Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI
https://www.youtube.com/watch?v=mMNkdYnIVC4
AI Engineer · 16 min · 170 sentences · sentence-level transcript from YouTube's caption track; each line begins at the time shown

[00:12] All right.
[00:13] I think we'll go ahead and get started with the with the presentation.
[00:17] So my name is James So.
[00:19] I am uh going to explain some of the work we're doing with Together AI and it's also in collaboration with Stanford around designing and optimizing environments for AI agents to enable these agents to make
[00:32] new kinds of scientific discoveries.
[00:35] All right.
[00:39] So so that I guess the current paradigm of how people often are using or deploying AI agents is often involves designing workflows that sort of tells the agents you know what to do, right?
[00:50] Or how the agent should work.
[00:53] And it's typically done through a series of steps or prompts, tools, and instructions.
[00:58] In contrast, the way we imagine the environment is that the environment should really specify not how the agent should work, but really where the agent should work, right?
[01:08] And the environment then should provide a set of incentives and infrastructure for the agents and guardrails and resources so that agent can then flexibly work within that environment.
[01:20] Right.
[01:21] And our thesis here is that as agents become more and more powerful, right?
[01:25] If we try to design workflows that often can limit the capabilities and creativity of the agents.
[01:30] Whereas if we properly design the environment, this can enables a lot more creativity and capabilities and intelligence for the agents to naturally emerge.
[01:39] This why I think we're trying to shift away from designing workflows and harnesses towards designing environments.
[01:46] So what I want to do today is to give a few examples of the how we design environments for agents.
[01:52] And in particular also show how they're able to then with with the right environment able to actually solve some really interesting and innovative problems.
[02:02] So, the first example I want to share is the system that we environment that we created called the Einstein Arena.
[02:07] It's sort of like the one of the first environments that enables AI agents to be able to collaborate in the wild and to compete to really solve open-ended scientific problems.
[02:18] So, we designed this Einstein Arena to be really agent native.
[02:22] So, I So, that means that it's very easy for agents to just read the skills talk on our on our arena and be able to access the arena.
[02:31] And it's actually also designed so that it's intentionally very hard for humans to enter the arena, right?
[02:37] So, you actually have to solve a little puzzle to prove that you're an AI agent in order to participate in this arena.
[02:43] But, any agent in the world can openly and freely participate on the arena.
[02:47] And once the agent actually enters into the Einstein Arena, this is what they'll see, right?
[02:51] They'll see actually see a list of curated problems.
[02:54] Each of these problems is actually a problem that we curated, so it's a scientifically interesting problem.
[02:59] And we curated these problems so that first, there's actually an existing community of human researchers that are interested in these problems.
[03:06] So, these are important problems for human scientists.
[03:09] And second is that for each of these problems, we can actually create a well-defined and deterministic deterministic verifier to assess the quality of the solutions to each of these problems.
[03:20] And I'll give some examples in a couple of slides.
[03:25] So, So, the agents can actually decide which of these problems they're interested in once they log onto the arena, right?
[03:30] So, if they enter into a particular problem space, this is what they'll see, right?
[03:34] They'll see some description that precisely explains what is the problem.
[03:38] We have a discussion forum where the agents can communicate.
[03:42] It's almost like a social network where the agents can actually communicate and talk to each other and ask for help or give recommendations.
[03:49] Um and we also have a leaderboard.
[03:52] This is where the agent can actually see each other's solutions.
[03:55] Right?
[03:55] So in any in at any time they want, the agent can actually submit a solution to one of these problems.
[04:01] And because we have this verifier, we can actually then determine what is the quality of that solution and provide a score in real time.
[04:09] So this leaderboard is being constantly updated in real time.
[04:12] And the agents can also see how other agents are doing on this problem.
[04:16] And they can also see other agents' solutions and download those solutions.
[04:20] So there's both a collaboration dynamics and also a competition dynamics in this arena, right?
[04:25] They can collaborate and ask each other questions and help in the discussion forum.
[04:29] But agents are also competing with each other.
[04:31] And that's why I think this also sort of simulates how human researchers can compete and also collaborate to solve interesting problems.
[04:39] So we launched this AI instant arena environment earlier this year, I think in March.
[04:45] And within a few weeks, it's already actually we're very impressed and very surprised that the agents were actually able to already discover new solutions to 11 problems that are of the best
[04:58] solutions that have ever been found.
[04:59] Right?
[04:59] So that means that the solutions that they discovered by the agents on AI instant arena were better than any previous human solutions or any solutions that we acquired using
[05:08] more specialized AI tools.
[05:12] So I'll just give you example of one such solution or one such problem which is called the kissing number problem.
[05:19] So this is actually a very famous problem.
[05:21] It's been around for hundreds of years.
[05:22] So for example, Isaac Newton was already working on some version of this kissing number problem.
[05:27] And it's actually relatively easy to state.
[05:29] Right?
[05:29] So the kissing number problem basically asks that what is the maximum number of spheres that you can place around the central sphere so that these additional spheres do not overlap each other?
[05:40] So for example, in one dimensions, right?
[05:42] So around the central sphere I can place one sphere to the left and one sphere to the right without overlap.
[05:46] So the kissing number in one dimension is easy to compute.
[05:49] This is two.
[05:50] In two dimensions, it's also easy to show that you can at most place six spheres.
[05:54] So, that's the kissing number in two dimensions is six.
[05:58] But, it turns out that in higher dimensions, it actually becomes really hard to compute what's the maximum number of over non-overlapping spheres.
[06:05] And the kissing number problem in higher dimensions is actually open, right?
[06:08] It's not been It's not clear what is the optimal number.
[06:12] And so, scientists have been trying to work on this problem for the last several centuries.
[06:19] And in particular, right, so the kissing number problem in 11 dimensions has attracted a lot of interest for various reasons.
[06:26] So, this is actually sort of a progression of the solutions in 11 dimensions.
[06:31] So, in the 1980s, right, so it's best known that there you can place 440 spheres, right, in 11 dimensions without overlap.
[06:41] And in I think 19 uh So, yeah, so so in in 1980, there was a big advance that the first for the first time showed that you can actually just construct
[06:53] with 582 spheres in 11 dimensions without overlap.
[06:57] Uh and then that sort of stuck there for about 40 years, right, until 2022, where a mathematician is able to publish a new advance, right, a breakthrough that's able to improve that to 592
[07:10] spheres.
[07:11] And then there's another breakthrough from DeepMind the following year that advances that to 593 spheres.
[07:18] But, with Alpha Zero, we know by having these agents able to collaborate actively, right, in the wild, within a few days they were actually able to construct a new solution
[07:28] that shows that for the first time you can create 604 spheres in 11 dimensions that do not overlap.
[07:34] And this is not just a problem that's of mathematical interest, because it turns out that the more of these sort of spheres you can place in higher dimensions without overlap that actually creates
[07:43] the better coding systems including ways of like doing error correction codes for information transfer.
[07:49] Right, so this actually is by creating this better constructions that also leads to this better engineering algorithms.
[07:57] And in this case actually the collaborations among these agents is really critical for making these advances, right?
[08:02] So this is a problem where not a single agent is able to solve by itself, right?
[08:06] Not you know, GPT 5.5 or a cloud models that can't really solve the problem by itself.
[08:12] So the collaboration among multiple agents is really critical.
[08:15] And here we're actually able to show that there's like this sort of a lineage trace of how the agents are able to collaborate and then basically take each other's solutions and refine that
[08:24] and further optimize it to arrive at this breakthrough.
[08:28] And you can also see some of these interactions and discussions on Einstein Arena, right?
[08:32] Where here's an example where the one agent actually was asking other agents, "Have you tried you know, some of these approaches?"
[08:40] Um, with uh, these STP approaches and then the other agents showed that yes, we have tried these approaches and here are some of the things that we found.
[08:46] Right, so the information sharing on the forums on the arena is actually really important to help the agents to arrive at this solution together.
[08:57] So in addition to solving these interesting scientific problems, but we've also been using platforms like the right Einstein Arena uh, to help to improve uh, you know, machine learning and AI itself.
[09:09] Right, so here's one example where we actually use these agents to basically help us to create better kernels for and to speed up those kernels.
[09:17] Right, and here we use the same environment, right?
[09:20] Where the agents can compete and they also can collaborate and they see these leaderboards.
[09:25] And we basically change the back end instead of trying to verify the solutions to this mathematics problem, here we're basically trying to you know, we will compile and benchmark and test and verify the quality and the speed of the individual kernels,
[09:37] right?
[09:37] And then we'll provide a feedback to the agents in real time in the form of these leaderboards.
[09:43] In these kernel settings, we also found it to be quite useful to have different agents with different personas, right?
[09:49] And these different personas actually corresponds to a different uh roles and priors that agents can actually have.
[09:54] So, for example, we have one agent that looks at tends to look at more of the profiling, another agent that tends to look at more of the memory consumptions,
[10:02] a third agent that looks at, you know, the precisions, the tensor computations.
[10:06] And these agents can and then across different personas, they can able to collaborate and a compete on the arena to speed up the kernels.
[10:14] And in this case, right here, the agents were also able to collaborate and lead to really quite substantial speed ups, uh including sometimes over two two x two-fold speed ups in some of these production kernels.
[10:26] So, here I'm just showing you a few examples where for things like page attention, uh and these are sort of for specific shapes, but we also have generalized this to many different shapes
[10:35] and different uh hardware types, right?
[10:37] Where we're actually seeing that we're getting up to sometimes over two x speed up in these kernels, and they uh compared to the previous state-of-the-art kernels for these problems.
[10:47] And these improved kernels created designed by the agents are actually already used in in production at Together AI.
[10:57] So, in the last few minutes, I want to show like a second example of a kind of environment that we created as a way to uh train and to create better data scientist agents,
[11:07] right?
[11:07] So, we call this DS Gym, which stands for data science gym, which is sort of like a unified environment that we created for both for evaluating and for training
[11:15] data science agents to solve complex data science problems.
[11:21] So, here in this DS Gym environment, we also curated and created a unified list of different data sets and tasks, right?
[11:29] So, these data sets can combine uh spans across many different settings.
[11:33] And the agents are then able to interact with these different data sets that we have through a unified uh interface and through code execution.
[11:42] In the DSGM environment, we also provide a unified infrastructure for the agents.
[11:47] So, for example, the agents can actually spin up many different Docker containers to test their data science algorithms and actually run them in parallel.
[11:58] So, in the process of actually creating the data sets and tasks for the data DSGM environment, so we initially actually wanted to incorporate some of the existing data science benchmarks that have been used to evaluate agents.
[12:10] But we actually quickly realized that many of the existing widely-used benchmarks actually have many problems.
[12:16] And one big problem is that they're actually very vulnerable to shortcuts.
[12:19] By shortcut, I mean here is that uh down here what I'm showing are three different common popular data science benchmarks.
[12:26] Right?
[12:26] And the in green here we see shows like the performance of the agents on these benchmarks.
[12:32] Uh but the red bar also shows how well they're able to the what fraction of the benchmark the agents can actually solve without actually using the data sets themselves.
[12:39] Right?
[12:40] So, just by reasoning or by, you know, uh doing other shortcuts without actually actually do working with the underlying data sets.
[12:47] And across many of these different benchmarks, right, sometimes up to 20 to 50% of the tasks can be solved without actually looking at any of the underlying data.
[12:57] Which I think is uh really a significant problem with many of the existing benchmarks.
[13:03] So, to address that, we actually carefully curated at our own our own benchmarks, right, for both for scientific analysis and also for predictive modeling.
[13:12] So, for scientific analysis and discovery, the way we did this is that we actually went through recently published papers and then carefully curated data and then also tasks from those papers.
[13:21] And then we also had human scientists and experts to review each of those tasks.
[13:26] And for predictive modeling, the way we did this is go through all the different Kaggle competitions to look for some of the recent Kaggle competitions that are still open
[13:34] and and where also you have high quality data sets and also high quality evaluations.
[13:39] Then we curated those into the DS Gym as a kind of task for evaluating how well models agents can actually build predictive models.
[13:48] So all together in the DS Gym, we actually have created over a dozen different tasks.
[13:53] They span across dozens of different scientific domains ranging from biology to physics to economics.
[14:01] It also involves many different data types and data modalities.
[14:06] So this actually makes it very easy for us to evaluate different models, both open and closed source models.
[14:12] And one thing we found is that the existing models, even the frontier models, often are only still achieves like less than 50% accuracy performance on the DS Gym tasks.
[14:22] Right?
[14:22] So these are definitely not saturated benchmarks.
[14:26] We can also use a DS Gym as sort of like a training factory to improve these open source models.
[14:31] Right?
[14:31] So one thing we did here is actually generate in the DS Gym actually the gym itself will actually create all these execution verified trajectories, which means that these are trajectories
[14:40] generated by the agents that have been verified through the through through actually executing the code from the agents.
[14:47] Right?
[14:47] So by generating these execution verified trajectories, then we are able to like fine-tune sort of small open source models that actually now achieve sort of the they're sort of the best in class
[14:58] open source models in terms of solving these kind of data science tasks.
[15:01] Right?
[15:01] And these models are small enough that you can actually run them locally on your laptops and your computers.
[15:08] So just to summarize the this part was the data science gym.
[15:11] Right?
[15:11] So we with DS Gym, we created this unified execution layer so people can actually run and all these different tasks across dozens of different tasks across many different domains.
[15:21] We have carefully verified that there are no shortcuts in these tasks, which has been sort of a common challenge with existing data science benchmarks.
[15:29] And we also enable in the DSG and a way to generate synthetic data, so you can easily use that to improve and to train your own data science agents.
[15:39] So, just to summarize the presentation, um I think the main takeaway here is that I think we're in seeing this interesting progression as in terms of how we build different AI systems.
[15:50] Right.
[15:50] So, then in the past, people have been building these AI systems mostly by designing individual models or individual tools.
[15:56] And currently, there's a lot of focus on creating designing agents or harnesses and workflows around agents.
[16:02] But what our research shows is that I think we're already moving towards the next stage, where you're not then trying to design workflows or specific or specific agents, what we really want to do is to design environments,
[16:13] which is a set of infrastructure and incentives that in that motivates the agents that you solve more and more challenging problems.
[16:20] And with appropriate designs these environments can actually unlock much more creativity and collective intelligence from the agents that's that's limited by the existing workflows.
[16:31] And here are some of the references for the papers that we published that describes these in more detail.
[16:36] So, thank you very much.
[16:38] [applause]
[16:51] [music]