Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI

AI Engineer · 16 min · 170 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:12All right.
00:13I think we'll go ahead and get started with the with the presentation.
00:17So my name is James So.
00:19I 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
00:32new kinds of scientific discoveries.
00:35All right.
00:39So 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?
00:50Or how the agent should work.
00:53And it's typically done through a series of steps or prompts, tools, and instructions.
00:58In 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?
01:08And 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.
01:20Right.
01:21And our thesis here is that as agents become more and more powerful, right?
01:25If we try to design workflows that often can limit the capabilities and creativity of the agents.
01:30Whereas if we properly design the environment, this can enables a lot more creativity and capabilities and intelligence for the agents to naturally emerge.
01:39This why I think we're trying to shift away from designing workflows and harnesses towards designing environments.
01:46So what I want to do today is to give a few examples of the how we design environments for agents.
01:52And 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.
02:02So, the first example I want to share is the system that we environment that we created called the Einstein Arena.
02:07It'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.
02:18So, we designed this Einstein Arena to be really agent native.
02:22So, 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.
02:31And it's actually also designed so that it's intentionally very hard for humans to enter the arena, right?
02:37So, you actually have to solve a little puzzle to prove that you're an AI agent in order to participate in this arena.
02:43But, any agent in the world can openly and freely participate on the arena.
02:47And once the agent actually enters into the Einstein Arena, this is what they'll see, right?
02:51They'll see actually see a list of curated problems.
02:54Each of these problems is actually a problem that we curated, so it's a scientifically interesting problem.
02:59And we curated these problems so that first, there's actually an existing community of human researchers that are interested in these problems.
03:06So, these are important problems for human scientists.
03:09And 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.
03:20And I'll give some examples in a couple of slides.
03:25So, So, the agents can actually decide which of these problems they're interested in once they log onto the arena, right?
03:30So, if they enter into a particular problem space, this is what they'll see, right?
03:34They'll see some description that precisely explains what is the problem.
03:38We have a discussion forum where the agents can communicate.
03:42It's almost like a social network where the agents can actually communicate and talk to each other and ask for help or give recommendations.
03:49Um and we also have a leaderboard.
03:52This is where the agent can actually see each other's solutions.
03:55Right?
03:55So in any in at any time they want, the agent can actually submit a solution to one of these problems.
04:01And because we have this verifier, we can actually then determine what is the quality of that solution and provide a score in real time.
04:09So this leaderboard is being constantly updated in real time.
04:12And the agents can also see how other agents are doing on this problem.
04:16And they can also see other agents' solutions and download those solutions.
04:20So there's both a collaboration dynamics and also a competition dynamics in this arena, right?
04:25They can collaborate and ask each other questions and help in the discussion forum.
04:29But agents are also competing with each other.
04:31And that's why I think this also sort of simulates how human researchers can compete and also collaborate to solve interesting problems.
04:39So we launched this AI instant arena environment earlier this year, I think in March.
04:45And 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
04:58solutions that have ever been found.
04:59Right?
04:59So 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
05:08more specialized AI tools.
05:12So I'll just give you example of one such solution or one such problem which is called the kissing number problem.
05:19So this is actually a very famous problem.
05:21It's been around for hundreds of years.
05:22So for example, Isaac Newton was already working on some version of this kissing number problem.
05:27And it's actually relatively easy to state.
05:29Right?
05:29So 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?
05:40So for example, in one dimensions, right?
05:42So around the central sphere I can place one sphere to the left and one sphere to the right without overlap.
05:46So the kissing number in one dimension is easy to compute.
05:49This is two.
05:50In two dimensions, it's also easy to show that you can at most place six spheres.
05:54So, that's the kissing number in two dimensions is six.
05:58But, it turns out that in higher dimensions, it actually becomes really hard to compute what's the maximum number of over non-overlapping spheres.
06:05And the kissing number problem in higher dimensions is actually open, right?
06:08It's not been It's not clear what is the optimal number.
06:12And so, scientists have been trying to work on this problem for the last several centuries.
06:19And in particular, right, so the kissing number problem in 11 dimensions has attracted a lot of interest for various reasons.
06:26So, this is actually sort of a progression of the solutions in 11 dimensions.
06:31So, in the 1980s, right, so it's best known that there you can place 440 spheres, right, in 11 dimensions without overlap.
06:41And 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
06:53with 582 spheres in 11 dimensions without overlap.
06:57Uh 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
07:10spheres.
07:11And then there's another breakthrough from DeepMind the following year that advances that to 593 spheres.
07:18But, 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
07:28that shows that for the first time you can create 604 spheres in 11 dimensions that do not overlap.
07:34And 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
07:43the better coding systems including ways of like doing error correction codes for information transfer.
07:49Right, so this actually is by creating this better constructions that also leads to this better engineering algorithms.
07:57And in this case actually the collaborations among these agents is really critical for making these advances, right?
08:02So this is a problem where not a single agent is able to solve by itself, right?
08:06Not you know, GPT 5.5 or a cloud models that can't really solve the problem by itself.
08:12So the collaboration among multiple agents is really critical.
08:15And 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
08:24and further optimize it to arrive at this breakthrough.
08:28And you can also see some of these interactions and discussions on Einstein Arena, right?
08:32Where here's an example where the one agent actually was asking other agents, "Have you tried you know, some of these approaches?"
08:40Um, 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.
08:46Right, so the information sharing on the forums on the arena is actually really important to help the agents to arrive at this solution together.
08:57So 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.
09:09Right, 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.
09:17Right, and here we use the same environment, right?
09:20Where the agents can compete and they also can collaborate and they see these leaderboards.
09:25And 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,
09:37right?
09:37And then we'll provide a feedback to the agents in real time in the form of these leaderboards.
09:43In these kernel settings, we also found it to be quite useful to have different agents with different personas, right?
09:49And these different personas actually corresponds to a different uh roles and priors that agents can actually have.
09:54So, 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,
10:02a third agent that looks at, you know, the precisions, the tensor computations.
10:06And these agents can and then across different personas, they can able to collaborate and a compete on the arena to speed up the kernels.
10:14And 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.
10:26So, 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
10:35and different uh hardware types, right?
10:37Where 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.
10:47And these improved kernels created designed by the agents are actually already used in in production at Together AI.
10:57So, 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,
11:07right?
11:07So, 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
11:15data science agents to solve complex data science problems.
11:21So, here in this DS Gym environment, we also curated and created a unified list of different data sets and tasks, right?
11:29So, these data sets can combine uh spans across many different settings.
11:33And the agents are then able to interact with these different data sets that we have through a unified uh interface and through code execution.
11:42In the DSGM environment, we also provide a unified infrastructure for the agents.
11:47So, for example, the agents can actually spin up many different Docker containers to test their data science algorithms and actually run them in parallel.
11:58So, 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.
12:10But we actually quickly realized that many of the existing widely-used benchmarks actually have many problems.
12:16And one big problem is that they're actually very vulnerable to shortcuts.
12:19By shortcut, I mean here is that uh down here what I'm showing are three different common popular data science benchmarks.
12:26Right?
12:26And the in green here we see shows like the performance of the agents on these benchmarks.
12:32Uh 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.
12:39Right?
12:40So, just by reasoning or by, you know, uh doing other shortcuts without actually actually do working with the underlying data sets.
12:47And 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.
12:57Which I think is uh really a significant problem with many of the existing benchmarks.
13:03So, to address that, we actually carefully curated at our own our own benchmarks, right, for both for scientific analysis and also for predictive modeling.
13:12So, 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.
13:21And then we also had human scientists and experts to review each of those tasks.
13:26And 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
13:34and and where also you have high quality data sets and also high quality evaluations.
13:39Then we curated those into the DS Gym as a kind of task for evaluating how well models agents can actually build predictive models.
13:48So all together in the DS Gym, we actually have created over a dozen different tasks.
13:53They span across dozens of different scientific domains ranging from biology to physics to economics.
14:01It also involves many different data types and data modalities.
14:06So this actually makes it very easy for us to evaluate different models, both open and closed source models.
14:12And 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.
14:22Right?
14:22So these are definitely not saturated benchmarks.
14:26We can also use a DS Gym as sort of like a training factory to improve these open source models.
14:31Right?
14:31So 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
14:40generated by the agents that have been verified through the through through actually executing the code from the agents.
14:47Right?
14:47So 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
14:58open source models in terms of solving these kind of data science tasks.
15:01Right?
15:01And these models are small enough that you can actually run them locally on your laptops and your computers.
15:08So just to summarize the this part was the data science gym.
15:11Right?
15:11So 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.
15:21We have carefully verified that there are no shortcuts in these tasks, which has been sort of a common challenge with existing data science benchmarks.
15:29And 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.
15:39So, 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.
15:50Right.
15:50So, then in the past, people have been building these AI systems mostly by designing individual models or individual tools.
15:56And currently, there's a lot of focus on creating designing agents or harnesses and workflows around agents.
16:02But 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,
16:13which is a set of infrastructure and incentives that in that motivates the agents that you solve more and more challenging problems.
16:20And 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.
16:31And here are some of the references for the papers that we published that describes these in more detail.
16:36So, thank you very much.
16:38[applause]
16:51[music]