Skip to main content
Noam Brown

Noam Brown

Sep 17, 2026

9:09
So I, I, I wanna understand if in six months people will have access to multi-agent systems, how should one model what it is like to collaborate with or hire a multi-agent system?
9:20
Yeah.
9:20
I should start by talking about how these multi-agent systems actually work, which is, I think, a very different way than a lot of multi-agent systems in, in other AIs.
9:29
So a lot of people that have approached multi-agents for things like LLMs tend to take this like very scaffolded approach where, you know, for example, there might be a, uh, coordinator agent that delegates work to a bunch of children and gives them a task, and the children work on it and then return their answer.
9:47
Um, and this seems like a very sensible setup, very sensible scaffold.
9:51
It definitely helps, but there are, are a bunch of limitations with these kinds of setups.
9:55
So for example, if, if in this setup you have, um, a coordinator that's sending tasks to children, the children work on it and then return their answers, well, what happens if two children are given similar tasks? Can they talk to each other? And usually the answer is no.

37 MINS LATER

46:43
Uh, and then there's the question of will they be capable to do it? And I think that comes back to this question which a listener might disagree with on is just like, will there be billions of human-level or above intelligences, many of which are physically embodied in the world, within a matter of ten years or less, right? So if those two things are true, this Hugging Face thing is just, like, like, extremely analogous structurally, um, even if it's, like, quite boring or why it happened to how we totally lose control of the world.
17:40
Mm-hmm.
17:40
So Deep Research came out, I think it was, like, probably early 2025 it came out.
17:45
And, um, it was able to write detailed reports on anything you wanted.
17:50
You know, you wanted to research the semiconductor industry, it would go around and do a ton of research.
17:54
It would compile this, like, really comprehensive report with citations and deliver it to you.
17:58
Now- Is that easily verifiable? I, I would think it's actually pretty hard to, to grade the quality of a detailed research report on an advanced topic.
18:10
Um, it's not like grading whether a math question is correct or incorrect.

23 MINS LATER

41:05
Mm-hmm.
1:36
What was the motivation for it?
1:37
Yeah.
1:37
The motivation was we released 5.5, and the initial reaction was kind of skepticism that it was a substantially better model.
1:47
To, to be fair, that only lasted for a few hours before people had some time to play around with it and, and try it out themselves, and they saw that it was actually substantially better.
1:53
Um, but I think a lot of the skepticism came from the benchmark grid that was published.
1:58
Basically, whenever a new model's released, there's this benchmark grid where they show all these different benchmarks on, on the X-axis and then the performance of different models on the Y-axis, and you can just, like, compare different models.
2:08
It's like a single number for a model on a single benchmark.

19 MINS LATER

21:37
What's between us and there, then?
12:50
But I'm just curious about the play-by-play of, you know, we think that we have a solution to this problem, and what does it take to actually sort of, you know, believe that we've gotten all the way?
13:01
Yeah, and basically what happened was we trained a new model, and it wasn't designed for math or anything, it wasn't, it's a general purpose reasoning model.
13:12
And afterwards, we have people with, you know, math PhDs at OpenAI, some people that were formerly math professors that are interested in math and they just decided like kind of as a side thing, like, okay, well, let's see where the capabilities of this new model are.
13:28
Let's run it on a bunch of unsolved problems and just see like if any of them, you know, if it thinks that any of them are correct.
13:37
And of course, like one of the challenges is like actually verifying these things.
13:40
So basically the model, so they ran it through a bunch of problems.
13:43
And this is one of the problems where it came back and said like, actually it's solved.

23 MINS LATER

36:32
But I was curious if you felt like any difference in kind on the AI research side between pointing the model at contest math versus pointing it at open research problems.
AlessioHOST
49:05
Mm-hmm.
49:05
... get a good profile of, like, who this player is, like, how they're playing, where their weaknesses are.
49:10
Now, I think with more recent technology, that has come down, um, but still, the sample efficiency has been a big challenge.
49:16
Now, what's interesting is that after working on poker, I worked on Diplomacy.
49:22
I think we talked about this earlier.
49:23
And Diplomacy is this...
49:24
You know, it's a seven-player negotiation game, and-When we started working on it, I took a very game theory approach to the problem.
BooksHOST
54:28
Do you think this will hold for multi-agents, like self-play to improve better than humans?
3:13
It's coming up.
3:17
The incredible progress in AI over the past five years can be summarized in one word, scale.
3:25
Yes, there have been algorithmic advances, but the frontier models of today are still based on the same transformer architecture that was introduced in twenty seventeen, and they are trained in a very similar way to the models that were trained in twenty nineteen.
3:42
The main difference is the scale of the data and compute that goes into these models.
3:49
In twenty nineteen, GPT-2 cost about five thousand dollars to train.
3:55
Every year since then, for the past five years, the models have gotten bigger, trained for longer on more data, and every year they've gotten better.
4:06
But today's frontier models can cost hundreds of millions of dollars to train, and there are reasonable concerns among some that AI will soon plateau or hit a wall.

13 MINS LATER

speaker_4ADVERTISER
17:41
Yes.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.