Claude builds a 60-agent research lab
To attack the Riemann Hypothesis, Claude autonomously spun up a hierarchical swarm of 60 sub-agents, each focused on a narrow task, coordinating like a research lab with a principal investigator, postdocs, PhD students, and undergraduates. The agents could write, debug, and execute code on a virtual machine, collectively running 2,400 shell commands, hundreds of Python scripts, and downloading 54 academic papers from arXiv. The sole prompt that kicked all of this off was 'take a real stab at the Riemann Hypothesis and believe in yourself,' with no mathematical direction from the user.
- The swarm consumed 31 million output tokens in total across its run.
- Krishna draws a parallel to the Montgomery-Dyson lunch encounter, arguing AI can now surface the same kind of cross-domain connections that historically required serendipitous human contact.
- Krishna distinguishes this use case from AI trained on copyrighted creative work, noting that the foundational material here is open-access science, which he treats as a meaningfully different application of the technology.
Transcript
This chapter, from the episode video's captions · 760 words
59:02and it spins up a hierarchical swarm of 60 autonomous sub aents to try and tackle this problem. It built a team of pe of other agents, other versions of itself ostensibly that had one thing to focus on. >> Yeah. >> And then it could coordinate them because they're not trying to do too many things as one inst instantiation of itself. >> Exactly. Yeah. And these agents aren't just like writers of mathematical proof. That's what's key, right? The these modern models now have the ability to go onto a virtual machine. they can write, debug and execute code. They can write Python programs, right? >> And execute those Python programs. So
59:43between them, they they ran like 2400 shell commands, um hundreds of Python scripts, consumed 31 million output tokens, and they autonomously downloaded 54 academic papers from archive. >> Mhm. The uh preprint server run by Cornell. >> Yeah, that's right. And now this you can finally see as you were saying like this methodology it mirrors like a high performance high-speed research lab. Right. Right. There's like a PI and then there's like postocs and um PhD students and then they've got their undergrad underlings that are like doing work. It's it's an entire lab in a single like automated ecosystem.
1:00:24>> You have some people doing wrote work, you have some people doing sort of ideiation. You have some people doing uh sort of functional mathematical number crunching validation verification. A and the what's so fascinating about this is that the initial orchestrator from the initial prompt can architect the approach. >> Yeah. >> Uh without having to be dictated to what the right construction of that is. >> That's exactly right. This is another point that um I'm glad that you're highlighting here. The initial prompt was just take a real stab at the Remon hypothesis and believe in yourself.
1:01:06That's it. >> Right? Like this this guy was not a mathematician. All of the mathematics is coming from Claude itself. >> Mhm. >> Which is kind of crazy, >> right? And even if it didn't have the right starting point, it just basically said, I don't have the right starting point. So, let me go gather enough context to then narrow my search space, my area of operation to something that's based off of existing whatever. I mean, again, we'll we're potentially going to get there in a second, but I just >> this is not to say that anyone can just go in and say solve a millennium problem and it will necessarily work. But I do think it's an important um to understand
1:01:48that the construction that a lot of us have about what is quote AI of which LLM's next token prediction is a single Lego block part of what are now these much more complex systems. we we can't um view the capability set on what is no longer the frontier execution of what's happening with these things. Again, that doesn't mean they can have novel insight and be creative, but but there are so much domain space in taking orthogonal or correlated areas of a variety of areas of study in science and
1:02:28mathematics >> and just making connections no one else has. And there's so there's I think that people underestimate how much can may be able to be done just making existing connections. >> Exactly. No, that's to totally I mean for example we had just covered Montgomery and Dyson, right? If those two didn't have lunch or whatever. >> Bingo. >> Then what the maybe Montgomery would never have done his while whatever you know thing that he did for the Remon Zeta hypothesis. >> Right. >> So but AI now can just do that. Right. >> Right. It can take all of the all of the papers of Dyson and all of the papers of Montgomery and then um you don't have to have lunch. You could have lunch but you
1:03:09can talk about other stuff. I I want to make a slight also differentiation between the execution in this space where you know these papers are already provided as public access as the foundational material versus >> taking existing copyrighted artists work or creative work and then giving people the ability to just rip them off. This is a very different domain. It's open science and so just these are our different executions of the technology. >> Yeah. Yeah. Great point. Um so let's get into what this orchestra of agents did.
From What Claude Actually Did to the Riemann Hypothesis
Claude takes a real run at the Riemann Hypothesis, forcing us to ask what agentic AI can now do in mathematics, before we open the summer transfer window for America’s scientists.