Scientists Found A Better Language For AI Agents

[00:00] The number of AI agents on the internet

[00:02] is increasing at such an insane rate. I

[00:06] don't think I've seen anything like

[00:07] this. This is crazy. And this is an area

[00:10] that is quite new, and the technology is

[00:12] still pretty rough. Improving rapidly,

[00:15] but pretty rough. And the promise of

[00:17] agents is incredible. It would book the

[00:19] cheapest plane ticket for you, or run 24

[00:22] hours a day to manage your schedule,

[00:24] submit insurance claims, continuously

[00:27] scan a codebase for vulnerabilities and

[00:29] patch it. Well, this is the good, but at

[00:31] the same time, you get so many news

[00:34] headlines about spam, security issues,

[00:36] and system breakdowns. And it gets even

[00:39] tougher when you have not one agent, but

[00:42] multiple agents. Imagine two agents

[00:45] organizing a holiday for you. The flight

[00:47] agent hallucinates a cheaper airport 400

[00:51] miles away from your real destination.

[00:53] Then, the hotel agent says, "Let's book

[00:56] something super cheap nearby." Well,

[00:59] super cheap is often non-refundable. And

[01:02] now congratulations.

[01:04] You now have a non-refundable room you

[01:07] will never see.

[01:09] And so many of these problems come from

[01:11] the fact that agent coordination is

[01:13] super difficult. Now, check out what

[01:16] this paper says we should do. Here is a

[01:18] math problem. First agent writes a plan.

[01:21] The next one critiques it, and the third

[01:24] one solves the problem. And at this

[01:26] point, I said, "Okay.

[01:28] I see nothing interesting here. This is

[01:31] what everyone does with agents." Yes,

[01:34] but here's the key. Most agents

[01:36] communicate a bit like we do, in words.

[01:40] Wait a second. Why should we do that?

[01:42] Look at this neural interface for

[01:45] brain-to-text communication. Yes, this

[01:48] really works. You just think about a

[01:50] letter in the alphabet, and it magically

[01:53] appears. And if you keep doing this a

[01:55] lot, you start asking. The alphabet is

[01:58] optimized for writing.

[02:00] Why use that? Why not use one that is

[02:03] optimized for thinking? And what would

[02:05] that even look like? Hint, it would look

[02:08] like this. We talked about this 500

[02:10] videos ago, paper in the description.

[02:13] Now, if you look at the agents, the

[02:15] first one does some work, packs it up,

[02:17] and passes it to the next one. So do the

[02:20] second and the third ones. Every

[02:22] [clears throat] time an agent wants to

[02:24] communicate something, it has to write

[02:26] out full sentences, decode tokens one by

[02:30] one, and the next guy has to read and

[02:33] re-encode the whole thing.

[02:36] Why are we doing that? Who said they

[02:38] should talk in plain English? And this

[02:41] is the part where I fell off the chair.

[02:43] Now, hold on to your papers, fellow

[02:44] scholars, because this work says, "Huh,

[02:47] forget English. You know what? Forget

[02:49] letters entirely." It says, "Instead,

[02:52] let's link up their brains." Kind of.

[02:56] Instead of using English words, they

[02:58] pass raw undecoded numbers directly to

[03:01] the next agent. Send raw brain signals,

[03:04] if you will. Call it cross-agent latent

[03:07] state transfer. So, the theory is that

[03:10] these three agents can work together

[03:12] round one, round two, and round three

[03:15] much cheaper than the text-based agents.

[03:18] They refine an answer, and you get

[03:20] better answers with the same amount of

[03:22] computation. So, is it better? Hmm,

[03:26] let's see. Dear fellow scholars, this is

[03:28] Two Minute Papers with Dr. Károly

[03:30] Zsolnai Fehér. Well, when given

[03:32] competition-level math questions, it

[03:34] goes from 73% to 86%.

[03:38] That is crazy.

[03:40] We are talking free sub-10 billion

[03:42] parameter models, not expensive frontier

[03:45] systems. And here is where it gets the

[03:48] Michelin star status. Look at that.

[03:52] Ooh.

[03:53] Token usage down 75%.

[03:57] They all evaporated into the latent

[04:00] space. Loving it. So, this can improve

[04:03] smaller systems to be in striking

[04:05] distance of much bigger, more expensive

[04:08] models on difficult math problems. So, I

[04:11] bet it costs a fortune to train, right?

[04:14] Well, look at that.

[04:17] Four bucks. Basically, you spend your

[04:19] coffee money on these agents and in

[04:21] return they punch a hole in space-time.

[04:25] Love it. Additionally, it might even

[04:27] unlock Wait, wait, wait. I shouldn't say

[04:30] unlock. That's AI speak. So, it might

[04:33] give us a new scaling law. More rounds,

[04:37] better results. And at this point, I

[04:39] thought we might have a deadly flaw

[04:41] here. And it's really subtle. So, the

[04:44] training for each agent's role is

[04:47] written by a giant AI model. So, if they

[04:50] perform well, you have to ask, are

[04:53] things better because of the brain

[04:55] linking or is it good distillation from

[04:58] an excellent teacher? So, which one is

[05:01] it? A good teacher or a good

[05:03] architecture? Well, fellow scholars, we

[05:06] are in luck. This is a really good

[05:08] paper. So, the scientists thought about

[05:11] this too. And look, goodness, a

[05:14] controlled comparison gives the same

[05:16] teacher to other architectures and this

[05:19] one. And the new one still outperforms.

[05:22] So, yes, the brain linking really works.

[05:26] What a time to be alive. Okay, now,

[05:29] let's not get too excited. This is

[05:31] two-minute papers and we respect the

[05:33] science here. Limitations. One, tests

[05:36] were on smaller models. We don't yet

[05:39] know how these insights scale up to

[05:41] bigger ones. If they don't, then this

[05:44] puts small models on steroids.

[05:47] Still good. If yes, potential huge

[05:50] game-changer. Two, there is an optimal

[05:52] latent thought length, and that is about

[05:55] 80 steps. This is somewhat of a limit on

[05:58] how much thinking an agent can do per

[06:01] round.

[06:02] >> [clears throat]

[06:02] >> I am thinking, you know, if it solves a

[06:04] mathematical Olympiad problems already,

[06:07] how bad can that be? And sure enough,

[06:10] after 80, you don't get a lot of value

[06:12] anyway, but I wanted to mention it.

[06:14] Okay? So, code and models are available

[06:17] for free. Note that this is still very

[06:20] rough, very early, but it shows

[06:22] potential. And this is still research.

[06:25] Please do not think you just plug this

[06:27] in and everything will fly immediately.

[06:30] We need new tools for the era of LLMs,

[06:33] and Weights & Biases now has Weave, a

[06:36] lightweight toolkit to confidently

[06:38] iterate on LLM applications. Use traces

[06:41] to debug how data flows through each

[06:43] step of your app, and use evaluations to

[06:46] measure your progress. It is the best.

[06:49] Try it out now at wnb.me/papers,

[06:53] or click the link in the description

[06:55] below.