[00:00] I appeared on camera for an interview not&nbsp; so long ago. And I was really surprised&nbsp;&nbsp; by how many of you Fellow Scholars said that&nbsp; you would like to see more. So first of all,&nbsp;&nbsp; Second, I thought let's try this and hope that you&nbsp; will enjoy it. Dear Fellow Scholars, this is Two&nbsp;&nbsp;
[00:20] Minute Papers with Dr. Károly Zsolnai-Fehér. Look, it only took 1,000 episodes. Now,&nbsp;&nbsp; I have an amazing paper for you because scientists&nbsp; at DeepMind did something pretty insane. Our&nbsp;&nbsp;
[00:33] question today is can an AI invent something that&nbsp; is fundamentally new and pushes humanity forward?&nbsp; Well, they said that their new AI agent can&nbsp; actually do research and even write research&nbsp;&nbsp;
[00:48] papers. Most of the core content anyway. Is that insane? Well…it’s not. A lot of other&nbsp;&nbsp; people have tried it and the only insane thing&nbsp; about it was how many poor papers they wrote.
[01:03] You see, I visited the research group that is&nbsp; behind this work last year. I flew to Mountain&nbsp;&nbsp; View into this crazy lab, and a grumpy&nbsp; guard didn’t even want to let me in first.
[01:18] Crazy town. So I was very surprised that they&nbsp; are guarding these secrets and they take them&nbsp;&nbsp; very seriously. What is even more surprising&nbsp; is that now they give some of those secrets&nbsp;&nbsp;
[01:30] away to all of us for free. Now that&nbsp; is insane! More on that in a moment. So I talked to these scientists, this was the&nbsp; research group of Quoc Le. They are brilliant.&nbsp;&nbsp;
[01:43] They wrote an AI that was able to do a gold&nbsp; medal worthy performance on the mathematical&nbsp;&nbsp; olympiad. This is serious business. Then they&nbsp; released this technique, anyone who is made out&nbsp;&nbsp;
[01:55] of money bags and pays for the Gemini Advanced&nbsp; can use it, it is called Deep Think. And now,&nbsp;&nbsp; this AI is even better than that. They call&nbsp; it Aletheia. Now that, once again is insane.
[02:11] Okay, so what does it do? Well, it&nbsp; promises that it does research. It&nbsp;&nbsp; solves novel problems. This is something&nbsp; that could push humanity forward. Now that is so much harder than the mathematical&nbsp; olympiad. Why is that? Well, in these contests,&nbsp;&nbsp;
[02:29] you have a not that huge piece of core&nbsp; knowledge you are supposed to have,&nbsp;&nbsp; and every problem can be guaranteed to&nbsp; be solved by those small set of tools. Every problem is nice, shiny, and polished. Tough,&nbsp;&nbsp;
[02:46] but polished. You know what is not&nbsp; polished at all? Real life problems.&nbsp; With these open problems, we don’t even know&nbsp; if they are solvable at all. Maybe they are&nbsp;&nbsp; impossible, or maybe possible, but not with our&nbsp; current tools. That’s the point: no one knows.
[03:03] When this technique is given a problem,&nbsp; the generator starts working on it,&nbsp;&nbsp; creates a candidate solution, and now here is&nbsp; one of the important parts of the paper. The&nbsp;&nbsp; verifier. This takes a look, and says, okay bro&nbsp; this is junk. Start again. This is essentially&nbsp;&nbsp;
[03:22] a filter. You know, that’s actually good life&nbsp; advice. Sometimes it’s good to have a filter,&nbsp;&nbsp; so you don’t just shoot those hot takes out&nbsp; there into the ether. Now every now and then,&nbsp;&nbsp;
[03:35] the solution looks pretty good, and could&nbsp; maybe pass with a few modifications. Then,&nbsp;&nbsp; it gets polished for another round of reviews,&nbsp; and so it goes. Sounds simple…maybe even trivial&nbsp;&nbsp;
[03:50] right? So what is so scientific about&nbsp; this? Why doesn’t every system do that? Well, that’s easier said than done. In fact,&nbsp; it is almost impossible to pull off. Why?
[04:02] One, when the AI is doing something&nbsp; fundamentally new, unfortunately,&nbsp;&nbsp; hallucinations still happen. Yup.&nbsp; It just makes stuff up. Fake papers,&nbsp;&nbsp; fictitious authors, you name&nbsp; it. All kinds of junk comes out.
[04:17] Two, when you want to compute&nbsp; 1+1 or other simple things,&nbsp;&nbsp; you have tons of training data about it out&nbsp; there. You can verify that easily. But if&nbsp;&nbsp;
[04:29] you want to do frontier research? There is&nbsp; no training data on what we don't even know&nbsp;&nbsp; yet. Of course there isn’t! You are trying&nbsp; to invent things no one understands yet.
[04:41] These two factors make it extremely&nbsp; difficult to get an AI to do something&nbsp;&nbsp; fundamentally new and useful. So how did&nbsp; they pull it off? With three key steps.
[04:53] First, Alethia does not use this formal rigid math&nbsp; language to check its own proofs. It uses natural&nbsp;&nbsp; English language. That is notoriously hard,&nbsp; because when the AI checks its own writing,&nbsp;&nbsp;
[05:07] it just blindly agrees with it.&nbsp; We humans do that too! Now here,&nbsp;&nbsp; the researchers found a way to separate the&nbsp; thinking part from the answer part. So the&nbsp;&nbsp;
[05:19] messy train of thought is hidden from the&nbsp; verifier, it cannot trick itself into just&nbsp;&nbsp; blindly agreeing with itself. Brilliant. Our&nbsp; brains would need something like that too.
[05:32] Then, two they let the computer think&nbsp; longer. That’s not new. However,&nbsp;&nbsp; they added some optimizations to this, so&nbsp; much so that the model they have now is&nbsp;&nbsp;
[05:45] just as smart as the one from 6 months ago.&nbsp; But hold on to your papers Fellow Scholars,&nbsp;&nbsp; because yes, same smarts, but it uses a 100&nbsp; times less compute. What! Crazy. They trained&nbsp;&nbsp;
[06:00] a much stronger base model which made it&nbsp; more efficient at reasoning. So this one,&nbsp;&nbsp; even without internet access, beats the&nbsp; mathematical olympiad gold AI easily. About&nbsp;&nbsp;
[06:13] 65% was improved to 95%. Wow. It went from&nbsp; a bit better than a coinfip to destroying&nbsp;&nbsp; the tasks made for some of the best human minds.&nbsp; All this in just a few months. I am out of words.
[06:27] Now three, they gave the AI the&nbsp; ability to search for stuff. We&nbsp;&nbsp; are talking about Google after all.&nbsp; Once again, that is easy. However,&nbsp;&nbsp; getting the AI to read and combine techniques&nbsp; from dozens and dozens of cutting-edge research&nbsp;&nbsp;
[06:43] papers without losing its mind. Now that is&nbsp; hard. You saw it earlier, this really happens! They heavily trained this AI to be&nbsp; able to use these tools and research&nbsp;&nbsp; works that are out there. That was what&nbsp; finally stopped it from making up junk.
[07:00] Okay, so how good is it? First I saw that&nbsp; it solved a few of these Erdős problems. It&nbsp;&nbsp; autonomously found the answer to 4 open math&nbsp; puzzles left behind by a legendary Hungarian&nbsp;&nbsp;
[07:14] mathematician. Is that insane? I asked&nbsp; a mathematician friend. He told me yeah,&nbsp;&nbsp; that’s pretty good, but there are&nbsp; so many of these problems out there,&nbsp;&nbsp; and not a ton of people work on them.&nbsp; In other words, they are fairly easy,&nbsp;&nbsp;
[07:29] they were just ignored by experts for&nbsp; years. So not nearly as good as I thought. But then, it stepped up its game and&nbsp; wrote the core contents of a research&nbsp;&nbsp; paper. On something new. Note that the final&nbsp; paper is written up by a human scientist.&nbsp;&nbsp;
[07:47] They had one paper on calculating constants&nbsp; in arithmetic geometry. And then it helped&nbsp;&nbsp; human scientists write 4 other papers, like&nbsp; finding new limits for interacting particles.
[07:59] Well, they are submitted for peer review&nbsp; and that’s going to take quite a while. So,&nbsp;&nbsp; many of them independent scientists. They&nbsp; checked it for correctness and novelty,&nbsp;&nbsp;
[08:16] and it checks out man. I think for the first&nbsp; time ever, an AI created core parts of a&nbsp;&nbsp; research work that is new, it has impact, it is&nbsp; useful. That is…wow. What a time to be alive!
[08:33] So I told you there is levels to this&nbsp; game. So where are we now? Level 0&nbsp;&nbsp; is negligible novelty work, it can do&nbsp; that. Level 1 is somewhat novel work,&nbsp;&nbsp;
[08:45] it can do that too. But now, it can help a&nbsp; person create publishable-level research.&nbsp;&nbsp; Levels 3 and 4, those are groundbreaking&nbsp; works, these are out of reach,&nbsp;&nbsp;
[09:00] but I ask you Fellow Scholars, given the pace&nbsp; of progress, for how long? For 6 more months?&nbsp; And I think that is something that&nbsp; needs to be talked about more.&nbsp;&nbsp; Research helping the people&nbsp; live a better life. Love it.
[09:15] And thank you so much to all of you Fellow&nbsp; Scholars for watching us over the years.&nbsp;&nbsp; We can only exist because of you Fellow&nbsp; Scholars. I really hope that you enjoyed&nbsp;&nbsp; this. It allows me to talk about papers&nbsp; where there is not a lot of visual content,&nbsp;&nbsp;
[09:31] and I really wanted to share this with you. Let&nbsp; me know in the comments if we should do more.