DeepMind's New AI Just Changed Science Forever

[00:00] I appeared on camera for an interview not  so long ago. And I was really surprised   by how many of you Fellow Scholars said that  you would like to see more. So first of all,   Second, I thought let's try this and hope that you  will enjoy it. Dear Fellow Scholars, this is Two  

[00:20] Minute Papers with Dr. Károly Zsolnai-Fehér. Look, it only took 1,000 episodes. Now,   I have an amazing paper for you because scientists  at DeepMind did something pretty insane. Our  

[00:33] question today is can an AI invent something that  is fundamentally new and pushes humanity forward?  Well, they said that their new AI agent can  actually do research and even write research  

[00:48] papers. Most of the core content anyway. Is that insane? Well…it’s not. A lot of other   people have tried it and the only insane thing  about it was how many poor papers they wrote.

[01:03] You see, I visited the research group that is  behind this work last year. I flew to Mountain   View into this crazy lab, and a grumpy  guard didn’t even want to let me in first.

[01:18] Crazy town. So I was very surprised that they  are guarding these secrets and they take them   very seriously. What is even more surprising  is that now they give some of those secrets  

[01:30] away to all of us for free. Now that  is insane! More on that in a moment. So I talked to these scientists, this was the  research group of Quoc Le. They are brilliant.  

[01:43] They wrote an AI that was able to do a gold  medal worthy performance on the mathematical   olympiad. This is serious business. Then they  released this technique, anyone who is made out  

[01:55] of money bags and pays for the Gemini Advanced  can use it, it is called Deep Think. And now,   this AI is even better than that. They call  it Aletheia. Now that, once again is insane.

[02:11] Okay, so what does it do? Well, it  promises that it does research. It   solves novel problems. This is something  that could push humanity forward. Now that is so much harder than the mathematical  olympiad. Why is that? Well, in these contests,  

[02:29] you have a not that huge piece of core  knowledge you are supposed to have,   and every problem can be guaranteed to  be solved by those small set of tools. Every problem is nice, shiny, and polished. Tough,  

[02:46] but polished. You know what is not  polished at all? Real life problems.  With these open problems, we don’t even know  if they are solvable at all. Maybe they are   impossible, or maybe possible, but not with our  current tools. That’s the point: no one knows.

[03:03] When this technique is given a problem,  the generator starts working on it,   creates a candidate solution, and now here is  one of the important parts of the paper. The   verifier. This takes a look, and says, okay bro  this is junk. Start again. This is essentially  

[03:22] a filter. You know, that’s actually good life  advice. Sometimes it’s good to have a filter,   so you don’t just shoot those hot takes out  there into the ether. Now every now and then,  

[03:35] the solution looks pretty good, and could  maybe pass with a few modifications. Then,   it gets polished for another round of reviews,  and so it goes. Sounds simple…maybe even trivial  

[03:50] right? So what is so scientific about  this? Why doesn’t every system do that? Well, that’s easier said than done. In fact,  it is almost impossible to pull off. Why?

[04:02] One, when the AI is doing something  fundamentally new, unfortunately,   hallucinations still happen. Yup.  It just makes stuff up. Fake papers,   fictitious authors, you name  it. All kinds of junk comes out.

[04:17] Two, when you want to compute  1+1 or other simple things,   you have tons of training data about it out  there. You can verify that easily. But if  

[04:29] you want to do frontier research? There is  no training data on what we don't even know   yet. Of course there isn’t! You are trying  to invent things no one understands yet.

[04:41] These two factors make it extremely  difficult to get an AI to do something   fundamentally new and useful. So how did  they pull it off? With three key steps.

[04:53] First, Alethia does not use this formal rigid math  language to check its own proofs. It uses natural   English language. That is notoriously hard,  because when the AI checks its own writing,  

[05:07] it just blindly agrees with it.  We humans do that too! Now here,   the researchers found a way to separate the  thinking part from the answer part. So the  

[05:19] messy train of thought is hidden from the  verifier, it cannot trick itself into just   blindly agreeing with itself. Brilliant. Our  brains would need something like that too.

[05:32] Then, two they let the computer think  longer. That’s not new. However,   they added some optimizations to this, so  much so that the model they have now is  

[05:45] just as smart as the one from 6 months ago.  But hold on to your papers Fellow Scholars,   because yes, same smarts, but it uses a 100  times less compute. What! Crazy. They trained  

[06:00] a much stronger base model which made it  more efficient at reasoning. So this one,   even without internet access, beats the  mathematical olympiad gold AI easily. About  

[06:13] 65% was improved to 95%. Wow. It went from  a bit better than a coinfip to destroying   the tasks made for some of the best human minds.  All this in just a few months. I am out of words.

[06:27] Now three, they gave the AI the  ability to search for stuff. We   are talking about Google after all.  Once again, that is easy. However,   getting the AI to read and combine techniques  from dozens and dozens of cutting-edge research  

[06:43] papers without losing its mind. Now that is  hard. You saw it earlier, this really happens! They heavily trained this AI to be  able to use these tools and research   works that are out there. That was what  finally stopped it from making up junk.

[07:00] Okay, so how good is it? First I saw that  it solved a few of these Erdős problems. It   autonomously found the answer to 4 open math  puzzles left behind by a legendary Hungarian  

[07:14] mathematician. Is that insane? I asked  a mathematician friend. He told me yeah,   that’s pretty good, but there are  so many of these problems out there,   and not a ton of people work on them.  In other words, they are fairly easy,  

[07:29] they were just ignored by experts for  years. So not nearly as good as I thought. But then, it stepped up its game and  wrote the core contents of a research   paper. On something new. Note that the final  paper is written up by a human scientist.  

[07:47] They had one paper on calculating constants  in arithmetic geometry. And then it helped   human scientists write 4 other papers, like  finding new limits for interacting particles.

[07:59] Well, they are submitted for peer review  and that’s going to take quite a while. So,   many of them independent scientists. They  checked it for correctness and novelty,  

[08:16] and it checks out man. I think for the first  time ever, an AI created core parts of a   research work that is new, it has impact, it is  useful. That is…wow. What a time to be alive!

[08:33] So I told you there is levels to this  game. So where are we now? Level 0   is negligible novelty work, it can do  that. Level 1 is somewhat novel work,  

[08:45] it can do that too. But now, it can help a  person create publishable-level research.   Levels 3 and 4, those are groundbreaking  works, these are out of reach,  

[09:00] but I ask you Fellow Scholars, given the pace  of progress, for how long? For 6 more months?  And I think that is something that  needs to be talked about more.   Research helping the people  live a better life. Love it.

[09:15] And thank you so much to all of you Fellow  Scholars for watching us over the years.   We can only exist because of you Fellow  Scholars. I really hope that you enjoyed   this. It allows me to talk about papers  where there is not a lot of visual content,  

[09:31] and I really wanted to share this with you. Let  me know in the comments if we should do more.