DeepMind’s New AI Found A Strange New Way To Think

[00:00] DeepMind's new AI just did something

[00:02] amazing, or did it? You see, there was a

[00:05] legendary mathematician called Paul

[00:07] Erdős, fellow Hungarian, who left more

[00:10] than a thousand open problems to the

[00:12] world to solve. Look, we Hungarians have

[00:16] a lot of problems.

[00:17] We got to contribute somehow, and this

[00:19] is our way of doing it. Now, DeepMind's

[00:22] new AI called AlphaProof Nexus tried to

[00:25] solve about 350 of them and came up with

[00:29] a 95.7%

[00:31] failure rate. Basically, it solved nine,

[00:34] and it only cost a couple hundred

[00:36] dollars per problem. Is that good? Well,

[00:39] I got to say, that is incredibly super

[00:42] good. Why? Well, these are decades-old

[00:45] problems that were not solved by anyone

[00:47] yet. The other line of criticism I hear

[00:49] is that this did not do fundamentally

[00:52] new things. Is that a problem?

[00:55] I think not. Why? Well, let's look back

[00:58] to 4 years ago. GPT-3. People said,

[01:01] "Well, it can't even add numbers

[01:04] together reliably." Then, 2 years ago,

[01:07] people said, "Well, it can't even solve

[01:09] high school competition problems

[01:11] reliably." Then, 1 year ago, people

[01:14] said, "It can't even win the

[01:16] Mathematical Olympiad gold medal

[01:18] reliably." And today, they are saying,

[01:20] "Well, it can't even solve 50-year-old

[01:24] unsolved problems reliably."

[01:27] Do you see where this is going? It is

[01:29] clear as day. Please apply the first law

[01:32] of papers here. It says, "Do not look at

[01:35] where we are. Look at where we will be

[01:37] two more papers down the line." And this

[01:39] result is absolutely amazing, stunning,

[01:42] even. So, how did they do it? How is

[01:45] that even possible? Dear fellow

[01:47] scholars, this is Two Minute Papers with

[01:49] Dr. Károly Zsolnai Fehér. Normally, you

[01:51] would reach out to some AI assistant to

[01:54] take a crack at it, but it won't solve

[01:56] it because they hallucinate and make

[01:58] things up. To avoid that, they make it

[02:00] use Lean, a formalized mathematical

[02:03] language where it's easy to check

[02:05] whether your proofs are correct. Is this

[02:07] new? Not at all.

[02:09] Everyone is doing that today. Okay, so

[02:12] what's new here? Look, first, a

[02:14] mathematician writes down the problem in

[02:16] Lean and the solution. The proof is left

[02:19] blank. Then, the AI agent tries to solve

[02:22] it. Of course, it fails. Too hard. Then,

[02:26] another AI checks it and says, "Mhm,

[02:28] this is not great." But, it also says

[02:31] why it's not great. But, here's the key,

[02:34] this guy right here. This is a cheaper

[02:37] judge AI that reads two previous

[02:39] solutions and picks a winner. Both

[02:42] solutions can be wrong, but it picks the

[02:45] one that is a bit better. Now, this is

[02:48] genius. Why? Well, hold on to your

[02:51] papers, fellow scholars, because it's

[02:53] kind of like a chess system where the

[02:55] solutions are the players and each of

[02:58] these players gets an ELO score, also

[03:01] named after Arpad Elo, fellow Hungarian.

[03:04] Look, sometimes we provide solutions,

[03:07] too. So, each proof now has a score. And

[03:10] now, we start again. But, not from

[03:13] scratch. No, no, no. We start out from

[03:16] the highest scoring bad solution. So,

[03:19] this is now a tournament. Do this over

[03:22] and over again. So cool. And now, we

[03:25] keep running and running this tournament

[03:27] until the validator says, "Yep, this one

[03:30] checks out." And then, we have a formal

[03:33] proof. Nailed it. This is incredible

[03:36] because it takes an unreliable AI, runs

[03:39] it over and over again, and it can lie

[03:42] its rear end off as much as it wants,

[03:45] and we still get a reliable system out

[03:48] of this. A reliable system built out of

[03:51] unreliable parts. I love that. And the

[03:54] fact that they put all this research out

[03:56] there in the open for free for all of

[03:59] us.

[04:00] Chef's kiss. Thank you so much for

[04:02] everyone who worked on this. What a time

[04:04] to be alive. But wait, interestingly,

[04:08] the story of AI so far has been that we

[04:11] make it smarter. Now, the story has

[04:13] changed. We don't need to make it

[04:15] smarter, we need to make the harness

[04:17] around it tighter. Give it a good judge.

[04:20] Let it a thousand times and it will

[04:23] slowly work out the right solution to

[04:26] incredibly hard problems. So here, the

[04:28] intelligence is not just in the model,

[04:31] but it is in the loop around it.

[04:34] Everyone is experimenting with different

[04:36] kinds of loops and it is super fun. I do

[04:39] it too on lambda. Okay, not even this

[04:41] technique is perfect. Limitations. In

[04:44] other words, the stuff that you don't

[04:45] hear about in mainstream media. So one,

[04:49] why not test on the full 1200 Erdős

[04:52] problems? Well, there is a little

[04:54] selection bias here. I think they took a

[04:57] subset of 350 that was easier to

[05:00] formalize. Is that a problem? In my

[05:03] eyes, not at all.

[05:04] You got to start somewhere. Let's not be

[05:06] one of those people that say, well, it

[05:09] can't even solve the 50-year-old

[05:11] unsolved problems reliably. What it has

[05:14] achieved is incredible. Now, two,

[05:17] smaller models solved zero problems.

[05:19] Zero. Nothing. You still need a beefy AI

[05:23] system at the core. That is an

[05:25] interesting case because people keep

[05:27] showing these benchmarks where the super

[05:29] fast cheap model is just a couple

[05:32] percentage points away from the

[05:33] frontier. And whenever I try them, they

[05:36] always seem a great deal weaker. This

[05:39] seems to reinforce that. Also, people

[05:41] will probably start thinking, do I use a

[05:44] larger model with fewer tournament

[05:47] rounds or do I use a smaller one with

[05:50] more? Assume that they cost the same.

[05:53] Interesting question. Now, where does

[05:55] this put us? Well, an AI just solved

[05:58] nine math problems that no human could

[06:00] crack in 56 years for a couple of

[06:04] hundred dollars each, and they did it by

[06:06] letting an unreliable AI fail thousands

[06:09] of times against a judge that cannot

[06:12] lie. And we went from can't even add

[06:15] numbers to solving decades-old open

[06:18] problems in the span of four years. And

[06:21] I think that is insane. But, limitations

[06:24] apply. Also, models used to be the only

[06:27] thing that matters. Now, harnesses,

[06:29] loops around them, also matter. Now, I

[06:32] recently talked to Pushmeet, one of the

[06:34] leaders of the project, and he's

[06:36] amazing. I am just a student who loves

[06:39] to travel the world and tries to learn

[06:41] from incredible scientists like him and

[06:44] bring that knowledge to you fellow

[06:45] scholars. And it [clears throat] is a

[06:47] huge honor for me to be able to talk

[06:49] about it to such a super smart audience

[06:51] as you fellow scholars. Subscribe and

[06:54] hit the bell if you feel that this is

[06:56] the way of doing it. Thank you so much

[06:58] for being with me all these years and

[07:00] over more than a thousand videos. We

[07:03] need new tools for the era of LLMs, and

[07:07] Weights & Biases now has weave, a

[07:09] lightweight toolkit to confidently

[07:11] iterate on LLM applications. Use traces

[07:14] to debug how data flows through each

[07:17] step of your app, and use evaluations to

[07:19] measure your progress. It is the best.

[07:22] Try it out now at wnb.me/papers,

[07:27] or click the link in the description

[07:29] below.