Googles New Breakthrough is BIGGER Than You Think! (Googles Q*)

So there was a recent breakthrough by one of the Frontier Labs, Google, DeepMind. And I think most people haven't realized the gravity of the situation here. DeepMind managed to be the first frontier lab to solve international Mathematical Olympiad problems at a silver medalist level. And this combines Alpha Proof, a new breakthrough model for reasoning, and Alpha Geometry 2, an improved version of their previous system. Now, through this was something that I think people did gloss over a little bit because they didn't realize the impact of this work. But I think that this is most certainly probably one of the top five most important breakthroughs this year. And I'm going to get into why that is and why you should be paying attention to this research and what it means, because not only was it incredible, there was also some other pieces of information surrounding. So essentially they state here, the breakthrough models Alpha Proof and alpha geometry 2 solve advanced reasoning problems in mathematics. Artificial general intelligence with advanced mathematical reasoning has the potential to unlock new frontiers in science and technology. We've made great progress in building AI systems that help mathematicians discover new insights, novel algorithms and answers to open problems. But current AI systems still struggle with solving general math problems because of limitations in reasoning skills and training data. And this is where the craziness starts. Today we present Alpha Proof, a new reinforcement learning based system for formal math reasoning, and Alpha Geometry 2, an improved version of our geometry solving system. Together they managed to solve four out of six problems from this year's International Mathematical Olympiad. Now, if you don't know what the International Mathematical Olympiad is, you probably won't realize why this is so crazy. This is the oldest and largest most prestigious competition for young mathematicians and held annually since 1959. And each year elite pre college mathematicians train, sometimes for thousands of hours to solve six exceptionally difficult problems in algebra, combinatorics, geometry and number theory. And basically this is one of the challenges for artificial intelligence. And many people do predict that once we have a system that's able to get gold, we'll know that truly capable systems are here. Now you can see here that this recent. So you can see here the score on the IMO 2024 problems. You can see the graph showing performance of our AI system relative to human competitors. At the IMO 2024, we earned 28 out of 42 total points, achieving the same level as a silver medalist in the competition. So you can see that it literally was one point away from getting the gold medal. I think actually if it's 28, then that means it's two points away, but nonetheless it is a fine margin to getting the gold. But this is absolutely insane because one of the things I realized that Google have done is they've gone back to some of their old architectures that they previously used to. Now, if you aren't familiar with what Google have previously done in the past, they've actually done a huge amount of different AI projects that have been really successful and they've even managed to create superhuman AI systems. Now, essentially the reason why I say that this is so crazy and the reason why this excited me was because if we actually take a look at what Google have done here, you can see that they actually described that this is a neurosymbolic hybrid system in which the language model was based on Gemini, which and trained from scratch on an order of magnitude more synthetic data than its predecessor. This helped the model tackle much more challenging geometry problems, including problems about movements of objects and equations of angles, ratio or distances. And Alpha Geometry 2 employs a symbolic engine that is two orders of magnitude faster than its predecessor. When presented with a new problem, a novel knowledge sharing mechanism is used to enable advanced combination of different search trees to tackle more convex problems. And the reason that this is crazy is because neurosymbolic AI, from what we've seen in early experiments, even on some of the hardest benchmarks, have proven to consistently generate results that surprise even the most well researched researchers. So that's why this is so crazy, because if Google managed to continue pushing the bounds on neurosymbolic AI, I do believe that they're likely to make an increasing number of breakthroughs and increasingly more powerful systems in terms of their reasoning capabilities. If you want to take a look at a kind of neuro symbolic system that was there before, if you remember AlphaGo Zero, there was AlphaGo Zero, which was a form of AlphaGo, which was a lot better. But essentially this model basically surpassed the previous AlphaGo and actually previously then went to surpass the AlphaGo master. So AlphaGo Zero was a system that basically managed to train itself and essentially managed to master go in just 21 days. This was a new apart, this was a new approach. And you can see that in 40 days that it surpasses all other versions of Go and becomes the best Go player in the world. And it does this entirely from self play. Now of course I'm not saying that you can apply this entire concept to LLMs, but the point here is that one of the things, if you remember and if you've been paying attention to some of the reports floating around, is that they've been saying that Look, AI is running out of training data, we're running out of data, what are we going to do, yada yada, yada. But, but one of the things that people have been slowly exploring is the fact that NeuroSymbolic AI improves AI's reasoning capabilities by using many different things. One of the things is of course, tool use and of course different ways to search and solve different reasoning problems. And I think that this method is something that we've seen time and time again in all the research papers that I've looked at. This is something that increases the reasoning ability of these models. So basically you have this researcher called Francis Chollet. So he's a French software engineer and computer scientist working at Google. Now, essentially he created a benchmark that is rather hard for current AI systems. And it's kind of benchmark that is not subject to contamination. So it's not something that is leaked in the training data where AIs can plan for it, and it's not something they can memorize either. So this is a really hard benchmark. Now what's crazy about this is that he basically said, to be clear, I've never claimed that solving ARC was equivalent to solving AGI. The first ARC solver is not going to be an AGI. But he basically said that, you know, this ARK challenge that he created, and I did talk about this before, but I just wanted to quickly gloss over this. But he says here that until we solve Ark, we don't have AGI since the AIs we have cannot adapt to simple tasks that they haven't seen before. And solving ARK will require figuring out how to make AI systems adapt on the fly to novel tasks. And this should be a major milestone on the way to AGI. Which is why I said that this is a major milestone, because solving the arch benchmark is going to be a major milestone, because whatever approach that you do use to solve that benchmark, it means that whatever reasoning engine that you're using, whether it be neurosymbolic, whether it be tree search, or whatever kind of approach it is that you do use, whatever approach that is going to be, it's going to be something that's remarkably effective if it can actually focus and solve this benchmark. And he says here the purpose of ARC is to get researchers to refocus on intelligence and pure away from memorization, because I believe this is how we will get to AGI. And basically you didn't know LLMs, they don't really have intelligence in the sense that they figure things out they do well on many tasks because they've been trained on various pieces of data. And there's a pure difference between, you know, humans that can see an image, like, of two cats, and then immediately they can, you know, recognize what a cat is outside in the wild. So this is the kind of reasoning where you're able to figure out what's going on on the fly in new and novel scenarios. Now, with that benchmark, though, the reason I actually brought this up was because someone basically decided to use LLMs and a Neuro symbol approach and basically managed to figure out how to do this. Now, it's crazy because, like I said before, the meme kind of memes neurosymbolic AI, but essentially the method that the person used, which was Ryan, to manage to get, you know, 72% with GPT4O, he actually used a neurosymbolic approach. So the meme here is kind of wrong. But essentially this was pretty crazy because it was something that many people thought would take quite a while. And you can see here, Francis Soleil said this has been the most promising branch of approaches so far, leveraging an LLM to help with discrete program search by using the LLM as a way to sample programs or branching decisions. This is exactly what neurosymbolic AI is, for the record. And that's why when you have AIs that can search over multiple things, it's something that results in a much more comprehensive system. Now, I wonder how effective this is going to be, because when you have an AI system like AlphaGo that can search over millions and millions of different positions, we can manage to filter out the bad decisions and then, of course, we can get to the real results. And although some people would argue that, oh, but that's not real intelligence. I mean, if it manages to get the result, then it doesn't really matter how it gets there, it just matters that it is there. And what's crazy about all of this is that, you know, Demis Hassabis tweeted about this. He says, we've long pioneered the use of these types of neurosymbolic Systems, starting with AlphaGo in 2016 through to AlphaZero. And we'll be bringing all the goodness of Alpha Proof and Alpha geometry 2 to our mainstream Gemini models very soon. Watch this space. So it means that Gemini models are potentially about to get really, really smart. Now, this whole thing was a little bit scary for some individuals. And I mean scary in the sense that this is something that some people did predict would initially Shrink the timeline. So if you're not familiar with what I'm talking about, you can see here that Paul Cristiano, the person who invented RLHF, now he basically said here that he's going to update his timeline if an AI got gold in any international Math Olympiad by the end of 20, and today, alpha proof came within one point of achieving that gold medal. You can see here his statements. He says that I think the IMO challenge would be significant direct evidence that powerful AI would be sooner or at least would be technologically possible sooner. I think this would be fairly significant evidence, Perhaps pushing my 2040 probability up from 25% to 40% or something like that. And I think that this would be significant evidence that the takeoff would will be limited by sociological facts and engineering effort rather than a slow march of smooth machine learning scaling. Maybe I'd move from 30% to a 50% chance of a hard takeoff. So basically what we have here is a situation where timelines are shrinking as we move throughout the year. And this is something that many people didn't even think would happen, considering how difficult these problems are. Now, Eliza Jukowski actually made a statement here that said, Paul Cristiano and I previously worked hard to pin down concrete disagreements. And one of our headers was that Paul put up an 8% probability on AI built before 2025, IMO reaches gold level on it. And I put it at least 16%. Now, the reason that this is so crazy is because Elijah Yudkowski has been someone who's basically stating that, look, creating super intelligent AI is just a stupid thing to do. Simply put, it's just stupid. It's just downright stupid. Because what's going to happen is it's going to do something that is going to cause irreversible damage or, or potentially human extinction. And what's crazy about all of this is that if you've seen and looked at some of the conversations that Eliza Yudkowski has, some of the arguments he does make are quite fascinating because he basically describes how a super intelligent AI is one that you simply cannot win against. If you think about it like this, for example, some people frequently say that, okay, tell me how the super intelligent AI is going to win and then we're going to solve that issue. But it's like this, okay? If you have someone who is the best chess player in the world, like Magnus Carlsen or Gary Kasparov, you can say that, look, if you put an average person against them, you know, 100% of the time, they are going to fail. But what we don't know is how they're going to fail. I can't tell you where they're going to place the pieces on the board. But what we do know is the end result. And the same situation is with AI. We don't know what this AI is going to do if it is a super intelligent system. What we do know is that the end result is that humans lose. Because we look at evolution, Anytime a new species comes on that is far more intelligent, the other species aren't around for much longer, or are essentially kept as entertainment, or just pretty much farmed for whatever resources they have. But just paint one or two possibilities.

Okay, so why is this hard? First, because you can't predict exactly where a smarter chess program will move. Maybe even more importantly than that, imagine sending the design for an air conditioner back to the 11th century. Even if they if it's enough detail for them to build it, they will be surprised when cold air comes out because the air conditioner will use the temperature pressure relation and they don't know about that law of nature. So if you want me to sketch what a superintelligence might do, I can go deeper and deeper into places where we think there are predictable technological advancements that we haven't figured out yet. And as I go deeper and deeper, it'll get harder and harder to follow. It could be super persuasive. That's relatively easy to understand. We do not understand exactly how the brain works. So it's a great place to exploit laws of nature that we do not know about rules of the environment, invent new technologies. Beyond that, can you build a synthetic virus that gives humans a cold and then bit of neurological change and they're easier to persuade? Can you build your own synthetic biology, synthetic cyborgs? Can you blow straight past that to covalently bonded equivalents of biology where instead of proteins that fold up and are held together by static cling, you've got things that go down much sharper potential energy gradients and are bonded together. People have done advanced design work.

So the point here, and I might even include a clip from Eliza Yudkowski, but the point here is that, you know, the meta point that Paul and I was arguing was, is the progress of AI smooth and therefore predictable and boundable? And the sharp prediction market movement on our bet suggests that this development was not smoothly predictable based on public information. Meaning that the kinds of research that's going on right now in Frontier Labs shows us that the kinds of systems that are currently being developed could far surpass our current estimates of what is even capable on the upper bounds of current intelligence. Which means that, look, capable systems are not far away and they're probably closer than we think, considering these incredible jumps in terms of prediction. Now, what was even crazier about this entire thing was that someone tweeted OpenAI has the opportunity to do the funniest thing. And then Sam Altman responded stating, lol, just stating, laughing out loud. Now the reason this is crazy is because not only did Sam Altman respond to this tweet saying lol, there has actually been some recent information regarding OpenAI's secret model. And if you remember what the secret model was actually focusing on, it was actually focusing on math and Sam Altman saying, lol, this basically is an indication that they perhaps might even be far ahead of where other systems currently are, which could mean that incredible systems are already here. And I don't think that Sam Altman would have responded to that tweet if that wasn't the case. Now for those of you that think, okay, this is just reaching, all he did was just put lol, you have to remember that the recent Q Star project or the Strawberry project you can see here from Reuters, which is a very reliable source, said that two sources described viewing earlier this year what OpenAI staffers told them were QStar demos capable of answering, remember, tricky science and math questions out of reach of today's commercially available models. And a different source briefed on the matter said that OpenAI has tested AI internally that scored over 90% on a math Datacent, a benchmark of championship math problems. And Reuters could not determine if this was the Strawberry project. So that's the reason that this is so crazy is because, you know, on one hand you've got OpenAI basically stating that, lol. We do have the ability to do something really funny, but at the same time we don't know what's going on with Strawberry because this was a model that is highly guarded and highly secret. But if the AI did score over 90% on a math data set, then that is pretty impressive because not only did OpenAI do it, but if you do remember, Google actually did release, not release, but release a paper which actually spoke about their specialized model called gemini math specialized 1.5 pro, and we can see that it scores 91.1%. And of course there is some RM at 256. I actually haven't read the paper, so I don't know what that means, but it's probably some kind of method that allows the AI system to get better results. So I think what's going on here is that we do have this inflection point on our hands where truly capable systems are just around the corner in terms of math and science. And I think that these kind of breakthroughs from Google and from OpenAI, managing to score over 90% on these benchmarks is going to be what actually drives the rest of technological progress, because this is the kind of, you know, research that can actually lead to new knowledge and, of course, AGI. So this was something that I wanted to talk about. Less about, you know, the information, just regarding the fact that, you know, it was able to get silver, but more around the fact that, you know, if we look at how timelines have now shrunk and. And the fact that OpenAI are hinting towards their systems being even more capable, this is something that is truly fascinating. So, that being said, if you did enjoy this video, let me know what your thoughts are on AI creating new knowledge and what you think about the International Math Olympiad and what you think about, you know, the varied experts on opinions on how we're going to get to AGI. There are many different things floating around, but of course, neurosymbolic AI so far does look to be very promising. With that being said, if you enjoyed the video, hopefully you guys have a wonderful day and I'll see you in the next AI update.