Will Quantum Computing Transform AI? Adam Brown of Google DeepMind & Stanford | AI Rabbit Hole 2025

kind of goes very well into hello hello so it goes very well into like what we just discussed you know world descriptor models and AGI so Adam is probably the smartest person I know and I do know a lot of very smart people so I wanted to ask you a few a few questions today about you know quantum computing AI I so the first elephant in you know elephant in the room is everybody hypes up quantum as the future of AI that it's going to make models require fewer weights and you know biases like fewer parameters it it's just going to speed up everything exponentially is it so well first of all thank you for having me here what an absolutely wonderful conference hello everybody basically there's a short answer and a long answer to that question and the short answer is is no. The the long answer is that quantum computing has a there are two different reasons at the moment why quantum computing is not going to impact AI very much.

one is that we simply don't currently have quantum computers or the quantum computers we do have are so bad that they're completely useless. that is a development project that's been going on for decades. will continue. People are making valiant efforts, but it's going to be a while before we have a quantum computer. The second half of the answer is that even if we had a fully functional error corrected, proper quantum computer that actually worked, which we do not currently have. even if we had a working quantum computer, it basically doesn't help very much with AI. Quantum computers are extremely powerful. Well, once they work, they're extremely powerful at the a special purpose set of tasks for which they've been specifically designed. So, the most famous such task is factoring large composites. So, if you could do that, you can break all public key crypto systems that are currently in widespread use.

It's also very good at simulating quantum systems themselves. A quantum computer is good at simulating systems for which quantum mechanics is significant. So it'll help a lot with quantum chemistry, but what it will not do is take you a large language model and somehow magically allow you to infer what the weight should be or to do inference any faster. This is super helpful. So should we all sell our quantum stocks? the best time to sell your quantum stocks would have been a couple of months ago. in terms of forward-looking statements, I don't have any I don't have any opinion, but there has been the quantum stocks have been going very strange of late. There was an an announcement that Google had made a a better quantum chip, still very very far from a working quantum computer. And all of the quantum stocks, including those that go would go out of business if Google succeeded, went up by about a factor of 10.

they they've now come back down again. But yeah, it's very difficult to reconcile that with the efficient market hypothesis. Well, I think understanding how stock markets works is is a little bit easier than understanding quantum computing. so, you know, a lot of people just go with the sentiment, oh, quantum computing, I heard about it, like what quantum stocks should I invest in? But shifting gears. So you once said that LLMs are just interpolators climbing the ladder of abstraction, meaning that they eventually learn to think in more abstract terms, but they cannot do out out of the box thinking. Do you think it's it's it's it's a it's a big limitation and what would it take for AI to actually do something useful? well, AI can certainly do something useful even merely via interpolation. I I do think I do think that it'll probably be just arbitrarily useful.

and there's nothing that they can't do. I think I think what we've learned over the last few years, the main takeaway for me is that what it's easy to dismiss these things, neural networks and and LLMs in particular, as just interpolators and pattern matches. I think what we've learned is how powerful pattern matching is and that when you discuss interpolation you need to be very clear what level of abstraction you're discussing or maybe I should just say that like less philosophically that you know to begin with they just learn the the exact training set that they've been given if you if you're not careful and then they learn to generalize from there at a sort of higher level abstraction and then they just keep going to higher and higher levels of abstraction and I think If if you look at what we do, my you know cut my teeth as a theoretical physicist studying black holes and quantum mechanics and similar if you look at what the day job of a theoretical physicist is a lot of it is just abstraction at at is just interpolation at different layers levels of abstraction.

So you know first of all you can solve the exact problem you're giving in a homework then you can solve all problems that are like that even if you've never seen them before. and then you just continue in that manner. If I examine any of my my physics papers, they tend to be in some sense just combining ideas that already exist in a novel fashion. But right so I don't I I actually am bullish on the ability of large language models to just just wrap up all of theoretical physics. When do you think AI can interpolate to the extent that it rediscovers something major that we have already discovered like general relativity? I think general relativity. So general relativity is is you know a subject I absolutely love. It's like invented by Einstein in 1915. The sort of perhaps the most beautiful thing that the human mind has ever created.

I have taught the general relativity course at Stanford. Like I'm absolutely love it. I probably that is the hardest thing that we will ever call at least in the physics domain maybe ever call on an LLM to do. So you're you're asking for the final boss in which maybe you plug in all physics as it was known in 1900 and no physics after that and also you know just mention that the perihelion the orbit of Mercury seems to be behaving anomalously anomalously in this way and see if see if the LLM can figure it out. I think that will be the once it can do that, it can basically do everything, at least in the physics domain. I think we're probably at least 10 years away from from that, but maybe no more than 10 years. It would already be able to figure out the mathematics and it would just need to make the connection.

I think it could easily be within 10 years. This is very cool. I I love the 10-year term because our funds are 10 year long. Okay, great. So, yeah, like we're safe. We're safe for 10 years, dear investors. so in one of your conversations you also reflected that AI development you know innovation in AI is happening much quicker than it happens in theoretical physics. do you think AI can speed this up since most of most of the papers are interpolation as you say? it is definitely the case that AI is cons advances considerably faster than theoretical physics. So I, you know, I work in both and I put out a physics paper last year that I had originally written on the train to Santa Barbara in 2013 and then just sort of filed away on a desk and thought about for a decade and then finally felt it was ready to to put out and see the world.

That does not fly in AI at all. If I look back on the stuff that we were doing a year ago even, it just seems like the stone age and everything's moving substantially faster. Partly that's just a reflection of the resources and the lowhanging fruit that's available in that field. In terms of turning it around and can we use LLMs to speed up theoretical physics, there's there's kind of two components to that question. One is can they just do it on their own without our help. and as we discussed, we're not quite there yet, but I think that's coming. What we can do today is use them as tools. And so for sure the things that they're you know unbelievably useful for for a working theoretical physicist is search in a way that's way better than than anything we've ever had before. And that sounds like a trivial use case but but really it's not.

The amount of stuff that humanity has figured out is so much larger that I know in my head that the ability to have access to this, you know, Russian physicist writing not in English in the 1990s figured this thing out that's super useful to me. Before that was just completely inaccessible. And now I have access to the brains of anybody who actually wrote anything down and put it on the internet. And these LLMs are able to search and find that. That's that's one level. and then there are other more there are other more sophisticated things they can do for you as well which I could describe to you. So I if AI can help you know if if it can do stuff on its own. Yeah. theoretically what is preventing us? Is it like do we need more data of some sort? Do we need new architecture?

Do we need like a breakthrough or or better chips or more compute? all of those things would help. my guess is we do not need a new architecture. The the transformer architecture will get us there. It's super powerful, super adaptable, maybe not the optimal thing to do, but the architecture itself is at this stage not really holding us back. We probably need better improved RL algorithms. I would say obviously the last year has seen a great improvement in applying these RL techniques that were so useful in the context of of chess and go now turning them back towards large language models. we probably need even more and better of those. We also need to just collect better data. I would say that that's one limit at the moment. that's a big thing that's going on is people are going and trying to mine all of this scientific high quality scientific data with which to do RL.

That's a so that's something else we need and more compute is always handy but actually the reinforcement learning is not really limited by compute at the moment. Pre-training is limited by compute where we're taking all of these resources more chips than we had in the entire world not very long ago and devoting them to doing massive pre-training runs. At the moment, even people who are figuring out how to scale RL compute are not particularly limited by by compute in in one run. One of the things I think it could also help is that right now I think people publish much less than they have in their heads. How do you think we can encourage the scientific community to put out more thoughts, ideas to help train the omnipotent LLM that would you know figure out all of the questions of the universe? Yes. So for scientists in academia, I think it is certainly the case, you know, why do you even want to be in an academic department?

mainly it's you know informal chats with my colleagues or things that that they they don't write down anywhere but that they just know in their head and that the actual amount that you publish in published journals is a tiny fraction of that you know 1% and I also have this you know having come through academia I have this feeling that I should only put it out if it's like this perfect thing that you know looks good to a nature referee and not just some random thought I had. of course that is completely antithetical to the LLM philosophy as you point out. What we should really be encouraging everybody to do is to take everything that they've ever thought, however low quality, and just dump it on the internet where an LLM can read it and mine it for good ideas and we should be encouraging people to be not perfectionists and putting it putting it all out there.

So yeah, maybe like a there's already the archive which is sort of one step in that direction where you can just put things pre-referee but there basically the quality is almost referee quality already. maybe you should make a website being a man of action where people can just dump any any old nonsense that they had and there's no social penalty for having incomplete thoughts. Yeah, that could be built. that that is easy to build but it's going to be hard to persuade everybody to kind of automate themselves out of the job. I think vanity is another reason people like want to get published. in theoretical physics it is because of the archive it is already the case that publication itself is not really considered a big event. I you know I consider the paper to be finished when I put it on the archive and when it's later published none of my colleagues notice because they read it when it went on the archive.

So you spoke that you know about intuition and creativity that some people like Feynman or Einstein brought to the field what do you think they could accomplish if they had LLMs in their age? I think that they would find them extremely useful for the reasons I said partly just as force multipliers they're we discussed semantic search which is extremely useful we also discussed they can just be like graduate students doing your calculations for you your low-level calculations for you and just force multiply all these people even if you look back at Einstein's discovery of general relativity a lot of that he had the correct idea a decade before and a lot of it was going and learning from the Eastern European mathematicians, the mathematics required and you know he you know like many of us didn't find them particularly great teachers or forbidding teachers and so he he kept on making mistakes.

I think a large language model would be a much more patient tutor for him about this topics that were known and technically in the literature but that he had trouble getting Hilbert to explain clearly to him. So 10 years, he would have saved 10 years possibly. Yeah. and he made some some errors that I think an LLM would find it easy to catch as well. I I also remember reading that Feynman was doing biology for summer and he almost discovered that all other proteins can be made out of other proteins. for which I think 30 years later somebody got a Nobel Prize. But he screwed up the lab protocol. So his experiment went wrong. But yeah, I I guess LLMs could could help solve lab protocols as well. Oh, the beautiful new world. And well, we're looking forward to working quantum computing. U like we we now know that it's not going to be probably a super big deal for LLMs.

it'll be a big deal for chemistry and it'll be a big deal for cryptography. It's just not going to be a big deal for AI. is going to be a very big deal for humanity. Yeah. But in all of the new materials, all of the new science that we can do, but 10 years is a is a good term. Thank you so much, Adam, for being with us. guys, Adam Brown. Okay. Thank you for having me.