Ex-OpenAI Employee Just Revealed it ALL!

So Leopold Aschenbrenner is someone who used to work at OpenAI until he was, quote, unquote, fired for leaking internal documents. Now, I do want to state that this video is arguably one of the most important videos because he details the decade ahead for how many of the companies around the world are going to get to AGI. And his insights are like no other. He made this tweet stating that virtually nobody is pricing in what's coming in AI. And he made an entire document about the stages that we will need to take in order to get to AGI and some of the things that you're going to witness in the coming years. I think you should at least watch the first 10 minutes of this video because it is remarkably insightful into some of the things that he is predicting. So there are seven different sections, and I've read this thing from top to bottom at least through three times. And I'm going to give you guys the most insightful sections from this entire essay because I do believe that this is a remarkable document that I think everyone needs to pay attention to. So without wasting any more time, let's get into situational awareness, the decade ahead. So he has this introduction here where he talks about how the talk of the town has shifted from 10 billion complete clusters to $100 billion compute clusters of to even trillion dollar clusters. And every six months, another zero is added to the boardroom plans. The AGI race has begun. We are building machines that can think and reason. And by 2025 to 2026, these machines will outpace college graduates. And by the end of the decade, they will be smarter than you or I. And we will have superintelligence in the true sense of the word. I'm going to say that again. By the end of the decade, okay, we will have superintelligence in the truest sense of the word, along with the way national security forces not seen in half a century will be unleashed. And before long, the project will be on. These are some very fascinating predictions, but just trust me, once we get into some of the charts and some of the data that he's been analyzing, I think it really does make sense. And this is why this document is called Situational Awareness. Just read this part before we get into everything he says. Before long, the world will wake up. But right now, there are perhaps a few hundred people, most of them in San Francisco and the AI labs that actually have situational awareness. Through whatever peculiar forces or fate I have found myself amongst them. And this is why this document is really important because information like this, we're really lucky that people could leave a company like OpenAI and then publish a piece of information which gives us the details on how superintelligence is likely to arise and and when that system is likely to arise. So this is the section one from GPT4 to AGI counting the orders of magnitudes. So when you see ooms, that's what it stands for. So he clearly states here his AGI prediction, AGI by 2027 is strikingly plausible. GPT2 to GPT4 took us from preschooler to smart hall high schooler abilities in just four, four years. And if we trace the trend lines of compute algorithmic efficiencies and unhobbling of gains, we should expect another preschooler to high schooler size qualitative jump by 2027. Now this is where we get into our first very important chart because this shows us exactly where things may go. He says, I make the following claim. It is strikingly plausible that by 2027 models will be able to do the work of an AI researcher engineer. That doesn't require believing in sci fi, it just requires believing in straight lines on a graph. What we can see here is a graph of the base scale up of effective compute, counting GPT2 all the way up to GPT4 and looking at the effective compute that we're going to continue to scale up. Now one thing that is fascinating from here is, is that I think there is going to be an even steeper curve for this. The reason I state that is because during the period from 2022 to 2023 there was something that I would like to call an awareness. Okay, this period right here marks the birth of GPT3 to GPT4. And this put a real giant spectacle on the AI era. GPT3 and GPT4 weren't just research products. I mean GPT3 was, but GPT4 and ChatGPT3 were actual products that were available for the public. And since then we've seen an explosion in terms of how many people are now intrigued by AI and how many different companies are now piling billions of dollars and billions of resources into the technology, into the compute clusters, just so that they can capture all of the economic value that's going to be happening during this area. Which is why I do believe that it wouldn't be surprising if during the period from 2024 to 2028 we do have a lot more growth than we've had in this period. Which means that having an automated AI research engineer by 2027 to 2028 is not something that is far, far off. Because if we're just looking at the straight lines and the effect of compute, then this is definitely where we could get to. And the implications of this are quite stark because if we can have an automated AI research engineer, that means that it wouldn't take that long to get to superintelligence after that. Because if we can automate AI research, then all better off, we're able to effectively recursively self improve, but just without that crazy loop that makes superintelligence explode. Now here's where he states one of the things that I think is really, really important to understand. I stated this in a video before this document was released, but it's glad to see that someone else is ushering one of the same concerns that I originally thought of. He stated that the next generation of models has been in the oven, leading some to proclaim that stagnation and that deep learning is hitting a wall. But by counting the orders of magnitude, we get a peek at what we should actually expect. In a video around three weeks ago, I clearly stated that look, things are slowing down externally, but things are not slowing down internally at all. Just because some of the top AI labs may not have presented their most recent research is that doesn't mean that breakthroughs aren't being made every single month. He states here that while the inference is simple, the implication is striking. Another jump like that very well could take us to AGI, to models as smart as PhDs or experts that can work beside us as a co worker. Perhaps most importantly, if these AI systems could automate AI research itself, that would set intense feedback loops. And that's of course where we get AI researchers to make breakthroughs in AI research. And then we apply those breakthroughs to the AI systems, they become smarter and then the loop continues from there. Basically recursive self improvement, but on a slower scale. And here he clearly states, even now, barely anyone is pricing this in. But the situational awareness on AI isn't actually that hard once you step back and look at the trends. If you keep being surprised by AI capabilities, just start counting the orders of magnitudes. So here, here's where we talk about the last four years. So you can see he speaks about GPT2 to GPT4. GPT2 was essentially like a preschooler. While it can string together a few plausible sentences. And These are the GPT2 examples people found very impressive at the time, but yet it could barely count to five without getting tripped up. Then of course, we had GPT3, which was in 2020. And this was as smart as an elementary schooler. And this was something that once again impressed people quite a lot. Then of course, this is where we get to GPT4 in 2023. And this is where we get a smart high schooler. While it can write some pretty sophisticated code and iteratively debug, it can write intelligently and sophisticatedly about complicated subjects, it can reason through difficult high school competition math, and it's beating the vast majority of high schoolers on whatever tests we give it. And remember, there was the Sparks of AGI paper which showed some capabilities that showed us that we weren't too far away from AGI and that this GPT4 level system were the first initial sparks of artificial general intelligence. The thing is here he clearly states, and I'm glad he's stating this because a lot of people don't realize this, that the limitation comes down to obvious ways that models are still hobbled. And basically he's talking about the way that models are used and the current frameworks that they have, the raw intelligence behind the model, the raw cognitive capabilities of these models, if you even want to call it that, as artificially constrained. And basically, in the future, if you calculate the fact that these are going to be unconstrained in the future, it's going to be very fascinating on how that raw intelligence applies across different applications. And one of the clear things I think that most people aren't realizing is that we're running out of benchmarks. As an anecdote, my friends Dan and Colin made a benchmark called the MMLU a few years ago. In 2020, they hoped to finally make a benchmark that would stand the test of time equivalent to all the hardest exams we give high school and college students. Just three years later, models like GPT4 and Gemini get around 90%. And then of course, GPT4 mostly cracks all the standard high school and college aptitude tests. And you can see here the test scores of AI systems on various capabilities relative to human performance. You can see that in the recent years there have been a stark, stark level of increases here. It's absolutely crazy with as to how many different areas that AI is increasing in terms of the capabilities. It's really, really fascinating to see and also potentially quite concerning. Now, one of the things that most people did actually miss about going from GPT4 to AGI was a benchmark that actually did shock me. So there is essentially this benchmark called the math benchmark. A set of difficult mathematic problems from a high school math competitions. And when the benchmark was released in 2021, GPT3 only got 5%. And basically the crazy thing about this was that researchers predicted at that time stating to have more traction on mathematical problem solving, who will likely need new algorithmic advancements from the broader research community. And we're going to need fundamental new breakthroughs to solve maths. Or as they thought, they predicted minimal progress over the coming years. But by mid 2022 we got to 50% accuracy. And basically now with the recent math Gemini 1.5 Pro, we know that this is now at 90%, which is absolutely incredible. And here's something that you can clearly screenshot and share to your friends or colleagues or whatever it is, whatever kind of community that you might be in. But you can see that the performance on the common exams percentile compared to human test takers, we can see that GPT4 ranks above 90% for pretty much all of them, except calculus and chemistry, which is a remarkable feat when we went from GPT3 to GPT4 and in such a short amount of time. This is a true, true jump in capabilities that many people just simply wouldn't have expected. Now here's where we start to get to some of the predictions that we can really start to make based on the nature of deep learning. So essentially the magic of deep learning is that it just works. And the trend lines have been astonishingly consistent, despite the naysayers at every turn. We can see here that these are screenshots from the scaling compute in the OpenAI's Sora technology. And at each level we can see an increase in the quality and consistency. The base compute results in a pretty terrible image video. 4 times compute results in something that is pretty coherent and consistent. But 30 times compute is something that is remarkable in terms of the quality, consistency and the level of video that we do get, which shows us that these trend lines are very, very consistent. And he says if we can reliably count the orders of magnitude that we're going to be training these models, we, we can therefore extrapolate the capability improvements. And that's how some people actually saw the GPT4 level of capabilities coming. And one of the things that he talks about is of course things like chain of thought tools and scaffolding. And therefore we can unlock significant latent capabilities basically when we have GPT4 or whatever the base cognitive capabilities are for this architecture, and then we can use that to unlock latent capabilities by adding different steps in front of that system. So, for example, when you use GPT4 with chain of thought reasoning, you significantly improve your ability to answer certain questions in different scenarios. And it's things like that where you can unlock more knowledge from the system by using different ways to interact with it. Which means that the raw data behind the system and the raw knowledge is a lot bigger than people think. So this is what you call unhobbling gains. Now, one of the things that's really important, and this is something that doesn't get enough attention, but this is going to make up a lot of the gains that you won't see, and this is the algorithmic efficiencies. So whilst massive investments into compute get all the attention, algorithmic progress is similarly an important driver of progress and is dramatically underrated. To see just how big of a deal algorithmic progress can be, consider the following illustration. This one right here for drop of the price to attain 50% accuracy on the math benchmark over just two years. And for comparison, a computer science PhD student who didn't particularly like math scored 40%. So this is already quite good. And the inference efficiency improved by nearly three orders of magnitude or 1000x in less than two years. So what we have here is something that is incredibly more efficient for the same result in just two years. That is absolutely incredible. These algorithmic efficiencies are going to drive a lot more gains than you think. And as someone who was looking at arcsive, which is where a lot of these research papers get published, just trust me, there are like probably 50 to 80 different research papers that get published every single day. And a few of those allow, you know, 10 to 20% gain, 30% gain. And if you calculate the fact that all of these algorithmic efficiencies are going to compound against each other, which we're really going to see more cases like this. Here's where you talk about the API cost and you basically look at how efficient it becomes to run these models. So GPT4 on release cost the same at GPT3 when it was released. But since the GPT4 release a year ago, the prices for GPT4 level models have fallen 6 times, 4 times for the input output with the release of GPT4O and GPT 3.75 level is basically Gemini 1.5 flash and this is 85 times cheaper than than what we previously used to have. So we can see here on this graph that if we want to calculate exactly how much progress we're going to be make, we can clearly See that there are two main things here, which is of course, the physical compute of scaling, which is going to be things like these data centers and the hardware that we throw at the problems. And then of course, the algorithmic progress, which is going to be the efficiencies, where people rewrite these algorithms in crazy ways that just drive efficiencies that we previously didn't know how to solve. And that's why in the future, where we do get an automated AI researcher to do that, this gap is going to widen even more. This is where we talk about unhobbling. This is of course something that we just spoke about before. But the reason that this is important is because this is where you can get gains from a model in ways that you couldn't previously see before. So imagine if when someone asked you a math problem, you had to instantly answer with the first thing that came to mind. It seems pretty obvious that you would have a very hard time and except for the simplest problems. But until recently, that's how we had LLM solve math problems instead. Most of us, when we do math problems, we work the problem step by step and are able to solve much more difficult problems that way. It's basically chain of thought, and that's what we do for LLMs. And despite the excellent RAW capabilities, they were much worse at math than they could be because they were hobbled in an obvious way. And it was a small algorithmic tweak that unlocked much greater capabilities. Essentially what he's stating here is that when these even better models get even more unhobbled, we're going to see even more compounded gains overall. And one of the craziest ones that we recently have is, of course GPT4 can only solve the software engineering bench 2% correctly, while with Devon's agent scaffolding, it jumps to 14%, which is pretty, pretty incredible. And this is something that is very, very small in terms of its infancies. And it says tools. Imagine if humans weren't allowed to use calculators or computers. We're literally only at the beginning here. ChatGPT can only now use a web browser, run some code, and so on. And of course, this is where we talk about the context length, which is, you know, it's gone from a 2k context length to 32 to literally a 1 million context length. And of course there's post training, which is substantially improving the models after you've trained the model, which is making huge gains. We went from 50% to 72% on math and 40% to 50% on the GPQA. And here's where we can see once again there is another stack in terms of the growth. So you can see here the raw chatbot from the chatbot to agent. These are things that most people just aren't factoring in. When we take a look at the future of AI growth, this is why, and this is why he says the improvements will be step changes compared to GPT6 and reinforcement learning with human feedback. By 2027, rather than a chatbot, you're going to have something that looks more like an agent and more like a co worker. Now one of the craziest things I saw here was that when you take in all of the information that was just stated, this is absolutely incredible because he basically says that by the end of 2027 this is absolutely insane. So we can see the gains made from GPT2 to GPT4 by physical compute algorithmic efficiencies, plus major unhobbling gains from the base model chatbot and the subsequent four years, we're going to see three to six orders of magnitude of base effective compute scale up, which is the physical compute and algorithmic efficiencies. But basically he says that with all of this combined, what this should look like. Suppose GPT4 training took three months in 2027, a leading AI lab will be able to train a GPT4 level model in a minute. That is a incredible prediction and I'm wondering if that is going to be true. I But then again, you have to think about it. In three years, with billions of dollars and that much more compute floating around the industry, I wouldn't be surprised if some of the things that we think right now are sci fi completely aren't. So here's where we can see everything visualized. We can literally see how we have the base knowledge, then we've got the chatbot framework, then once we have the agentic framework, and of course once we have the more orders of magnitude, we can see that the intelligence by 2027, become an automated AI research engineer. Now that you actually look at all the information from every different single point, it doesn't seem that crazy. Like if you count everything, every single thing into account, this doesn't seem like something that is too, too far away. Especially with the kind of jumps that we've seen before. And maybe, just maybe with the GPT5 release and subsequent AI models, we're going to start to see that 2027 and 2026 are going to be incredible time periods. He says we are on course for AGI by 2027 and that these AI systems will basically be able to automate basically all cognitive jobs, think any job that can be done remotely. That is a crazy, crazy statement. And I think that that is something that you need to, you know, bear in mind AGI by 2027 is not something that is out of the picture and it's something that definitely could happen. So one of the most interesting things as well that I think is really important is that the reason this period is so important is because this is the decisive period. This is the period where the growth occurs and we really get to see what is capable. So he says right here, in essence, we're in a middle of a huge scale up. And this is where we are reaping one time gains this decade and progress through the orders of magnitude will be multiples slower thereafter. If this scale up doesn't get us to AGI in the next five to 10 years, it might be a long way out. So the reason that this is going to be, you know, so interesting is because it's this decade or bust. So you can see right here that the effective scale up of compute is going to become harder and harder the larger it gets. Because think about it like this, it's like, well, I don't actually have a great example, but we just have to think about how hard it is to invest billions and billions of dollars more to scale up systems even more. So it's like of course you can scale up a model from 10 billion to 100 million, but the scale from 100 million to 10 billion is really huge. It takes a lot of investment. You're going to have to data centers, multiple data centers, you're going to have to make them really huge. You're going to have to think about all the cooling, there's a lot of power requirements. And then to get to a trillion dollar clusters or even $100 billion or even $500 billion clusters, that's even more incredible. So basically he's stating once we get to the $100 billion level and above, if we aren't at AGI at that level, then it means that realistically we're going to have to wait for some kind of algorithmic breakthrough or an entirely new architecture. Because with the gains that are going to be made by being able to have so much more compute being thrown at the problem, it is very hard for us to make gains based on compute after that. He basically says here that spending a million dollars on a model used to be outrageous. But by the end of the decade, we will likely have $100 billion or 1 trillion dollar clusters. And going much higher than that is going to be a lot harder. So it's going to be basically the feasible limit, both in terms of what businesses can actually afford and even just as a fraction of the gdp. And he also states that the large gains that we're getting from CPUs to GPUs will be likely gone by the end of the decade because we're going to have AI specific chips. And without much further beyond more zlo, there's not going to be much more gains possible. And the reason this is important, because for those of you who are trying to navigate this entire thing, you're trying to figure out, okay, where AI capability is going to stop, where is the next growth going to be at? It's basically the fact that right now we're scaling up our systems and once we reach the top of $100 billion to 100 trillion clusters, if we don't have superintelligence or AGI by that limit, that then we'll know that maybe we're using the wrong architecture and things are going to have to change significantly. So it's either going to be a long, slow slog or we're going to get there relatively soon. And by the looks of things, it looks like we're going to get there relatively soon. Now here's where we talk about AGI to superintelligence, the intelligence explosion. And basically this is where he talks about how AI progress will not stop a human level. Hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress, which adds five orders of magnitude into one year. And we would rapidly go from human level to vastly superhuman AI systems. And the power and the peril of super intelligence would be dramatic. So here's where we have basically, I think, the most important graph. Okay. If there's one that you want to screenshot and keep on your phone, I think it's this one. Okay. The reason that it is is because once you have the GPT, 4, GPT, 3 GPT two timelines mapped out, we can clearly see that this intersection here is at 2023. But of course, as the trends continue, we can see that once we do get to this period right here, this is where things start to get interesting. Because this is of course, the period of automated AI research. And that's why once this does happen, and this is not something that's like a fairy tale, this is something that Sam Altman has said that's his entire goal. That's what OpenAI are trying to build. They're not really trying to build super intelligence, but they define AGI as a system that can do automated AI research. And once that does occur, and I don't think it's going to take that long, that's when we're going to get that recursive self improvement loop where superintelligence is not going to take that long after. Because if you can deploy 5,000 agents, okay, that are essentially all super intelligent, not super intelligent, but at the level of a standard AI researcher, and we can deploy them on certain problems and keep them running 24, seven, that is going to just compress years of AI research into a very short time frame. Which is why you can see that the graph during this purple period here, it starts to go up rapidly. And that's why the next decade is so important. Because once this area actually happens, once we get to that breakthrough level where, okay, we've automated AI research, then all bets are off because we know that superintelligence will just be around the corner. And that's when we have the intelligent explosion. Because every time an AI researcher manages to make a breakthrough, the AI research breakthrough is then applied to that AI researcher and then the progress continues again. Because now the AI researcher is just that, more efficient or even smarter. And here's the crazy thing, this is one of the craziest implications about this entire thing. We don't need to automate everything, just AI research. I'm gonna say that again. We don't need to automate everything. It's just AI research. A common objection to transformative impacts of AGI is that it will be hard for AI to do everything. Look at robotics for instance. The doubters say that there will be a gnarly problem even if AI is cognitively at the level of PhDs, or take automating biology research and design, which might require lots of physical lab work and human experiments, but we don't actually need robotics. We don't need many things for AI to automate AI research. The jobs of AI researchers and engineers at leading labs can be done fully, virtually and don't run into real world bottlenecks the same way that robotics does. And of course this is going to still be limited by compute, which is addressed later. And basically that's things whereby, like the literal hardware issues that you get when you're trying to scale these systems like it's not that hard. Well, I say it's not that hard. But theoretically it should be easier to read ML literature and come up with new questions and ideas implement these experiments, test those ideas interpret the results and then of course repeat it and all it takes is for once we get to that level that's where we have this insane feedback loop and this is where 2027 we should expect GPU fleets in the tens of millions training clusters alone approaching three altitudes larger already putting us at 10 million A100 equivalents and this is going to be running millions of copies of our automated AI researchers perhaps 100 million human researcher equivalents running day and night that is absolutely incredible and of course some of the GPUs are going to be you know used for training new models but just think about that guys imagine 100 million human researcher equivalents running 247 what kind of breakthroughs are going to be made at that stage? I mean it's very hard to conceptualize but it's important to take into account what is truly coming because like he said nobody's really pricing this in.