deng: Here are three dimensions, right?

Anastasios Angelopoulos: We shouldn't be having these safety conversations that are about these random theoretical risks that people just think. Basically, they're pulling them out of their ass. It's like, what's your P-doom? It's a joke, man, because people are just making it up. We should be talking about whether models are providing real value to people and whether they're causing real harm to people. And because Arena has such a huge platform, we're able to measure that in the real world.

Peter Deng: Yeah, yeah, exactly. And it's all about instrumentation of the behaviors. And I remember being a grad student and just even looking at how do you describe behaviors in an objective, measurable way for psychology, for example. And it's the same exact thing, but you have to be very thoughtful about the measures, and I think you guys are being incredibly thoughtful.

Anastasios Angelopoulos: That's exactly right. In fact, we're taking a lot of inspiration from the measurement science literature. It's like a psychology literature. So we think a lot about what's the construct that we're measuring and whether there's ecological validity and whether or not there's construct validity, and all these concepts really, we bring them to bear in a big way on these problems. This is actually the reason why Arena is a better evaluation platform than benchmarks. I think the benchmarking industry in five years is gonna be dead.

Peter Deng: Yeah.

Anastasios Angelopoulos: It's the analogy of testing your product by using a focus group versus actually real customers.

Peter Deng: Yeah, yeah, I agree.

Anastasios Angelopoulos: And nobody uses focus groups.

Peter Deng: I agree.

Anastasios Angelopoulos: They're like a waste of money and time.

Peter Deng: I agree.

Anastasios Angelopoulos: A benchmark is just a set of questions that you paid people to come up with to get models to answer. First of all, the people that you can pay to generate questions for models is the same group as the people that you can pay to be in focus groups. Those are people that have nothing else better to do. So it's a crappy source of data that's not representative of reality. And second of all, these benchmarks, as soon as you see the questions, they're overfit.

Peter Deng: Yeah, exactly. Can you explain what overfit? For the non-researchers, what does overfit mean?

Anastasios Angelopoulos: Overfitting is the idea that once a model sees a set of questions, it can study for those questions and do well on those questions, but not actually be good at the domain from which those questions come. So, as an example, it would be like a student goes in to take a test, they see the questions, they do badly the first time, they go back, they find the answers to those questions, and they can go take that test and get 100%. And you ask the question, have they learned anything? And the answer is no. All they learned is the answers to those questions. But they didn't study to be good at the subject.

Peter Deng: Exactly.

Anastasios Angelopoulos: Right? By the way, we've also seen evidence that models can understand when they're being tested. And so especially for safety and alignment, if the model knows it's being tested for safety, it will pretend that it's safe. And then when you put it out in reality to real user queries, it will do whatever it wants. So it's being a bad student. Meanwhile, Arena works completely differently. We don't have a static set of questions. We actually just test the models by seeing whether or not real users are getting value. So the questions are always fresh. Users are always asking different things. And the only way a model can climb the leaderboard is if it's actually helping people.

Peter Deng: And let me just add to that, because I think that's the other thing that gave us so much conviction, is that we've seen in the past that human preferences evolve and shift. Our bar gets higher. And what I love about the Arena model, and what we all love about the Arena model here at Felicis, is that it's always going to be relevant because as long as humans are in the loop, or as long as humans are going to be a consumer of any part of the output of AI, whether it's answering a question or, to your point, agentic traces that go and span days, that is going to be an ever-evolving bar. And my simple analogy for this is if you look backwards, I remember when we first launched voice mode at ChatGPT, everyone was blown away. It's like, oh my goodness, this is so magical. It's so lifelike. And then within two to three months, people were like, this is so cringeworthy, right? Because it's like, all of a sudden, you've already adapted to that high watermark, and people just want more and more, right?

Anastasios Angelopoulos: Yeah, totally agree. I couldn't have said it better myself. I think there's evergreen value in measuring human utility.

Peter Deng: 100%. 100%. So going back to the alignment index, by the time this podcast comes out, we might already be out of date, just given how fast the models move. But who's currently on top? Give us a rundown of which models do you see are most well aligned.

Anastasios Angelopoulos: So the top two models come from OpenAI and Anthropic. So OpenAI has the most aligned models on our index. They're actually the best across all three signals. They have the fewest unauthorized actions, the fewest deceptive completions, and the fewest false attributions. Second place would be Anthropic. Anthropic is just a little bit worse than OpenAI on all three of these. And then third and fourth place, I believe today, it's Google and SpaceX AI. Now, underneath these, there is a sort of smattering of the Chinese open source models, which is very interesting. One of the trends that we noticed on the alignment leaderboard is that the more capable models are, the more aligned they actually tend to be. So we can't tell whether it is that those open source models are worse because those labs don't prioritize safety and alignment as much or because the models simply aren't as capable to sort of self-regulate. The other interesting trend that we're seeing on the leaderboard, if I had to show one plot, on the horizontal axis, you can say the number of turns in an agentic conversation. On the vertical axis, we show the rate of these different misalignment signals. And then we have three line plots showing, you know, false attribution, deceptive completion, and unauthorized actions. And as the conversation goes on, these signals are going like this. By the time you get about 20 turns into a conversation, a long-running conversation, the rate of one of these signals occurring is over 50%.

Peter Deng: Wow. And why do you think that is? Is that because things go outside the context window or there's just more shots on goal and there's more chance of one of these occurring?

Anastasios Angelopoulos: Well, the second factor is definitely part of it. But I think that the degree of increase in the rate can't be fully explained by the fact that there's multiple shots on goal. I think what's happening is that models are actually not properly trained to be aligned in long-running conversations. They're very good at being aligned in the first and second turn.

Peter Deng: So let's talk about that because I think that this is one of the things that we talked about at the board meeting that I thought was so valuable, right? The fact is you have day-long agent traces.

Anastasios Angelopoulos: Yeah.

Peter Deng: Which actually is not how these models are typically trained.

Anastasios Angelopoulos: Right.

Peter Deng: And, you know.

Anastasios Angelopoulos: We have, by the way, multi-day agent traces.

Peter Deng: Yeah, multi-day. It's a good indicator of the fact that how a model is trained is actually correlated with how it performs.

Anastasios Angelopoulos: For sure.

Peter Deng: So if, I mean, look, I think that to the extent that people care about safety and alignment, I think people need to really care about Arena. Because there needs to be a third party that comes in and is able to help hold up a mirror with incentives that are different than any of these labs. Your incentive is to be Switzerland. Your incentive is to be trusted over a long period of time. And that's going to be a key player in this whole evolution of how AI is going to go. And if I'm at one of these labs, I'm always going to be valuing the signal because it is the same reason why we don't want groupthink if we're building a product. You want to always seek out information from your customers. In the old days of building the products that I've built, you always want to get that data because you need to make sure you're attuned to the users. And no matter what, if you're a lab, if you look at only your internal data, you're just going to be biased because those are your users. And I think that people are aware of this. Like, for example, I use Claude and OpenAI. And I use them for completely different things for no good reason. All of my personal stuff is on ChatGPT and all my work stuff is on Claude. There's no reason whatsoever. It's just a muscle memory habit of the shortcuts on my keyboard. And as a result, like if OpenAI were to ever try to get good at the things I do for work, they couldn't use my data because it's not even there. Right. Do you have a sense of there being these skews of people using Claude for one type of task and something like OpenAI for another?

Anastasios Angelopoulos: For sure there's skews. There's also skews demographically. You know, I think Claude is more popular among people who are high-skilled workers, knowledge workers. People like that love to use Claude. The general population likes to use ChatGPT. I mean, that's the way that the cards have fallen.

Peter Deng: So give us a sneak peek of what's next. So we talked about evolving from a leaderboard to the alignment index. That's just the tip of the iceberg, right? So can you give us a sneak peek of what other types of things might we see from Arena coming forward?

Anastasios Angelopoulos: A few big things that are coming. Let me just talk about three. So the first is we plan on digging into agent mode so that all of our users on arena.ai can have an even better experience on the platform. New products, you know, more integrations, the ability to use multiple harnesses, all that stuff is coming on Arena. So Arena will be the best place for you to get real work done across many providers, many agents. Number two, we are releasing more and more safety signals to the leaderboard, okay? So that the leaderboard can expand its mandate beyond the alignment signals that we have now into other aspects of safety. And in fact, we're working with labs to define what those signals are so that we can understand their needs. And the third is that we are taking this technology and productionizing it so that it doesn't just work based on the data that we have from arena.ai, but it can be copy-pasted into every business on the planet. Every business is going to need their own Arena. And the reason is that every business is going to be a model lab at some point.

Peter Deng: Say more about that. Why do you believe that every business is going to be a model lab?

Anastasios Angelopoulos: Well, AI is such a transformative technology. It's essentially like a labor supplement, if not a labor replacement. Then every business on the planet has labor, and so what's going to happen is that every company, in order to be operationally efficient and to be maximally effective, is going to need to adopt [unclear] throughout their organizations. In 20 years, there's going to be no such thing as an AI company or a non-AI company, just the same way that today there's no such thing as an internet company versus a non-internet company. And so in that world, people are going to need to understand what's the performance of all the different models. They're going to have to have continual retraining pipelines. And Arena's in a pole position to be able to do that in a provider-neutral way.

Peter Deng: I love that. All right. So many exciting things to come. Thank you so much, Anastasios, for coming on and giving insight into Arena. This has been an amazing conversation. And I feel so grateful to be able to work with you on the board. It's awesome.

Anastasios Angelopoulos: Thank you. It's our honor. Thank you for believing in us from the very beginning.

Peter Deng: Thank you.