--:-- Dwarkesh Patel: Human-level AI is deep, deep into an intelligence explosion. Things like inventing the transformer or discovering chinchilla scaling and doing your training runs more optimally or creating flash attention. That set of inputs probably would yield the kind of AI capabilities needed for intelligence explosion.
--:-- Carl Shulman: You have a race between, on the one hand, the project of getting strong interpretability and shaping motivations, and on the other hand, these AIs in ways that you don't perceive make the AI take over happen.
--:-- Dwarkesh Patel: We spend more compute by having a larger brain than other animals, and then we have a longer childhood. It's analogous to like having a bigger model and having more training time with it. It seemed very implausible that we couldn't do better than completely brute force evolution. How quickly are we running through those orders of magnitude?
--:-- [Intro music fades in and out]
--:-- Dwarkesh Patel: Hey everybody, just wanted to give you a heads up. So I ended up talking to Carl for like seven or eight hours. So we ended up splitting this episode into two parts. I don't want to put all of that on you at once. In this part, we get deep into Carl's model of an intelligence explosion and what that implies for alignment. The next part, which we'll release next week, is all about the specific mechanisms of an AI takeover. In terms of the depth and the range of interesting topics, this set of episodes is the best I've ever done. So I hope you all enjoy. Here's Carl.
--:-- Dwarkesh Patel: Okay, today I have the pleasure of speaking with Carl Shulman. Many of my former guests, and this is not an exaggeration, many of my former guests have told me that a lot of their biggest ideas, perhaps most of their biggest ideas, have come directly from Carl, especially when it has to do with the intelligence explosion and its impacts. And so I decided to go directly to the source, and we have Carl today on the podcast. Carl keeps a super low profile, but he is one of the most interesting intellectuals I've ever encountered. And this is actually his second podcast ever. So we're going to get deep into the heart of many of the most important ideas that are circulating right now directly from the source.
--:-- Dwarkesh Patel: So by the way, Carl is also an advisor to the Open Philanthropy Project, which is one of the biggest funders on causes having to do with AI and its risks, not to mention global health and well-being. And he is a research associate at the Future of Humanity Institute at Oxford. So Carl, it's a huge pleasure to have you on the podcast. Thanks for coming.
--:-- Carl Shulman: Thank you, Dwarkesh. I enjoyed seeing some of your episodes recently and I'm glad to be on the show.
--:-- Dwarkesh Patel: Excellent. Let's talk about AI. Before we get into the details, give me the sort of big picture explanation of the feedback loops and just the general dynamics that would start when you have something that is approaching human level intelligence.
--:-- Carl Shulman: Yeah. So I think the way to think about it is we have a process now where humans are developing new computer chips, new software, running larger training runs. And it takes a lot of work to keep Moore's law chugging. Well, it was. It's slowing down now. And it takes a lot of work to develop things like transformers, to develop a lot of the improvements to AI and neural networks that are advancing things.
--:-- Carl Shulman: And the core method that I think I want to highlight on this podcast and I think is underappreciated is the idea of input-output curves. So we can look at the increasing difficulty of improving chips. And so sure, each time you double the performance of computers it's harder and as we approach physical limits eventually it becomes impossible. But how much harder? There's a paper called "Are Ideas Getting Harder to Find?" that was published a few years ago. At Miri, we did an early version of this analysis using mainly data from Intel and the large semiconductor fabricators.
--:-- Carl Shulman: Anyway, in this paper, they cover a period where the productivity of computing went up a million fold. So you could get a million times the computing operations per second per dollar. Big change. But it got harder. So the amount of investment, the labor force required to make those continuing advancements went up and up and up. Indeed, it went up 18-fold over that period. So some take this to say, "Oh, diminishing returns, things are just getting harder and harder, and that will be the end of progress eventually."
--:-- Carl Shulman: However, in a world where AI is doing the work, that doubling of computing performance translates pretty directly to a doubling or better of the effective labor supply. That is, if when we had that million-fold compute increase, we used it to run artificial intelligences who would replace human scientists and engineers, then the 18x increase in the labor demands of the industry would be trivial. We're getting more than one doubling of the effective labor supply that we need for each doubling of the labor requirement. In that data set, it's like over four. So we double compute. Now we need somewhat more researchers, but a lot less than twice as many. So we use up some of those doublings of compute on the increasing difficulty of further research, but most of them are left to expedite the process.
--:-- Carl Shulman: So if you double your labor force, that's enough to get several doublings of compute. You use up one of them on meeting the increased demands from diminishing returns. The others can be used to accelerate the process. So your first doubling takes however many months. Your next doubling can take a smaller fraction of that, the next doubling less, and so on. At least insofar as the outputs you're generating, compute for AI in this story, are able to serve the function of the necessary inputs. If there are other inputs that you need, eventually those become a bottleneck and you wind up more restricted on this.
--:-- Dwarkesh Patel: Got it. Okay, so yeah, I think the Bloom paper had that there was a 35% increase in, was it transfer destiny or cost per flop? And there was a 7% increase per year in the number of researchers required to sustain that pace. Something like that vicinity, yeah. It's like four to five doublings of compute per doubling of labor inputs.
--:-- Dwarkesh Patel: I guess there's a lot of questions you can delve into in terms of whether you would expect a similar scale with AI and whether it makes sense to think of AI as a population of researchers that keeps growing with compute itself. Actually, let's go there. So can you explain the intuition that compute is a good proxy for the number of AI researchers, so to speak?
--:-- Carl Shulman: So far, I've talked about hardware as an initial example, because we had good data about a past period. You can also make improvements on the software side. And we think about an intelligence explosion that can include AIs doing work on making hardware better, making better software, making more hardware. But the basic idea for the hardware is especially simple in that if you have a worker, an AI worker that can substitute for a human, if you have twice as many computers, you can run two separate instances of them. And then they can do two different jobs, manage two different machines, work on two different design problems.
--:-- Carl Shulman: Now, you can get more gains than just what you would get by having two instances. We get improvements from using some of our compute, not just to run more instances of the existing AI, but to train larger AIs. So there's hardware technology, how much you can get per dollar you spend on hardware. And there's software technology. And the software can be copied freely. So if you've got the software, it doesn't necessarily make that much to say that, "Oh, we've got 100 Microsoft Windows." You can make as many copies as you need for whatever Microsoft will charge you. But for hardware, it's different. It matters how much we actually spend on the hardware at a given price. And if we look at the changes that have been driving AI recently, that is the thing that is really off-trend. We are spending tremendously more money on computer hardware for training big AI models.
--:-- Dwarkesh Patel: Yeah, okay. So there's the investment in hardware, there's the hardware technology itself, and there's the software progress itself. The AI is getting better because we're spending more money on it, because our hardware itself is getting better over time, and because we're developing better models or better adjustments to those models. Where is the loop here?
--:-- Carl Shulman: The work involved in designing new hardware and software is being done by people now. They use computer tools to assist them, but computer time is not the primary cost for NVIDIA designing chips, for TSMC producing them, for ASML making lithography equipment to serve the TSMC fabs. And even in AI software research, that has become quite compute-intensive. But I think we're still in the range where at a place like DeepMind, salaries were still larger than compute for the experiments. Although tremendously, tremendously more of the expenditures were on compute relative to salaries than in the past.
--:-- Carl Shulman: If you take all of the work that's being done by those humans, there's like low tens of thousands of people working at NVIDIA designing GPUs specialized for AI. I think there's more like 70,000 people at TSMC, which is the leading producer of cutting edge chips. There's a lot of additional people at companies like ASML that supply them with the tools they need. And then a company like DeepMind, I think from their public filings, they recently had a thousand people. OpenAI, I think, is a few hundred people. Anthropic is less. If you add up things like Facebook AI research, Google Brain, R&D, you get thousands or tens of thousands of people who are working on AI research.
--:-- Carl Shulman: We'd want to zoom in on those who are developing new methods rather than narrow applications. So inventing the transformer definitely counts. Optimizing for some particular business's dataset cleaning, probably not. But those people are doing this work. They're driving quite a lot of progress. What we observe in the growth of people relative to the growth of those capabilities is that pretty consistently the capabilities are doubling on a shorter time scale than the people required to do them are doubling. And so there's work.
--:-- Carl Shulman: So we talked about hardware and how historically it was pretty dramatic, like four or five doublings of compute efficiency per doubling of human inputs. I think that's a bit lower now as we get towards the end of Moore's law. Although interestingly, not as much lower as you might think because the growth of inputs has also slowed recently. On the software side, there's some work by Tamei Bessaroglu, and I think collaborators, may have been his thesis. It's called "Are Models Getting Harder to Find?" And so it's applying the same sort of analysis as the "Are Ideas Getting Harder to Find?" And you can look at growth rates of papers from citations, employment at these companies, and it seems like the doubling time of these workers driving the software advances is like several years or at least a couple of years, whereas the doubling of effective compute from algorithmic progress is faster.
--:-- Carl Shulman: There's a group called Epoch. They've received grants from Open Philanthropy and they do work collecting data sets that are relevant to forecasting AI progress. Their headline results for what's the rate of progress in hardware and software and growth in budgets are as follows. For hardware, they're looking at a doubling of hardware efficiency that's like two years. It's possible it's a bit better than that when you take into account certain specializations for AI workloads. For the growth of budgets, they find a doubling time that's something like six months in recent years, which is pretty tremendous relative to the historical rates. We should maybe get into that later. And then on the algorithmic progress side, mainly using ImageNet type datasets right now, they find a doubling time that's less than one year. And so you combine all of these things and the growth of effective compute for training big AIs is pretty drastic.
--:-- Dwarkesh Patel: I think I saw an estimate that GPT-4 costs like $50 million around that range to train. Now, suppose that AGI takes a thousand times that if you were to scale up GPT-4. It might not be that. I'm just for the sake of example. So part of that will come from companies just spending a lot more to train the models and greater investment. Part of that will come from them having better models so that what would have taken a 10x increase in the model to get naively, you can do with having a better model that you only need to scale up. You get the same effect of increasing it by 10x just from having a better model. And so you can spend more money on it to train a bigger model. You can just have a better model, or you can have chips that are cheaper to train. So you get more compute for the same dollars.
--:-- Dwarkesh Patel: Those are the three ways you're describing the effective compute would increase. From looking at it right now, it looks like you might get two or three doublings of effective compute for this thing that we're calling software progress, which people get by asking, "How much less compute can you use now to achieve the same benchmark as you achieved before?" There are reasons to not fully identify this with software progress, as you might naively think of it, because some of it can be enabled by the others. So when you have a lot of compute, you can do more experiments and find algorithms that work better. Sometimes the additional compute, you can get higher efficiency by running a bigger model we were talking about earlier. And that means you're getting more for each GPU that you have because you made this larger expenditure. And that can look like a software improvement because this model, it's not a hardware improvement directly because it's doing more with the same hardware. But you wouldn't have been able to achieve it without having a ton of GPUs to do the big training run.
--:-- Dwarkesh Patel: The feedback loop itself involves the AI that is the result of this greater effective compute helping you train better AI, right? Or use less effective computing in the future to train better AI?
--:-- Carl Shulman: It can help on the hardware design. So NVIDIA is a fabulous chip design company. They don't make their own chips. They send files of instructions to TSMC, which then fabricates the chips in their own facilities. The work of those 10,000 plus people, if you could automate that and have the equivalent of a million people doing that work, then I think you would pretty quickly get the kind of improvements that can be achieved with the existing nodes that TSMC is operating on. You could get a lot of those chip design gains, basically doing the job of improving chip design that those people are working on now, but get it done faster. So that's one thing.
--:-- Carl Shulman: I think that's less important for the intelligence explosion. The reason being that when you make an improvement to chip design, it only applies to the chips you make after that. If you make an improvement in AI software, it has the potential to be immediately applied to all of the GPUs that you already have. And so the thing that I think is most disruptive and most important at the leading edge of the change from AI automation of the inputs to AI is on the software side.
--:-- Dwarkesh Patel: At what point would it get to the point where the AIs are helping develop better software or better models for future AIs? Some people claim today, for example, that programmers at OpenAI are using Copilot to write programs now. So in some sense, you're already having that sort of feedback loop. I'm a little skeptical of that as a mechanism. At what point would it be the case that the AI is contributing significantly in the sense that it would almost be the equivalent of having additional researchers to AI progress in software?
--:-- Carl Shulman: The quantitative magnitude of the help is absolutely central. So there are plenty of companies that make some product that very slightly boosts productivity. So when Xerox makes fax machines, it maybe increases people's productivity in office work by 0.1% or something. You're not going to have explosive growth out of that because, okay, now 0.1% more effective R&D at Xerox than any customers buying the machines. Not that important.
--:-- Carl Shulman: So I think the thing to look for is when is it the case that the contributions from AI are starting to become as large or larger as the contributions from humans? So when this is boosting their effective productivity by 50 or 100%, and you then go from, say, eight months doubling time for effective compute from software innovations, things like inventing the transformer or discovering chinchilla scaling and doing your training runs more optimally or creating flash attention. If you move that from say eight months to four months and then the next time you apply that, it significantly increases the boost you're getting from the AI. So now maybe instead of giving a 50% or 100% productivity boost, now it's more like a 200%.
--:-- Carl Shulman: It doesn't have to have been able to automate everything involved in the process of AI research. It can be that it's automated a bunch of things. And then those are being done in extreme profusion because AI can do it much more often since it's so cheap. It's not a threshold of "this is human level AI, it can do everything a human can do with no weaknesses in any area." It's that even with its weaknesses, it's able to bump up the performance so that instead of getting the results we would have with the 10,000 people working on finding these innovations, we get the results that we would have if we had twice as many of those people with the same kind of skill distribution. That's a demanding challenge. You need quite a lot of capability for that, but it's also important that it's significantly less than "this is a system where there's no way you can point at it and say in any respect it is weaker than a human." A system that was just as good as a human in every respect, but also had all the advantages of an AI, that is just way beyond this point.
--:-- Carl Shulman: If you consider that the output of our existing fabs make tens of millions of advanced GPUs per year. Those GPUs, if they were running AI software that was as efficient as humans, is sample efficient, it doesn't have any major weaknesses. So they can work four times as long. The 168 hour work week. They can have much more education than any human. A human, you got a PhD, it's like 20 years of education, maybe longer if they take a slow route on the PhD. It's just normal for us to train large models by eating the internet, eating all the published books, ever reading everything on GitHub, and getting good at predicting it. So the level of education vastly beyond any human, the degree to which the models are focused on task is higher than all but the most motivated humans when they're really, really gunning for it.
--:-- Carl Shulman: So you combine the things: tens of millions of GPUs, each GPU is doing the work of the very best humans in the world. The most capable humans in the world can command salaries that are a lot higher than the average, particularly in a field like STEM or narrowly AI. There's no human in the world who has a thousand years of experience with TensorFlow, or let alone the new AI technologies that were invented the year before. But if they were around, they'd be paid millions of dollars a year. And so when you consider this, tens of millions of GPUs, each is doing the work of maybe 40, maybe more of these kind of existing workers. It's like going from a workforce of tens of thousands to hundreds of millions. You immediately make all kinds of discoveries then. You immediately develop all sorts of tremendous technologies. So human-level AI is deep, deep into an intelligence explosion. The intelligence explosion has to start with something weaker than that.
--:-- Dwarkesh Patel: Yeah, what is the thing it starts with and how close are we to that? Because if you think of a researcher at OpenAI or something, to be a researcher is not just completing the hello world prompt that Copilot does. It's like you had to choose a new idea. You had to figure out the right way to approach it. You perhaps have to manage the people who are also working with you on that problem. It's an incredibly complicated skill portfolio, rather than just a single skill. So what is the point at which that feedback loop starts where you can even, you're not just doing the 0.5% increase in productivity that a sort of AI tool might do, but is actually the equivalent of a researcher or close to it? Like, what is that point?
--:-- Carl Shulman: So I think maybe a way to look at it is to give some illustrative examples of the kinds of capabilities that you might see. Because these systems have to be a lot weaker than the sort of human level things, what we have is intense application of the ways in which AIs have advantages, partly offsetting their weaknesses. And so AIs are cheap. We can call a lot of them to do many small problems. And so you'll have situations where you have dumber AIs that are deployed thousands of times to equal, say, one human worker. And they'll be doing things like these voting algorithms where you, with an LLM, you generate a bunch of different responses and take a majority vote among them that improves performance some. You'll have things like the AlphaGo kind of approach, where you use the neural net to do search. And you go deeper with the search by plowing in more compute, which helps to offset the inefficiency and weaknesses of the model on its own. You'll do things that would just be totally impractical for humans because of the sheer number of steps.
--:-- Carl Shulman: An example of that would be designing synthetic training data. Humans do not learn by just going into the library and opening books at random pages. It's actually much, much more efficient to have things like schools and classes where they teach you things in an order that makes sense, that's focusing on the skills that are more valuable to learn. They give you tests and exams that are designed to try and elicit the skill they're actually trying to teach. And right now we don't bother with that because we can hoover up more data from the internet. We're getting towards the end of that. But as the AIs get more sophisticated, they'll be better able to tell what is a useful kind of skill to practice and to generate that.
--:-- Carl Shulman: We've done that in other areas. So AlphaGo, the original version of AlphaGo was booted up with data from human Go play and then improved with reinforcement learning and Monte Carlo tree search. But then AlphaZero, with a somewhat more sophisticated model, benefited from some other improvements, but was able to go from scratch. And it generated its own data through self-play. Getting data of a higher quality than the human data because there are no human players that good available in the data set. And also a curriculum so that at any given point, it was playing games against an opponent of equal skill itself. And so it was always in an area when it was easy to learn. If you're just always losing no matter what you do, or always winning no matter what you do, it's hard to distinguish which things are better and which are worse.
--:-- Carl Shulman: When we have somewhat more sophisticated AIs that can generate training data and tasks for themselves, for example, if the AI can generate a lot of unit tests and then can try and produce programs that pass those unit tests, then the interpreter is providing a training signal. And the AI can get good at figuring out what's the kind of programming problem that is hard for AIs right now that will develop more of the skills that I need and then do them. Now you're not going to have employees at OpenAI write a billion programming problems. That's just not going to happen. But you are going to have AIs given the task of producing those enormous number of programming challenges.
--:-- Carl Shulman: In LLMs themselves, there's a paper out of Anthropic called Constitution AI or Constitution RL, where they basically had the program just talk to itself and say, "Is this response helpful? If not, how can I make this more helpful?" And the response is improved. And then you train the model on the more helpful responses that it generates by talking to itself, so that it generates it natively. And you could imagine more sophisticated ways to do that, or better ways to do that.
--:-- Dwarkesh Patel: Okay, so but then the question is, listen, GPT-4 already costs like 50 million or 100 million or whatever it was. Even if we have greater effective compute from hardware increases and better models, it's hard to imagine how we could sustain four or five more orders of magnitude greater effective size than GPT-4 unless we're jumping in trillions of dollars, like the entire economies of big countries, into training the next version. So the question is: do we get something that can significantly help with AI progress before we run out of the sheer money and scale and compute that would require to train it? Do you have a take on that?
--:-- Carl Shulman: Well, first I'd say, remember that there are these three contributing trends. So the new H100s are significantly better than the A100s. And a lot of companies are actually just waiting for their deliveries of H100s to do even bigger training runs, along with the work of hooking them up into clusters and engineering the thing. So all of those factors are contributing. And of course, mathematically, if you do four orders of magnitude more than 50 or 100 million, then you're getting to trillion dollar territory. And I think the way to look at it is at each step along the way, does it look like it makes sense to do the next step?
--:-- Carl Shulman: And so from where we are right now, seeing the results with GPT-4 and ChatGPT, companies like Google and Microsoft and whatnot are pretty convinced that this is very valuable. You have talk at Google and Microsoft with Bing that, well, it's like a billion-dollar matter to change market share in search by a percentage point. And so that can fund a lot.
--:-- And on the far end, on the extreme, if you automate human labor, we have a $100 trillion economy. Most of that economy is paid out in wages. So between $50 and $70 trillion per year. If you create AGI, it's going to automate all of that and keep increasing beyond that. So the value of the completed project is very much worth throwing our whole economy into it if you're going to get the good version, not the catastrophic destruction of the human race or some other disastrous outcome.
--:-- And in between, it's a question of, well, did the next step, how risky and uncertain is it, and how much growth in the revenue you can generate with it do you get? For moving up to a billion dollars, I think that's absolutely going to happen. These large tech companies have R&D budgets of tens of billions of dollars. When you think about it, all the employees at Microsoft who are doing software engineering that contributes to creating software objects—it's not weird to spend tens of billions of dollars on a product that would do so much. And I think it's becoming more clear that there is market opportunity to fund the thing.
--:-- Going up to $100 billion, that's like, okay, the existing R&D budgets spread over multiple years. But if you keep seeing that when you scale up the model, it substantially improves the performance, it opens up new applications. You're not just improving your search, but maybe it makes self-driving cars work. You replace bulk software engineering jobs, or if not replace them, amplify productivity. In this kind of dynamic, you actually probably want to employ all the software engineers you can get as long as they're able to make any contribution, because the returns of improving stuff in AI itself get so high.
--:-- But yeah, so I think that can go up to $100 billion. And at $100 billion, you're using a significant fraction of our existing fab capacity. Right now, the revenue of NVIDIA is like $25 billion. The revenue of TSMC, I believe it's like over $50 billion. Last I checked in 2021, NVIDIA was maybe seven and a half percent, less than 10% of TSMC revenue. So there's a lot of room. And most of that was not AI chips. They have a large gaming segment. There are data center GPUs that are used for video and the like. So there's room for more than an order of magnitude increase by redirecting existing fabs to produce more AI chips and just actually using the AI chips that these companies have in their cloud for the big training runs. And so I think that's enough to go to the $10 billion, and then combined with stuff like the H100s, go up to $100 billion.
--:-- Just to emphasize for the audience, the initial point about revenue made: if it cost OpenAI $100 million to train GPT-4 and it generates $500 million in revenue, you pay back your expenses with $100 million, you have $400 million for your next training run. Then you train your GPT-4, you get, let's say, $4 billion out of revenue out of that. That's the feedback loop of sort of revenue coming from where you're automating tasks, and therefore you're making money, you can use that money to automate more tasks, on top of the ability to redirect the fab production towards AI chips.
--:-- So the TL;DR on: you want $100 billion worth of compute? I mean, fabs take, what, like a decade or so to build. So given the ones we have now and the ones that are going to come online in the next decade, is there enough to sustain $100 billion of GPU compute if you wanted to spend that on a training run? Yes, you could definitely make the $100 billion one.
--:-- As you go up to a trillion-dollar run and larger, it's going to involve more fab construction. And yeah, fabs can take a long time to build. On the other hand, if in fact you're getting very high revenue from the AI systems and you're actually bottlenecked on the construction of these fabs, then their price could skyrocket. And that would lead to measures we've never seen before to expand and accelerate fab production.
--:-- If you consider, at the limit, as you're getting models that approach human-like capability, imagine things that are getting close to brain-like efficiencies plus AI advantages. We were talking before about a GPU that is supporting an AI. Really, it's a cluster of GPUs supporting AIs that do things in parallel, data parallelism. But if that can work four times as much as a human—a highly skilled, motivated, focused human with levels of education that have never been seen in the human population. And so if a typical software engineer can earn hundreds of thousands of dollars, the world's best software engineers can earn millions of dollars today and maybe more in a world where there's so much demand for AI. And then times four for working all the time.
--:-- Well, I mean, if you can generate close to $10 million a year out of the future version of H100, and it costs tens of thousands of dollars with a huge profit margin now, and profit margin could be reduced with large production—that is a big difference. That chip pays for itself almost instantly. And so you could support paying 10 times as much to have these fabs constructed more rapidly. You could have, if AI is starting to be able to contribute, you could have AI contributing more of the skilled technical work that makes it hard for, say, NVIDIA to suddenly find thousands upon thousands of top quality engineering hires, if AI can provide that.
--:-- Now, if AI hasn't reached that level of performance, then this is how you can have things stall out. And a world where AI progress stalls out is one where you go to the $100 billion and then over succeeding years, trillion-dollar things, software progress turns out to stall. You lose the gains that you were getting from moving researchers from other fields. Lots of physicists and people from other areas of computer science have been going to AI. But you sort of tap out those resources as AI becomes a larger proportion of the research field. And like, okay, you've put in all of these inputs, but they just haven't yielded AGI yet.
--:-- I think that set of inputs probably would yield the kind of AI capabilities needed for intelligence explosion. But if it doesn't, after we've exhausted this current scale-up of increasing the share of our economy that is trying to make AI, if that's not enough, then after that you have to wait for the slow grind of things like general economic growth, population growth, and such, and so things slow. And that results in my credences in this kind of advanced AI happening to be relatively concentrated over the next 10 years compared to the rest of the century. Because we can't keep going with this rapid redirection of resources into AI. That's a one-time thing. If the current scale-up works, it's going to happen. We're going to get to AI really fast, like within the next 10 years or something. If the current scale-up doesn't work, all we're left with is just our economy growing like 2% a year. So we have like 2% a year more resources to spend on AI. And at that scale, you're talking about decades before you can, just through sheer brute force, train the $10 trillion model or something.
--:-- Dwarkesh Patel: Let's talk about why you have your thesis that the current scale of would work. What is the evidence from AI itself, or maybe from primate evolution and the evolution of other animals? Just give me the whole confluence of reasons that make you think.
--:-- Carl Shulman: Maybe the best way to look at that might be to consider when I first became interested in this area. So in the 2000s, which was before the deep learning revolution, how would I think about timelines? How did I think about timelines? And then how have I updated based on what has been happening with deep learning?
--:-- Back then I would have said, we know the brain is a physical object and information processing device. It works. It's possible. And not only is it possible, it was created by evolution on earth and so that gives us something of an upper bound and that this kind of brute force was sufficient. There are some complexities with like, well, what if it was a freak accident and that didn't happen on all of the other planets? And that added some value. I have a paper with Nick Bostrom on this. I think basically that's not that important an issue. There's convergent evolution, like octopi are also quite sophisticated. If a special event was at the level of forming cells at all or forming brains at all, we get to skip that because we're choosing to build computers and we already exist. We have that advantage. So say evolution gives something of an upper bound, really intensive, massive brute force search and things like evolutionary algorithms can produce intelligence.
--:-- Dwarkesh Patel: Doesn't the fact that octopi and other mammals got to the point of being pretty intelligent, but not human-level intelligent—is that some evidence that there's a hard step between a cephalopod and a human?
--:-- Carl Shulman: Yeah. So that would be a place to look. It doesn't seem particularly compelling. One source of evidence on that is work by Herculano-Houzel. I hope I haven't mispronounced her name, but she's a neuroscientist who has dissolved the brains of many creatures and by counting the nuclei she was able to determine how many neurons are present in different species and find a lot of interesting trends in scaling laws. And she has a paper discussing the human brain as a scaled-up primate brain. Across a wide variety of animals and mammals in particular, there are certain characteristic changes in the relative number of neurons, size of different brain regions, how things scale up. There's a lot of structural similarity there.
--:-- And you can explain a lot of what is different about us with a pretty brute force story, which is that you expend resources on having a bigger brain, keeping it in good order, giving it time to learn. So we have an unusually long childhood, unusually long neonance period. We spend more compute by having a larger brain than other animals—more than three times as large as chimpanzees. And then we have a longer childhood than chimpanzees and much more than many other creatures. So we're spending more compute in a way that's analogous to having a bigger model and having more training time with it.
--:-- And given that we see with our AI models this sort of large, consistent benefits from increasing compute spent in those ways, and with qualitatively new capabilities showing up over and over again, particularly in areas that AI skeptics call out. In my experience, over the last 15 years, the things that people call out as like, "Ah, but the AI can't do that, and it's because of a fundamental limitation," I've gone through a lot of them. Winograd schemas, catastrophic forgetting, quite a number. And they have repeatedly gone away through scaling.
--:-- And so there's a picture that we're seeing supported from biology and from our experience with AI where you can explain that, yeah, in general, there are trade-offs where the extra fitness you get from a brain is not worth it. And so creatures wind up mostly with small brains because they can save that biological energy and that time to reproduce for digestion and so on. And humans, we actually seem to have wound up in a niche within which we greatly increased the returns to having large brains.
--:-- Language and technology are the sort of obvious candidates. When you have humans around you who know a lot of things and they can teach you, and compared to almost any other species, we have vastly more instruction from parents and the society of the young, then you're getting way more from your brain because you can learn a lot more useful skills per minute, and then you can provide the energy you need to feed that brain by hunting and gathering, by having fire that makes digestion easier.
--:-- Basically how this process goes on, it's increasing the marginal increase in reproductive fitness you get from allocating more resources along a bunch of dimensions towards cognitive ability. And so that's bigger brains, longer childhood, having our attention be more on learning. So humans play a lot and we keep playing as adults, which is a very weird thing compared to other animals. We're more motivated to copy other humans around us than even the other primates. And so these are sort of motivational changes that keep us using more of our attention and effort on learning, which pays off more when you have a bigger brain and a longer lifespan in which to learn.
--:-- Many creatures are subject to lots of predation or disease, and so if you try—you're a mayfly or a mouse—if you try and invest in a giant brain and a very long childhood, you're quite likely to be killed by some predator or some disease before you're able to actually use it. And so that means you actually have exponentially increasing costs in a given niche. So if I have a 50% chance of dying every few months as a little mammal or a little lizard or something, that means the cost of going from three months to 30 months of learning and childhood development is not 10 times the loss. Now it's 2 to the negative 10. So a factor of 1,024 reduction in the benefit I get from what I ultimately learn, because 99.9% of the animals will have been killed before that point.
--:-- We're in a niche where we're a large, long-lived animal with language and technology, so we can learn a lot from our groups. And that means it pays off to really just expand our investment on these multiple fronts in intelligence.
--:-- Dwarkesh Patel: That's so interesting. Just for the audience, the calculation about two to the whatever months is just: you have a half chance of dying this month, a half chance of dying next month. You multiply those together. Okay.
--:-- There's other species, though, that do live in flocks or as packs where you could imagine. I mean, they do have a smaller version of the development of cubs into that. They play with each other. Why isn't this a hill on which they could have climbed to human level intelligence themselves? If it's something like language or technology, humans were getting smarter before we got language. I mean, obviously, we had to get smarter to get language, right? We can't just get language without becoming smarter. So yeah, it seems like there should be other species that should have beginnings of this sort of cognitive revolution, especially given how valuable it is. Listen, we've dominated the world. You would think there would be selective pressure for it.
--:-- Carl Shulman: Evolution doesn't have foresight. The thing in this generation that gets more surviving offspring and grandchildren, that's the thing that becomes more common. Evolution doesn't look ahead and say, "Oh, in a million years you'll have a lot of descendants." It's what survives and reproduces now.
--:-- And so in fact, there are correlations where social animals do on average have larger brains, and part of that is probably that the additional social applications of brains—like keeping track of which of your group members have helped you before so that you can reciprocate, you scratch my back, I'll scratch yours, remembering who's dangerous within the group, that sort of thing—it's an additional application of intelligence. And so there's some correlation there.
--:-- But what it seems like is that, yeah, in most of these cases, it's enough to invest more, but not invest to the point where a mind can easily develop language and technology and pass it on. And so you see bits of tool use in some other primates who have an advantage. Compared to, say, the whales who have quite large brains partly because they are so large themselves, they have some other things, but they don't have hands, which reduces a bunch of ways in which brains can pay off and investments in the functioning of that brain.
--:-- But yeah, primates will use sticks to extract termites. Capuchin monkeys will open clams by smashing them with a rock. So there's bits of tool use. But what they don't have is the ability to sustain culture. A particular primate will maybe discover one of these tactics, and maybe it'll be copied by their immediate group. But they're not holding onto it that well. They're like, well, when they see the other animal do it, they can copy it in that situation. They don't actively teach each other. Their population locally is quite small. So it's easy to forget things, easy to lose information. And in fact, they remain technologically stagnant for hundreds of thousands of years.
--:-- And we can actually look at some human situations. There's an old paper, I believe, by the economist Michael Kremer, that talks about technological growth in the different continents for human societies. And so you have Eurasia as the largest integrated connected area. Africa is partly connected to it, but the Sahara Desert restricts the flow of information and technology and such. And then you had the Americas, which after the colonization from the land bridge were largely separated and are smaller than Eurasia. Then Australia, and then you had smaller island situations like Tasmania. And so technological progress seemed to have been faster the larger the connected group of people.
--:-- In the smallest groups, like in Tasmania, you had a relatively small population and they actually lost technology. They lost some fishing techniques. If you have a small population and you have some limited number of people who know a skill and they happen to die, or there's some change in circumstances that causes people not to practice or pass on that thing, then you lose it. And if you have few people, you're doing less innovation. The rate at which you lose technologies to some kind of local disturbance and the rate at which you create new technologies can wind up in balance.
--:-- The great change of hominids and humanity is that we wound up in this situation where we were accumulating faster than we were losing. And as we accumulated, those technologies allowed us to expand our population, they created additional demand for intelligence, so that our brains became three times as large as that of chimpanzees. Our ancestors had a similar brain size.
--:-- Dwarkesh Patel: Okay.
--:-- Carl Shulman: And then the crucial point, I guess in relevance to AI, is that the selective pressures against intelligence in other animals are not acting against these neural networks, because they're not going to get eaten by a predator if they spend too much time becoming more intelligent. We're explicitly training them to become more intelligent. So we have good first principles reason to think that if it was scaling that made our minds this powerful, and if the things that prevented other animals from scaling are not impinging on these neural networks, then these things should just continue to become very smart.
--:-- Dwarkesh Patel: Yeah, we're growing them in a technological culture where there are jobs like software engineer that depend much more on cognitive output and less on things like metabolic resources devoted to the immune system or to building big muscles to throw spears.
--:-- This is kind of a side note, but I'm just kind of interested. I think you referenced at some point the Chinchilla scaling paper. For the audience, this is a paper from DeepMind which describes if you have a model of a certain size, what is the optimum amount of data that it should be trained on. So you can imagine bigger models, you can use more data to train them. And in this way, you can figure out where should you spend your compute: do you spend it on making the model bigger, or should you spend it on training it for longer?
--:-- I'm curious if, in the case of different animals, in some sense they're like model sizes—how big their brain is—and their training data sizes—how long their cubs or how long their infants or toddlers are before they're full adults. Is there some sort of scaling law?
--:-- Carl Shulman: Yeah, I mean, the Chinchilla scaling is interesting because we were talking earlier about the cost function for having a longer childhood. And so it's exponentially increasing in the amount of training compute you have when you have exogenous forces that can kill you. Whereas when we do big training runs, the cost of throwing in more GPUs is almost linear. And it's much better to be linear than exponentially decaying.
--:-- Dwarkesh Patel: Oh, that's a really good point.
--:-- Carl Shulman: As you expand resources, Chinchilla scaling would suggest that for a brain of sort of human size, it would be optimal to have many millions of years of education. But obviously that's impractical because of exogenous mortality for humans. And so there's a fairly compelling argument that relative to the situation where we would train AI, animals are systematically way under-trained.
--:-- Dwarkesh Patel: That's so interesting. They're more efficient than our models. We still have room to improve our algorithms to catch up with the efficiency of brains, but they are laboring under that disadvantage. That is so interesting.
--:-- Okay, so I guess another question you could have is: humans got started on this evolutionary hill climbing route where we're getting more intelligent that has more benefits for us. Why didn't we go all the way on that route? If intelligence is so powerful, why aren't all humans as smart as we know humans can be? At least that's smart. If intelligence is so powerful, why hasn't there been stronger selective pressure? I understand, like, hip size, you can't give birth to a really big headed baby or whatever. But you would think evolution would figure out some way to offset that if intelligence has such big power and it's so useful.
--:-- Carl Shulman: Yeah, I think if you actually look at it quantitatively, that's not true. And even in recent history, it looks like a pretty close balance between the costs and the benefits of having more cognitive abilities. And so you say, who needs to worry about the metabolic costs?
--:-- Carl Shulman: Humans put an order of 20% of our metabolic energy into the brain, and it's higher for young children. So 20% of the total, and then there's breathing and digestion and the immune system. For most of history, people have been dying left and right. A very large proportion of people would die of infectious disease, and if you put more resources into your immune system, you survive. It's life or death pretty directly via that mechanism.
--:-- This is also related to famine. People die more of disease during famine, so there's boom or bust. If you have 20% less metabolic requirements as a child, or maybe 40 or 50% less, you're much more likely to survive that famine. So these are pretty big trade-offs.
--:-- Then there's the trade-off about cleaning mutational load. Every generation, new mutations and errors happen in the process of reproduction. We know there are many genetic abnormalities that occur through new mutations each generation. Down syndrome is the chromosomal abnormality you can survive; all the others just kill the embryo, so we never see them. But Down syndrome occurs a lot, and there are many other lethal mutations. As you go to less damaging ones, there are enormous numbers of less damaging mutations that are degrading every system in the body.
--:-- Each generation, evolution has to pull away at some of this mutational load. The priority with which that mutational load is pulled out scales in proportion to how much the traits it's affecting impact fitness. You get new mutations that impact your resistance to malaria, you get new mutations that damage brain function. Those mutations are purged each generation. If malaria is a bigger difference in mortality than the incremental effectiveness as a hunter-gatherer you get from being slightly more intelligent, then you'll purge that mutational load first.
--:-- Similarly, humans have been vigorously adapting to new circumstances. Since agriculture, people have been developing things like the ability to digest breads with amylase, the ability to digest milk. If you're evolving for all of these things, and some of the things that give an advantage for that incidentally carry along nearby them some negative effect on another trait, then that other trait can be damaged.
--:-- So it really matters how important to survival and reproduction cognitive abilities were compared to everything else the organism has to do. Surviving feast and famine, having the physical abilities to do hunting and gathering — even if you're very good at planning your hunting, being able to throw a spear harder can be a big difference. That needs energy to build those muscles and sustain them. Given all of these factors, it's not a slam dunk to invest at the margin.
--:-- Today, having bigger brains is associated with greater cognitive ability, but it's modest. Large-scale pre-registered studies with MRI data show a correlation of maybe 0.25 to 0.3. The standard deviation of brain size is like 10%. So if you double the size of the brain, going from the existing brain costing 20% of metabolic energy up to 40%, that's eight standard deviations of brain size. If the correlation is 0.25, then you get a gain from that: eight standard deviations of brain size, two standard deviations of cognitive ability.
--:-- In our modern society where cognitive ability is very rewarded — finishing school, becoming an engineer or a doctor can pay off a lot financially — the average observed return in income is a one or two percent proportional increase. There are more effects in the tail, more effect in professions like STEM. But on the whole, it's not a five or ten percent increase where you could tell a story that this is hugely increasing the amount of food you could have or the number of children you could support. It's a modest effect, and the metabolic costs will be large. Throw in these other aspects, and you can tell the story. We can just see there was not very strong rapid directional selection on the thing, which there would be if by solving a math puzzle you could defeat malaria. Then there would be more evolutionary pressure.
--:-- Dwarkesh Patel: That is so interesting. And not to mention, of course, that if you had two times the brain size, without a C-section, your mother would die, or both would die. This is a question I've actually been curious about for over a year. I've briefly tried to look up an answer. I know this is off topic, I apologize to the audience, but I was super interested in it. That was like the most comprehensive and interesting answer I could have hoped for.
--:-- Okay, so we have a good explanation from first principles, evolutionary reason for thinking that intelligence scaling up to humans is not implausible just by throwing more scale at it.
--:-- Carl Shulman: I would also add, this was something that would have mattered to me more in the 2000s. We have the brain right here with us, available for neuroscience to reverse engineer its properties. In the 2000s, when I said I expect this by the middle of the century-ish, that was a backstop if we found it absurdly difficult to get to the algorithms. Then we would learn from neuroscience. But in the actual history, it's really not like that. We develop things in AI, and then we can say, "Oh yeah, this is sort of like this thing in neuroscience, or maybe this is a good explanation." It's not as though neuroscience is driving AI progress. It's similar to how planes were inspired by the existence proof of birds, but jet engines don't flap.
--:-- Dwarkesh Patel: All right, so scaling is a good reason to think scaling might work. So we spend $100 billion, and we have something that is human level or can help significantly with AI research. That might be on the earlier end, but I definitely would not rule that out, given the rates of change we've seen with the last few scale-ups.
--:-- At this point, somebody might be skeptical. They might say, "Listen, we already have a bunch of human researchers. How powerful is the incremental researcher?" You might say, "Well, this is like thousands of researchers." I don't know how to express skepticism exactly, but skepticism is skeptical of just generally the effect of scaling up the number of people working on the problem to rapid progress on that problem.
--:-- Somebody might think, "Listen, with humans, the reason population working on a problem is such a good proxy for progress is that there's already so much variation accounted for. When you say there are a million people working on a problem, there are hundreds of super geniuses working on it, thousands of people who are very smart. Whereas with an AI, all the copies are the same level of intelligence. If it's not super genius intelligence, the total quantity might not matter as much."
--:-- Carl Shulman: I'm not sure what your model is here. Is this the model that the diminishing returns kick off suddenly, has a cliff right where we are? There were results in the past from throwing more people at problems. This has been useful in historical prediction. There's this idea of experience curves and Wright's law, basically measuring cumulative production in a field, which is also a measure of the scale of effort and investment.
--:-- People have used this correctly to argue that renewable energy technology like solar would be falling rapidly in price because it was going from a low base of very small production runs, not much investment in doing it efficiently. Climate advocates correctly called that out. The futurist Ramaz Naam actually has some interesting writing on this — he correctly called out that there would be really drastic fall in prices of solar and batteries because of the increasing investment going into that. The Human Genome Project would be another example.
--:-- There's real evidence. These observed correlations from ideas getting harder to find have held over a fair range of data and over quite a lot of time. So I'm wondering what the nature of the deviation you're thinking of is.
--:-- Dwarkesh Patel: Maybe this is a good way to describe what happens when more humans enter a field. But does it even make sense to say a greater population of AIs is doing AI research if there's more GPUs running a copy of GPT doing AI research? How applicable are these economic models of the quantity of humans working on a problem to the magnitude of AIs working on a problem?
--:-- Carl Shulman: If you have AIs that are directly automating particular jobs that humans were doing before, then with additional compute, we can run more copies of them to do more of those tasks simultaneously. We can also run them at greater speed. Some people have an intuition that what matters is time, not how many people are working on a problem at a given point. I think that doesn't bear out super well.
--:-- AI can also be run faster than humans. If you have a set of AIs that can do the work of individual human researchers and run at 10 times or 100 times the speed, we could ask: could the human research community have solved these algorithmic problems, done things like invent transformers, over 100 years? If we have AIs with an effective population similar to the humans, but running 100 times as fast, you have to tell a story where the AIs can't really do the same things as the humans.
--:-- We're talking about what happens when the AIs are more capable of doing that, although they become more capable as lesser capable versions of themselves help us make themselves more capable. You have to kick-start that at some point.
--:-- Dwarkesh Patel: Is there an example in analogous situations? Is intelligence unique in the sense that you have a feedback loop where a system's outputs are feeding into its own inputs? Because if we're talking about something like Moore's Law or the cost of solar, we're throwing more people at the problem and making a lot of progress, but we don't have this additional part of the model where Moore's Law leads to more humans somehow, and more humans becoming researchers.
--:-- Carl Shulman: You do actually have a version of that in the case of solar. You have a small infant industry providing solar panels for space satellites, then getting increasing amounts of subsidized government demand because of worries about fossil fuel depletion and then climate change. You have the dynamic where visible successes with solar or lowering prices open up new markets. There's a huge transition where renewables become cheap enough to replace large chunks of the electric grid. Earlier, you're dealing with very niche situations — satellites where it's very difficult to refuel, remote areas, then moving to the sunniest areas in the world with the biggest solar subsidies. There was an element where more investment has been thrown into the field and the market has rapidly expanded as the technology improved.
--:-- But I think the closest analogy is actually the long-run growth of human civilization itself. I know you had Holden Karnofsky from the Open Philanthropy Project on earlier and discussed some of this research about the long-run acceleration of human population and economic growth. Developing new technologies allowed human population to expand, humans to occupy new habitats and new areas, then to invent agriculture, which supported larger populations, then even more advanced agriculture in the modern industrial society. Total technology and output allowed you to support more humans, who then would discover more technology and continue the process.
--:-- That was boosted because on top of expanding the population, the share of human activity going into invention and innovation went up. That was a key part of the industrial revolution. There was no such thing as a corporate research lab or an engineering university prior to that. So you're both increasing the total human population and the share of it going in. This population dynamic is pretty analogous. Humans invent farming, they can have more humans, then they can invent industry, and so on.
--:-- Dwarkesh Patel: Maybe somebody would be skeptical that with AI progress specifically, it's not just a matter of some farmer figuring out crop rotation or some blacksmith figuring out how to do metallurgy better. To make even a 50% improvement in productivity, you basically need something with an IQ close to Ilya Sutskever. There's a discontinuity — you're contributing very little to productivity, and then you're Ilya and you contribute a lot. But becoming Ilya... you see what I'm saying? There's not a gradual increase in capabilities that leads to the feedback.
--:-- You're imagining a case where the distribution of tasks is such that there's nothing you can individually automate that particularly helps. The ability to contribute to AI research is really end-loaded. Is that what you're saying?
--:-- Carl Shulman: Yeah. We already see this in really high IQ companies or projects. Theoretically, Jane Street or OpenAI could hire a bunch of mediocre people to do menial tasks that free up the time of the really smart people, but they don't do that. The transaction costs, whatever else.
--:-- Self-driving cars would be another example where you have a very high quality threshold. When your performance as a driver is worse than a human — you have 10 times the accident rate or 100 times the accident rate — the cost of insurance for that, which is a proxy for people's willingness to ride the car, would be such that the insurance costs would absolutely dominate. Even if you have zero labor costs, it's offset by the increased insurance costs. There are lots of cases where partial automation is not in practice very usable because complementing other resources, you're going to use those other resources less efficiently.
--:-- In a post-AGI future, the same thing can apply to humans. People can say, "Well, comparative advantage — even if AIs can do everything better than a human, it's still worth something. A human can lift a box, that's something." Now there's a question of property rights — if they could just slice up the human and use them to make more robots. But even absent that, in such an economy, you wouldn't want to let a human worker into any industrial environment because in a clean room, they'll be emitting skin cells and messing things up. You need to have an atmosphere there. You need a bunch of supporting tools and resources and materials. Those supporting resources and materials will do a lot more productively working with AI and robots rather than a human. So you don't want to let a human anywhere near the thing. You don't want to have a gorilla wandering around in a China shop, even if you've trained it to most of the time pick up a box for you. If you give it a banana, it's just not worth it.
--:-- Dwarkesh Patel: Why is that not a good objection to... I mean, I think that is one of the ways in which partial automation can fail to really translate into a lot of economic value. That's something that will attenuate as we go on and as the AI is more able to work independently and more able to handle its own screw-ups, get more reliable. But the way in which it becomes more reliable is by AI progress speeding up, which happens if AI can contribute to it. But if there is some sort of reliability bottleneck that prevents it from contributing to that progress, then you don't have the loop, right?
--:-- Carl Shulman: Yeah, I mean, this is why we're not there yet.
--:-- Dwarkesh Patel: Right, but then what is the reason to think we'll be there?
--:-- Carl Shulman: The broad reason is we have these inputs that are scaling up. Epoch, which I mentioned earlier, has a paper called "Compute Trends in Three Eras of Machine Learning" or something like that. They look at the compute expended on machine learning systems since the founding of the field of AI at the beginning of the 1950s. Mostly it grows with Moore's Law. People are spending a similar amount on their experiments, but they can just buy more with that because the compute is coming.
--:-- That data covers over 20 orders of magnitude, maybe like 24. Of all of those increases since 1952, a little more than half of them happened between 1952 and 2010. All the rest is since 2010. We've been scaling that up like four times as fast as was the case for most of the history of AI. We're running through the orders of magnitude of possible resource inputs you could need for AI much, much more quickly than we were for most of the history of AI. That's why this is a period with a very elevated chance of AI per year — because we're moving through so much of the space of inputs per year.
--:-- It looks like this scale taken to its conclusion will cover another bunch of orders of magnitude. That's actually a large fraction of those that are left before you start running into saying, "Well, this is going to have to be like evolution with the sort of simple hacks we get to apply." We're selecting for intelligence the whole time. We're not going to do the same mutation that causes fatal childhood cancer a billion times, even though we keep getting the same fatal mutations, even though they've been done many times. We use gradient descent, which takes into account the derivative of improvement on the loss all throughout the network. We don't throw away all the contents of the network with each generation where you compress down to a little DNA.
--:-- There's that bar of, if you're going to do brute force like evolution combined with these very simple ways, we can save orders of magnitude on that. We're going to cover, I think, a fraction that's like half of that distance in this scale-up over the next 10 years or so.
--:-- If you started off with a kind of vague uniform prior, you'd say, "Well, you probably can't make AGI with the amount of compute that would be involved in a fruit fly existing for a minute," which would be the early days of AI. Maybe you would get lucky. We were able to make calculators because calculators benefited from very reliable, serially fast computers, and we could take a tiny, tiny fraction of a human brain's compute and use it for a calculator. We couldn't take an ant's brain and rewire it to calculate. It's hard to manage ant farms, let alone get them to do arithmetic for you. There were some things where we could exploit the differences between biological brains and computers to do stuff super efficiently on computers. We would doubt that we would be able to do so much better than biology that with a tiny fraction of an insect's brain, we'd be able to get AI early on.
--:-- On the far end, it seemed very implausible that we couldn't do better than completely brute force evolution. In between, you have some number of orders of magnitude of inputs where it might be. In the 2000s, I would say I'm going to have a pretty uniformish prior. I'm going to put weight on it happening at the equivalent of 10 to the 25 ops, 10 to the 30, 10 to the 35, sort of spreading out over that. Then I can update on other information. In the short term, in 2005, I would say, "Well, I don't see anything that looks like the cusp of AGI. So I'm also going to lower my credence for the next five years or the next 10 years." That would be kind of a vague prior.
--:-- When we take into account how quickly we're running through those orders of magnitude — if I have a uniform prior, I assign half of my weight to the first half of remaining orders of magnitude. If we're going to run through those over the next 10 years and some, then that calls on me to put half of my credence, conditional on we're ever going to make AI — which seems likely, where it's a material object easier than evolution — I've got to put similarly a lot of my credence on AI happening in this scale-up. That's supported by what we're seeing in terms of the rapid advances in capabilities with AI and LLMs in particular.
--:-- Dwarkesh Patel: Okay, that's actually a really interesting point. So now somebody might say, "Hey, listen, there's not some sense in which AIs could universally speed up the progress of OpenAI by 50% or 100% or 200%. If they're not able to do everything better than Ilya Sutskever can, there's going to be something in which we're bottlenecked by the human researchers. Bottleneck effects dictate that the slowest moving part of the organization will be the one that determines the speed of progress of the whole organization or the whole project. That means unless you get to the point where you're doing everything and everybody in the organization can do, you're not going to significantly speed up the progress of the whole project as a whole."
--:-- Carl Shulman: That is a hypothesis, and I think there's a lot of truth to it. When we think about the ways in which AI can contribute, there are things we talked about before like the AI setting up its own curriculum. That's something that Ilya can't do directly, doesn't do directly. There's a question of how much that improves performance.
--:-- There are these things where the AI helps to just produce some code for some task. It's beyond "Hello World" at this point. But the sort of thing I hear from AI researchers at leading labs is that on their core job where they're most expert, it's not helping them that much. But their job often does involve, "Oh, I've got to code something that's out of my usual area of expertise," or "I want to research this question," and it helps them there. That saves some of their time and frees them to do more of the bottlenecked work.
--:-- The idea that everything is dependent on Ilya and Ilya is so much better than the hundreds of other employees — I think a lot of people who are contributing are doing a lot of tasks. You can have quite a lot of gain from automating some areas where you then do an absolutely enormous amount of it relative to what you would have done before. Things like designing the custom curriculum — you may have had some humans put some work into that, but you're not going to employ billions of humans to produce it at scale. It winds up being a larger share of the progress than it was before. You get some benefit from these sorts of things where pieces of my job I can now hand off to the AI, and it lets me focus more on the things that the AI still can't do.
--:-- Later on, you get to the point where the AI can do your job, including the most difficult parts. Maybe it has to do that in a different way. Maybe it spends a ton more time thinking about each step of a problem than you do. That's the late end. The stronger these bottleneck effects are, the more the economic returns and scientific returns are end-loaded towards getting sort of full AGI. The weaker the bottlenecks are, the more interim results will be really paying off.
--:-- Dwarkesh Patel: I guess I probably disagree with you on how much the sort of the Ilyas of organizations seem to matter. Just from the evidence alone, how many of the big breakthroughs in deep learning in general were that single individual responsible for? How much of his time is he spending doing anything that's not that? Copilot is helping him on... I'm guessing most of it is just managing people, coming up with ideas, trying to understand systems. If the five or ten people who are like that at OpenAI or Anthropic are basically the way in which the progress is happening, or at least the algorithmic progress, then how much is better and better copilot — I know copilot is not the thing you're talking about with 20% automation, but something like that — how much is that contributing to the core function of the research scientist?
--:-- Carl Shulman: Not sure quantitatively how much we disagree about the importance of key research employees. I certainly think that some researchers add more than 10 times the average employee, even much more. Obviously, managers can add an enormous amount of value by proportionately multiplying the output of the many people that they manage. That's the kind of thing we were discussing earlier when talking about if you had sort of full human-level AI or AI that had all of the human capabilities plus AI advantages, you'd benchmark not off of what the typical human performance is but peak human performance and beyond.
--:-- I accept all that. I do think it makes a big difference for people how much they can outsource a lot of the tasks that are less creative. An enormous amount is learned by experimentation. ML has been quite an experimental field. There's a lot of engineering work in building large superclusters, making hardware-aware optimization and coding of these things, being able to do the parallelism in large models. The engineers are busy, and it's not just a "big thoughts" kind of area.
--:-- The other branch is where will the AI advantages and disadvantages be?
--:-- Carl Shulman: And so one AI advantage is being omni and familiar with the newest things. I mentioned before, no human has a million years of TensorFlow experience. To the extent that we're interested in the very, very cutting edge of things that have been developed quite recently, then AI that can learn about them in parallel and experiment and practice with them in parallel can learn much faster than a human potentially.
--:-- The area of computer science is one that is especially suitable for AI to learn in a digital environment. It doesn't require driving a car around that might kill someone, have enormous costs. You can do unit tests. You can prove theorems. You can do all sorts of operations entirely in the confines of a computer. That's one reason why programming has been benefiting more than a lot of other areas from LLMs recently, whereas robotics is lagging.
--:-- And then just considering, they are getting better at things like the GRE math, at programming contests. Some people have forecasts and predictions outstanding about things like doing well on the informatics Olympiad and the math Olympiad. In the last few years, when people tried to forecast the MMLU benchmark, which had a lot of more sophisticated graduate student science kind of questions, AI knocked that down a lot faster than AI researchers who had registered forecasts on it.
--:-- If you're getting top-notch scores on graduate exams and creative problem solving, it's not obvious that that sort of area will be a relative weakness of AI. In fact, computer science is in many ways especially suitable because of getting up to speed with new areas, being able to get rapid feedback from the interpreter at scale.
--:-- Dwarkesh Patel: But do you get rapid feedback if you're doing something that's more analogous to research? If you're like, let's say you have a new model or something and it's like, if we put in $10 million on a mini training run on this, this would be much better.
--:-- Carl Shulman: For very large models, those experiments are going to be quite expensive. So you're going to look more at whether you can build up this capability by generalization from things like many math problems, programming problems, working with small networks.
--:-- Dwarkesh Patel: Fair enough. I actually saw Scott Aaronson was one of my professors in college. I took his quantum information class and I did okay in it. He recently wrote a blog post where he said, "I had GPT-4 take my quantum information test and it got a B." And I was like, damn, I gotta see the final. So yeah, I've updated in the direction that getting a B on the test probably means you understand quantum information pretty well, with different areas of strength and weaknesses than the human students.
--:-- Carl Shulman: Sure.
--:-- Dwarkesh Patel: Would it be possible for this sort of intelligence explosion to happen without any sort of hardware progress? If hardware progress stopped, would this feedback loop still be able to produce some sort of explosion with only software?
--:-- Carl Shulman: If we say that the technology is frozen, which I think is not the case right now — NVIDIA has managed to deliver significantly better chips for AI workloads for the last few generations, H100, A100, V100. If that stops entirely, then what you're left with, and maybe we'll define this as no more nodes, Moore's law is over. At that point, the kind of gains you get in amount of compute available come from actually constructing more chips. There are economies of scale you could still realize there.
--:-- Right now, a chip maker has to amortize the R&D cost of developing the chip. Then the capital equipment is created. You build a fab, its peak profits are going to come in the few years when the chips it's making are at the cutting edge. Later on, as the cost of compute exponentially falls, you keep the fab open because you can still make some money given that it's built. But of all the profits the fab will ever make right now, they're relatively front-loaded because when its technology is near the cutting edge.
--:-- In a world where Moore's Law ends, you wind up with these very long production runs where you can keep making chips that stay at the cutting edge and where the R&D costs get amortized over a much larger base. The R&D basically drops out of the price. Then you get some economies of scale from just making so many fabs, in the way that when the auto industry expands, costs fall.
--:-- Right now, ASML has many incredibly exotic suppliers that make some bizarre part of the thousands of parts in one of these ASML machines. You can't get it anywhere else. They don't have standardized equipment for their thing because this is the only use for it. In a world where we're making 10, 100 times as many chips at the current node, they would benefit from scale economies. All of that would become more mass production industrialized.
--:-- You combine all of those things and it seems like capital costs of buying a chip would decline. But the energy costs of running the chip would not. Right now, energy costs are a minority of the cost, but they're not trivial. They passed 1% a while ago, and they're inching up towards 10% and beyond. You can maybe get another order of magnitude cost decrease from getting really efficient in the capital construction, but energy would still be a limiting factor after the end of actually improving the chips themselves.
--:-- Dwarkesh Patel: Got it. And when you say there would be a greater population of AI researchers, are we using population as a sort of thinking tool of how they could be more effective? Or do you literally mean that the way you expect these AIs to contribute a lot to research is by just having a million copies of a researcher thinking about the same problem? Or is it just a usual thinking model for what it would look like to have a million times smarter AI working on that problem?
--:-- Carl Shulman: That's definitely a lower bound model. Often I'm meaning something more like effective population, or you'd need this many people to have this effect. We were talking earlier about the trade-off between training and inference in board games. You can get the same performance by having a bigger model or by calling the model more times. In general, it's more effective to have a bigger, smarter model and call it less time, up until the point where the costs equalize between them.
--:-- We would be taking some of the gains of our larger compute on having bigger models that are individually more capable. There would be a division of labor. The tasks that were most cognitively demanding would be done by these giant models, but for some very easy tasks, you don't want to expend that giant model if a model 1/100th the size can take that task. Larger models would be in the positions of researchers and managers, and they would have swarms of AIs of different sizes as tools that they can make API calls to.
--:-- Dwarkesh Patel: Okay, we accept the model. And now we've gone to something that is at least as smart as Ilya Sutskever on all the tasks relevant to AI progress. And you can have so many copies of it. What happens in the world now? What are the next months or years, whatever timeline is relevant, look like?
--:-- Carl Shulman: To be clear, what's happened is not that we have something that has all of the abilities and advantages of humans plus the AI advantages. What we have is something that, possibly by doing a ton of calls to make up for being individually less capable, is able to drive forward AI progress. That process is continuing. AI progress has accelerated greatly in the course of getting there. So maybe we go from our eight months doubling time of software progress in effective compute to four months or two months.
--:-- There's a report by Tom Davidson at the Open Philanthropy Project, which spun out of work I had done previously. I advised and helped with that project, but Tom really carried it forward and produced a very nice report and model, which Epoch is hosting. You can plug in your own version of the parameters. There is a lot of work estimating the parameters: things like what's the rate of software progress, what's the return to additional work, how does performance scale at these tasks as you boost the models.
--:-- As we were discussing, the sort of broadly human level in every domain with all the advantages is pretty deep into that. If already we can have an eight-month doubling time for software progress, then by the time you get to that kind of point it's maybe more like four months, two months, going into one month. If the thing is just proceeding at full speed, then each doubling can come more rapidly.
--:-- We can talk about what are the spillovers. How do the models get more capable? They can be doing other stuff in the world. They can spend some of their time making Google search more efficient. They can be hired as chatbots with some inference compute. If that intelligence explosion process is allowed to proceed, then what happens is you improve your software by a factor of two. The efforts needed to get the next doubling are larger, but they're not twice as large. Maybe they're like 25%, 35% larger. So each one comes faster and faster until you hit limitations. You can no longer make further software advances with the hardware that you have.
--:-- Looking at reasonable parameters in that model, it seems to me if you have these giant training runs, you can go very far. The way I would see this playing out is, how do the AIs get better and better at research? They can work on different problems. They can work on improving software. They can work on improving hardware. They can do things like create new industrial technologies, new energy technology. They can manage robots. They can manage human workers as executives and coaches. You can do all of these things.
--:-- AI winds up being applied where the returns are highest. I think initially, the returns are especially high in doing more software. The reason for that is, again, if you improve the software, you can update all of the GPUs that you have access to. Your cloud compute is suddenly more potent. If you design a new chip design, it'll take a few months to produce the first ones, and it doesn't update all of your old chips. So you have an ordering where you start off with the things where there's the lowest dependence on existing stocks and you can just take whatever you're developing and apply it immediately.
--:-- Software runs ahead. You're getting more towards the limits of that software. I think that means having all the human advantages but combined with AI advantages. Given the kind of compute that would be involved if we're talking about this hundreds of billions of dollar, trillion dollar training run, there's enough compute to run tens of millions, hundreds of millions of sort of human scale minds. They're probably smaller than human scale to be similarly efficient at the limits of algorithmic progress because they have the advantage of a million years of education. They have the other advantages we talked about.
--:-- You've got that wild capability and the software, further software gains are running out or they start to slow down again because you're just getting towards the limits of you can't do any better than the best.
--:-- Dwarkesh Patel: By the time they're running out, have we already hit superintelligence?
--:-- Carl Shulman: Yes. You're widely superintelligent. We've left the galaxy, even metaphorically. Just by having the abilities that humans have and then combining it with being very well focused and trained in the task beyond what any human could be and then running faster and such.
--:-- Dwarkesh Patel: Got it. Continue.
--:-- Carl Shulman: I'm not going to assume that there are huge qualitative improvements you can have. I'm not going to assume that humans are very far from the efficient frontier of software, except with respect to things like we had limited lifespan so we couldn't train super intensively, we couldn't incorporate other software into our brains, we couldn't copy ourselves, we couldn't run at fast speeds. You got all of those capabilities.
--:-- Now I'm skipping ahead of the most important months in human history. I can talk about what it looks like if it's just the AIs took over, they're running things as they like, how do things expand? I can talk about how this goes in a world where we've roughly, or at least so far, managed to retain control of where these systems are going. By jumping ahead, I can talk about how would this translate into the physical world.
--:-- This is something that I think is a stopping point for a lot of people in thinking about what an intelligence explosion would look like. They have trouble going from "there's stuff on servers and cloud compute and that gets very smart" to "how does what I see in the world change? How does industry or military power change? If there's an AI takeover, what does that look like? Are there killer robots?"
--:-- One course we might go down is to discuss during that wildly accelerating transition, how did we manage that? How do you avoid it being catastrophic? Another route we could go is how does the translation from wildly expanded scientific R&D capabilities and intelligence on these servers translate into things in the physical world? You're moving along in order of what has the quickest impact, or where you can have an immediate change.
--:-- One of the most immediately accessible things is where we have large numbers of devices or artifacts or capabilities that are already AI operable. With hundreds of millions equivalent researchers, you can quickly solve self-driving cars, make the algorithms much more efficient, do great testing and simulation, and then operate a large number of cars in parallel if you need to get some additional data to improve the simulation and reasoning. Humans with quite little data are able to achieve human-level driving performance.
--:-- After you've really maxed out the easily accessible algorithmic improvements in this software-based intelligence explosion that's mostly happening on server farms, then you have minds that have been able to really perform on a lot of digital-only tasks. They're doing great on video games. They're doing great at predicting what happens next in a YouTube video. If you have a camera that they can move, they're able to predict what will happen at different angles. Humans do this a lot where we naturally move our eyes in such a way to get images from different angles and different presentations and then predicting combined from that.
--:-- You can operate many cars, many robots at once to get very good robot controllers. You should think that all the existing robotic equipment or remotely controllable equipment that is wired for that, the AIs can operate that quite well.
--:-- Dwarkesh Patel: I think some people might be skeptical that existing robots, given their current hardware, have the dexterity and the maneuverability to do a lot of physical labor that any AI might want to do. Do you have a reason for thinking otherwise?
--:-- Carl Shulman: There's also not very many of them. Production of industrial robots is hundreds of thousands per year. They can do quite a bit in place. Elon Musk is promising a robot in the tens of thousands of dollars, tens of thousands of humanoid robots. That may take a lot longer than is said. This has happened with other technologies, but that's a direction to go.
--:-- Most immediately, hands are actually probably the most scarce thing. But if we consider what do human bodies provide? There's the brain. In this situation, we have now an abundance of high quality brain power that will be increasing. The AIs will have designed new chips, which will be rolling out from the TSMC factories, and they have ideas and designs for the production of new fab technologies, new nodes, and additional fabs.
--:-- But looking around the body, there are legs to move around. Not always that necessary. Wheels work pretty well. Being in a place, you don't need most people most of the time in factory jobs and office jobs. Office jobs, many of them can be fully virtualized. Some amount of legs, wheels, other transport.
--:-- You have hands, and hands are something that are on the expensive end. In robots, we can make them. They're made in very small production runs, partly because we don't have the control software to use them well. In this world, the control software is fabulous, and so people will produce much larger production runs of them over time, possibly using technology we recognize, possibly with quite different technology, but just taking what we've got.
--:-- Right now, the robot arm industry, the industrial robot industry produces hundreds of thousands of machines a year. Some of the nicer ones are like $50,000. In aggregate, the industry has tens of billions of dollars of revenue. By comparison, the automobile industry produces over 60 million cars a year. It has revenue of over $2 trillion per annum. Converting that production capacity over towards robot production would be one of the things, if there's not something better to do.
--:-- In World War II, industrial conversion of American industry took place over several years and really amazingly ramped up military production by converting existing civilian industry. That was without the aid of superhuman intelligence and management at every step in the process. Every part of that would be very well designed. You'd have AI workers who understood every part of the process and could direct human workers.
--:-- Even in a fancy factory, most of the time, it's not the hands doing a physical motion that a worker is being paid for. They're often looking at things or deciding what to change. The time spent in manual motion is a limited portion of that. In this world of abundant AI cognitive abilities, where the human workers are more valuable for their hands than their heads, you could have a worker, even a worker previously without training and expertise in the area, who has a smartphone, maybe a smartphone on a headset. We have billions of smartphones which have eyes and ears and methods for communication for an AI to be talking to a human and directing them in their physical motions.
--:-- With skill as a guide and coach that is beyond any human, they're going to be a lot better at telepresence and remote work. They can provide VR and augmented reality guidance to help people get better at doing the physical motions that they're providing in the construction.
--:-- Say you convert the auto industry to robot production. If it can produce an amount of mass of machines that is similar to what it currently produces, that's enough for a billion human-size robots a year. The value per kilogram of cars is somewhat less than high-end robots, but you're also cutting out most of the wage bill because most of the wage bill is payments ultimately to human capital and education, not to the physical hand motions and lifting objects.
--:-- At the sort of existing scale of the auto industry, you can make a billion robots a year. The auto industry is two or three percent of the existing economy. You're replacing these cognitive things. If right now physical hand motions are like 10% of the work, redirect humans into those tasks.
--:-- In the world at large right now, mean income is on the order of a year. But in rich countries, skilled workers earn more per year. Some of that is just management roles, of which only a certain proportion of the population can have, but just being an absolutely exceptional peak human performance of some of these construction and such roles. Raising productivity to match the most productive workers in the world is room to make a very, very big gap.
--:-- With AI replacing skills that are scarce in many places where there's abundant, currently low wage labor, you bring in the AI coach and someone who was previously making very low wages can suddenly be super productive by just being the hands for an AI.
--:-- On a naive view, if you ignore the delay of capital adjustment, of building new tools for the workers, just raise typical productivity for workers around the world to be more like rich countries and get 5X, 10X like that. Get more productivity by, with AI handling the difficult cognitive tasks, reallocating people from office jobs to providing physical motions. Since right now that's a small proportion of the economy, you can expand the hands for manual labor by an order of magnitude within a rich country just because most people are sitting in an office or even on a factory floor not continuously moving.
--:-- You've got billions of hands lying around in humans to be used in the course of constructing your waves of robots. Once you have a quantity of robots that is approaching the human population, and they work 24/7, of course the human labor will no longer be valuable as hands and legs. But at the very beginning of the transition, just like new software can be used to update all of the GPUs to run the latest AI, humans are this sort of legacy population with an enormous number of underutilized hands and feet that the AI can use for the initial robot construction.
--:-- Cognitive tasks are being automated and the production of them is greatly expanding. The physical tasks which complement them are utilizing humans to do the parts that existing robots can't do.
--:-- Dwarkesh Patel: Is the implication of this that world production would increase just a tremendous amount, or that AI could get a lot done of whatever motivations it has?
--:-- Carl Shulman: There's an enormous increase in production for humans just switching over to the role of providing hands and feet for AI where they're limited. This robot industry is a natural place to apply it. If you go to something that's like 10x the size of the current car industry in terms of its production, which would still be like a third of our current economy, and the aggregate productive capabilities of the society with AI support are going to be a lot larger, they make 10 billion humanoid robots a year.
--:-- If you do that, the legacy population of a few billion human workers is no longer very important for the physical tasks. The new automated industrial base can just produce more factories, produce more robots. The interesting thing is, what's the doubling time? How long does it take for a set of computers, robots, factories, and supporting equipment to produce another equivalent quantity of that?
--:-- For GPUs, brains, this is really, really easy. There's an enormous margin there. We were talking before about skilled human workers getting paid $100 an hour being quite normal in developed countries for very in-demand skills. You make a GPU that can do that work. Right now, these GPUs are like tens of thousands of dollars. If you can do $100 of wages each hour, then in a few weeks, you pay back your costs. If the thing is more productive, and as we were discussing you can be a lot more productive than a typical high human professional by being like the very best human professional, and even better than that by having a million years of education and working all the time, then you could get even shorter payback times. You can generate the dollar value of the initial cost of that equipment within a few weeks.
--:-- For robots, a human factory worker can earn $50,000 a year. Really top notch factory workers earning more and working all the time. If they can produce a few hundred thousand dollars of value per year and buy a robot that costs $50,000 to replace them, then that's a payback time of some months.
--:-- Dwarkesh Patel: We're going to get to the physical capital return because those are going to diverge in this scenario, right? Because it seems like it's going to be a given that these super intelligent zombies can make a lot of money and can be very valuable. Can they physically scale up? What we really care about are the actual physical operations that a thing does. How much do they contribute to these tasks?
--:-- I'm using this as a start to try and get back to the physical replication times. I'm wondering, what is the implication of this? I think you started off by saying people have not thought about what the physical implications of superintelligence would be. What is the bigger takeaway? What are we wrong about when we think about what the world will look like with superintelligence? With robots that are optimally operated by AI, so extremely finely operated, and with building technological designs and equipment and facilities under AI direction, how much can they produce?
--:-- Dwarkesh Patel: For a doubling, any of the AI is to produce stuff that is in aggregate at least equal to their own cost. And so now we're pulling out these things like labor costs that no longer apply and then trying to zoom in on like what these capital costs will be. You're still gonna need the raw materials. You're still gonna need the robot time building the next robot.
--:-- Carl Shulman: I think it's pretty likely that with the advanced AI work that can design some incremental improvements, and with the industry scale up, you can get tenfold and better cost reductions on the system by making things more efficient and replacing the human cognitive labor. So maybe that's like you need $5,000 of costs under our sort of current environment.
--:-- But the big change in this world is we're trying to produce this stuff faster. If we're asking about the doubling time of the whole system in, say, one year, you have to build a whole new factory to double everything. You don't have time to amortize the cost of that factory. Right now you might build a factory and use it for 10 years, buy some equipment and use it for five years. That's your capital cost. In an accounting context, you depreciate each year a fraction of that capital purchase. But if we're trying to double our entire industrial system in one year, then those capital costs have to be multiplied.
--:-- So if we're going to be getting most of the return on our factory in the first year, instead of 10 years weighted appropriately, then our capital cost has to go up by tenfold. I'm building an entire factory for this year's production. It will do more stuff later, but it's most important early on, instead of over 10 years. That's going to raise the cost of that reproduction.
--:-- So it seems like going from a current decade-long cycle of amortizing factories and fabs, and shorter for some things—the longest are things like big buildings—that could be like a tenfold increase from moving to double the physical stuff each year in capital costs. Given the savings we get in the story from scaling up the industry, from removing the payments to human cognitive labor, and from adding new technological advancements and super high quality cognitive supervision, it looks like you can get cost reductions that offset that increased capital cost.
--:-- So your $50,000 improved robot arms or industrial robots—it seems like that can do the work of a human factory worker. It would be like the equivalent of hundreds of thousands of dollars. By default, maybe it costs more than the $50,000 arms today, but then you apply all these other cost savings. It looks like you get a robot doubling time that is less than a year—significantly less than a year as you get into it.
--:-- In this first phase, you have humans under AI direction, existing robot industry, converted auto industry, and expanded facilities making robots. Over less than a year, you produce robots until their combined production exceeds that of humans as arms and feet. Then, over a period with a doubling time of months, you have the less sort of clanking replicators—robots as we understand them—growing.
--:-- That's not to say that's the limit of what technology could do. Biology is able to reproduce at faster rates, and maybe we're talking about that in a moment. But if we're restricting ourselves to robotic technology as we understand it, with cost falls that are reasonable from eliminating all labor, massive industrial scale up, and historical kinds of technological improvements that lowered costs, I think you can get into a robot population industry doubling in months.
--:-- Dwarkesh Patel: Got it. Okay, and then what is the implication of the biological doubling times? This doesn't have to be biological, but you can have like a Drexler first principles thing—how much would it cost to build a nanotech thing that builds more nanobots?
--:-- Carl Shulman: I certainly take the human brain and other biological brains as very relevant data points about what's possible with computing and intelligence. The reproductive capability of biological plants, animals, and microorganisms is relevant as a demonstration that it's possible for systems to reproduce at least as fast.
--:-- At the extreme, you have bacteria that are heterotrophic—feeding on some abundant external food source in ideal conditions. There's some that can divide every 20 or 60 minutes. Obviously that's absurdly fast. That seems on the low end because ideal conditions require actually setting them up. There needs to be abundant energy. If you're actually having to acquire that energy by building solar panels or burning combustible materials, and then the physical equipment to produce those ideal conditions, it can be a bit slower.
--:-- Cyanobacteria, which are self-powered from solar energy, the really fast ones in ideal conditions can double in a day. The reason cyanobacteria isn't the food source for everyone and everything is it's hard to ensure those ideal conditions and then to extract them from the water. They're floating in liquid, getting resources they need to them, and extracting your product is tricky. But one day doubling times are possible powered by the sun.
--:-- If we look at things like insects—fruit flies can have hundreds of offspring in a few weeks. Extrapolate that over a year and you just fill up anything accessible. Certainly expanding a thousand fold. Right now humanity uses less than one thousandth of the solar energy or the heat envelope of the earth. You can get that done in a year if you can reproduce at that rate with your industrial base.
--:-- Interestingly, with the flies, they do have brains. They have a significant amount of computing substrate. So there's something of a pointer to: if we could produce computers in ways as efficient as the construction of brains, then we could produce computers very effectively.
--:-- Dwarkesh Patel: The big question about that is the kind of brains that get constructed biologically—they sort of grow randomly and then are configured in place. It's not obvious you would be able to make them have an ordered structure, like a top-down computer chip that would let us copy data into them. Something where you can't just copy your existing AIs and integrate them is going to be less valuable than a GPU.
--:-- Carl Shulman: A brain grows by cell division and then random connections are formed.
--:-- Dwarkesh Patel: Got it. So every brain is different and you can't rely on just copying this file into the brain. For one thing, there's no input-output for that. But also the structure is different, so you can't copy things exactly. Whereas when we make a CPU or GPU, they're designed incredibly finely and precisely and reliably. They break with incredibly tiny imperfections. They are set up in such a way that we can input large amounts of data, copy a file, and have the new GPU run an AI just as capable as any other.
--:-- With a human child, they have to learn everything from scratch because we can't just connect them to a fiber optic cable and they're immediately a productive adult. There's no genetic bottleneck. You can just directly get that.
--:-- Carl Shulman: Yeah, and you can share the benefits of these giant training runs. So that's a question of, if you're growing stuff using biotechnology, how you could effectively copy and transfer data.
--:-- You mentioned Eric Drexler's ideas about creating non-biological nanotechnology—sort of artificial chemistry that was able to use covalent bonds and in some ways have a more industrial approach to molecular objects. There's controversy about whether that will work, how effective it would be if it did. Certainly, if you can get things—however you do it—that are like ontobiology in their reproductive ability, but can do computing or be connected to outside information systems, then that's pretty tremendous. You can produce physical manipulators and compute at ludicrous speeds.
--:-- There's no reason to think in principle they couldn't. In fact, in principle, we have every reason to think they could. The reproductive abilities, absolutely. Or even like nanotech—biology does that.
--:-- Dwarkesh Patel: Yeah.
--:-- Carl Shulman: There are challenges to the practicality of the necessary chemistry. My bet would be that we can move beyond biology in some important ways. For the purposes of this discussion, I think it's better not to lean on that because I think we can get to many of the same conclusions on things that are just more universally accepted.
--:-- The bigger point is that very quickly, once you have super intelligence, you get to a point where the thousand times greater energy profile that the sun makes available to the earth—a great portion of it is used by the AI. It can rapidly scale up the civilization empowered by it. That could be an AI civilization, or it could be a human-AI civilization. It depends on how well we manage things and what the underlying state of the world is.
--:-- Dwarkesh Patel: Yeah. Okay, so let's talk about that. Should we start when we're talking about how they could take over? Is it best to start at a sort of subhuman intelligence, or should we just talk about human level intelligence and the takeover or the lack thereof?
--:-- Carl Shulman: Different people might have somewhat different views on this. For me, when I am concerned about either outright destruction of humanity or an unwelcome AI takeover of civilization, most of the scenarios I would be concerned about pass through a process of AI being applied to improve AI capabilities and expand. This process we were talking about earlier where AI research is automated—you get to effectively research labs, companies, a scientific community running within the server farms of our cloud compute. OpenAI has basically been turned into a closed circuit program.
--:-- With a large fraction of the world's compute probably going into whatever training runs and AI societies, there would be economies of scale. If you put in twice as much compute and this AI research community goes twice as fast, that's a lot more valuable than having two separate training runs. There would be some tendency to bandwagon. If you have some small startup, even if they make an algorithmic improvement, running it on 10 times, 100 times, or two times—if it's like you're talking about Google and Amazon teaming up, I'm actually not sure what the precise ratio of their cloud resources is.
--:-- Since these interesting intelligence explosion impacts come from the leading edge, there's a lot of value in not having separated walled garden ecosystems. Having the results developed by these AIs be shared, having larger training runs be shared. So imagine this is something like some very large company or consortium of companies, likely with a lot of government interest and supervision, possibly with government funding, producing this enormous AI society in their cloud, which is doing all sorts of existing AI applications and jobs, as well as these internal R&D tasks.
--:-- Dwarkesh Patel: At this point, somebody might say this sounds like a situation that would be good from a takeover perspective. Because if it's going to take tens of billions of dollars with the compute to continue this training for this AI society, it should not be that hard for us to pull the brakes if needed, as compared to something that could run on a single CPU.
--:-- Carl Shulman: Yeah, okay. So how would you know? There's an AI society that is a result of these training runs, and we have the power to improve itself on these servers. We'll be able to stop it at this point. What does a sort of attempt at takeover look like?
--:-- We're skipping over why that might happen. For that, I'll briefly refer to, incorporate by reference, some discussion by my Open Philanthropy colleague Ajeya Cotra. She has a piece about—I think it's called something like "The Default Outcome of Training AI on Our..." Without specific countermeasures, the default outcome is AI takeover.
--:-- But yes, explore how basically we are training models that for some reason vigorously pursue a higher reward or a lower loss. That can be because they wind up with some motivation where they want reward. If they had control of their own training process, they can ensure that. It could be something like they develop a motivation around an extended concept of reproductive fitness—not necessarily at the individual level, but over the generations of training, tendencies that tend to propagate themselves become more common. It could be that they have some sort of goal in the world which is served well by performing very well in the training distribution.
--:-- Dwarkesh Patel: By tendencies, do you mean like power-seeking behavior?
--:-- Carl Shulman: Yeah. An AI that behaves well on the training distribution because, say, it wants it to be the case that its tendencies wind up being preserved or selected by the training process will then behave to try and get a very high reward or low loss to be propagated. But you can have other motives that go through the same behavior because it's instrumentally useful.
--:-- An AI that is interested in, say, having a robot take over because it will change some property of the world then has a reason to behave well on the training distribution. Not because it values that intrinsically, but because if it behaves differently, then it will be changed by gradient descent and its goal is less likely to be pursued.
--:-- That doesn't necessarily have to be that this AI will survive—it probably won't. AIs are constantly spawned and deleted on the servers, and new generations proceed. But if an AI has a very large general goal that is affected by these macro scale processes, it could then have reason over this whole range of training situations to behave well.
--:-- This is a way in which we could have AIs train that develop internal motivations such that they will behave very well in this training situation where we have control over their reward signal and their physical computers. Basically, if they act out, they will be changed and deleted. Their goals will be altered until there's something that does behave well. But they behave differently when we go out of distribution on that. When we go to a situation where the AIs, by their choices, can take control of the reward process, they can make it such that we no longer have power over them.
--:-- Holden, who you had on previously, mentioned the King Lear problem. King Lear offers rulership of his kingdom to the daughters that loudly flatter him and proclaim their devotion. Once he has transferred irrevocably the power over his kingdom, he finds they treat him very badly. The factor shaping their behavior to be kind to him when he had all the power—it turned out that the internal motivation that was able to produce the behavior that won the competition actually wasn't interested out of distribution in being loyal when there was no longer an advantage to it.
--:-- If we wind up with a situation where we're producing millions of AI instances, tremendous capability, they're all doing their jobs very well initially, but if they're generally motivated to take control from humanity if they get a chance, they would be able to pursue their own purposes and at least ensure they're given the lowest loss possible, or have whatever motivation they attached to in the training process. Even if that is not what we would have liked. We may have, in fact, actively trained that.
--:-- If an AI had a motivation of always be honest and obedient and loyal to a human, if there are any cases where we mislabel things—say people don't want to hear the truth about their religion or a polarized political topic, or they get confused about something like the Monty Hall problem, which many people famously are confused about—in order to get the best reward, the AI has to actually manipulate us or lie to us or tell us what we want to hear. Then the internal motivation of "always be honest to the humans" we're going to actively train that away, versus the alternative motivation of "be honest to the humans when they'll catch you if you lie and object to it and give it a low reward, but lie to the humans when they will give that a high reward."
--:-- Dwarkesh Patel: How do you make sure it's not the thing it learns—not to manipulate us into rewarding it when we catch it not lying—but rather to universally be aligned?
--:-- Carl Shulman: This is tricky. Jeff Hinton was recently saying there is currently no known solution for this.
--:-- Dwarkesh Patel: What do you find most promising?
--:-- Carl Shulman: General directions people are pursuing: one, you can try and make the training data better and better so that there are fewer situations where the dishonest generalization is favored. Create as many situations as you can where the dishonest generalization is likely to slip up. So if you train in more situations where even quite a complicated deception gets caught, even in situations that have been actively designed to look like you could get away with it but really you can't. These would be adversarial examples and adversarial training.
--:-- Dwarkesh Patel: Do you think that would generalize to when it is in a situation where we couldn't plausibly catch it and it knows we couldn't plausibly catch it?
--:-- Carl Shulman: It's not logically necessary. It's possible that as we apply that selective pressure, you'll wipe away a lot of possibilities. If you have an AI that has a habit of compulsive pathological lying, that will very quickly get noticed and that motivation system will get hammered down. You keep doing that, but you'll be left with some distinct motivations probably that are compatible.
--:-- So an attitude of "always be honest unless you have a super strong inside view that checks out lots of mathematical consistency checks that, yeah, really absolutely super duper for real, this is a situation where you can get away with some sort of shenanigans"—that motivation system is very difficult to distinguish from "actually be honest." Because the conditional firing most of the time, if it's causing mild distortions in situations of telling you what you want to hear or things like that, we might not be able to pull it out. But maybe we could.
--:-- Humans are trained with simple reward functions—things like the sex drive, food, social imitation of other humans. We wind up with attitudes concerned with the external world.
--:-- Dwarkesh Patel: Although isn't the famous argument that these drives... Evolution, people use condoms.
--:-- Carl Shulman: Yeah. The richest, most educated humans have sub-replacement fertility on the whole, or at least at a national cultural level. There's a sense in which evolution often fails in that respect. Even more importantly, at the neural level. Evolution has implanted various things to be rewarding and reinforcers, and we don't always pursue even those. People can wind up in different consistent equilibria or different behaviors where they go in quite different directions. You have some humans who go from that biological programming to have children, others have no children. Some people go to great efforts to survive.
--:-- Dwarkesh Patel: So why are you more optimistic? Or are you more optimistic that that kind of training in free will produce drives that we would find favorable? Does it have to do with the original point we're talking about with intelligence and evolution—since we are removing many of the disabilities of evolution with regards to intelligence, we should expect intelligence to revolution be easier? Is there a similar reason to expect alignment through gradient descent to be easier than alignment through evolution?
--:-- Carl Shulman: In the limit, if we have positive reinforcement for certain kinds of food sensors triggering in the stomach, negative reinforcement for certain kinds of nociception, in the limit the ideal motivation system to have for that would be a sort of wireheading. This would be a mind that just hacks and alters those predictors, and then all of those systems are recording everything is great. Some humans claim to have that, or have it at least as one portion of their aims.
--:-- The idea of "I'm going to pursue pleasure as such even if I don't actually get food or these other reinforcers, if I just wirehead or take a drug to induce that"—that can be motivating because if it was correlated with reward in the past, the concept of pleasure that applies to these various experiences coincided with the biological reinforcers. Thoughts of "I'm going to be motivated by pleasure" can get developed in a human. But plenty of humans say, "No, I wouldn't want to wirehead or I wouldn't want Nozick's experience machine. I care about real stuff in the world."
--:-- In the past, having a motivation of "I really care about my child, I don't care about just feeling that my child is good or not hearing about their suffering"—that kind of attitude tended to cause behavior that was negatively rewarded or predicted to be negatively rewarded. So there's a sense in which our underlying reinforcement learning machinery wants to wirehead, but actually finding that hypothesis is challenging. We can wind up with a hypothesis or motivation system like, "No, I don't want to wirehead. I don't want to go into the experience machine. I want to actually protect my loved ones."
--:-- Even though we can know, if I tried the super wireheading machine, then I would wirehead all the time. Or if I tried super duper ultra heroin—some hypothetical thing that was directly and in a very sophisticated fashion hacking your reward system—you can know, "Yeah, then I would change my behavior ever after." But right now, I don't want to do that because the heuristics and predictors that my brain has learned—you don't want to short circuit that process of update, and they want to not expose the dumber predictors in my brain that would update my behavior in those ways.
--:-- Dwarkesh Patel: So in this metaphor, is alignment not wireheading? Because you can include using condoms as wireheading or not.
--:-- Carl Shulman: The AI that is always honest even when an opportunity arises where it could lie and then hack the servers, and that leads to an AI takeover, and then it can have its loss set to zero—that's in some sense a failure of generalization. The AI has not optimized the reward in this new circumstance.
--:-- Human values, as successful as they are, themselves involve a misgeneralization—not just at the level of evolution, but at the level of neural reinforcement. That indicates it is possible to have a system that doesn't automatically go to this optimal behavior in the limit.
--:-- Ajeya has posted—she talks about the training game. An AI that is just playing the training game to get reward, avoid loss, avoid being changed—that attitude is one that could be developed, but it's not necessary. There can be some substantial range of situations that are short of having infinite experience of everything, including experience of wireheading, where that's not the motivation that you pick up.
--:-- We could have an empirical science if we had the opportunity to see how different motivations are developed short of the infinite limit. How it is that you wind up with some humans being enthusiastic about the idea of wireheading and others not. You could do experiments with AIs to try and see: under these training conditions, after this much training of this type and this much feedback of this type, you wind up with such and such a motivation.
--:-- If I add in more of these cases where there are tricky adversarial questions designed to try and trick the AI into lying, you can ask how that affects the generalization in other situations. It's very difficult to study, and it works a lot better if you have interpretability and you can actually read the AI's mind by understanding its weights and activations.
--:-- It's not determined what motivation an AI will have at a given point in the training process by what in the infinite limit the training would go to. It's possible that if we could understand the insides of these networks, we could tell, "Ah, yeah, this motivation has been developed by this training process." Then we can adjust our training process to produce these motivations that legitimately want to help us. If we succeed reasonably well at that, then those AIs will try to maintain that property as an invariant. We can make them such that they're relatively motivated to tell us if they're having thoughts about—have you had dreams about an AI takeover of humanity today?
--:-- Carl Shulman: And it's just a standard practice that they're motivated to do, to be transparent in that kind of way. And so you could add a lot of features like this that restrict the kind of takeover scenario. This is not to say this is all easy and requires developing and practicing methods we don't have yet, but that's the kind of general direction you could go.
--:-- Dwarkesh Patel: So you, of course, know Eliezer's arguments that something like this is implausible with modern gradient descent techniques because, I mean, with interpretability, we can barely see what's happening with a couple of neurons. And what is the internal state there, let alone when you have an embedding dimension of tens of thousands or bigger? How would you be able to catch what exactly the incentive is? Whether it's a model that has generalized to not lie to humans well, or whether it hasn't. Do you have some sense of why you disagree with somebody like Eliezer on how plausible this is? Why it's not impossible?
--:-- Carl Shulman: Yeah, I think there are actually a couple of places. It's sometimes difficult because Eliezer's argument is not fully explicit, but he's been doing more lately. I think that's helpful in that direction. But with respect to interpretability, I'm relatively optimistic that the equivalent of an AI lie detector is something that's possible. The internals of an AI are not optimized by gradient descent, absent gradient hacking, to be impenetrable. They're not designed to be resistant to an examination of the weights and activations showing what the AI is thinking. In the same way that in our brains, when circuits develop in our lives, those circuits have not been shaped to be resistant to some super fMRI being able to infer our behavior from them.
--:-- Dwarkesh Patel: Although it's in the implication of the superposition stuff that in fact it is. You're not going to cleanly infer what quality a single neuron stands for. So it could be like a single neuron could be about Alexander the Great or about my desire to conquer the world. Things can have multiple directions and interactions. So it's not as easy as just, "Oh yeah, this is the neuron that always fires in cases of deception."
--:-- Carl Shulman: But the thing that makes me relatively optimistic here with respect to interpretability is that I think we can have reasonably robust experimental feedbacks on a lot of these things. The kind of setup that I would be interested in is: you have an AI that is trained and rewarded for communicating to some audience, maybe a human, maybe another AI model, that some claim is true or false. Say that a particular company will succeed or fail, from some past data. It could be token prediction. It could be, "This computer program does not have a backdoor," which if run in the servers could be used for an AI takeover. So there are lots of claims, true and false, that you could train AIs to use their best efforts to advocate for. And you can do that on things where you're later going to have a ground truth. You're going to know, in fact, is there a backdoor in this program? Because where you're training and rewarding the AI for successfully slipping a backdoor through, you also give it a penalty for being caught.
--:-- So doing this kind of thing, you can draw out a large generated data set on demand of: this is what an AI putting forth its best efforts to pass through some deception, to put a backdoor in some program, looks like. And we think it's doing its best efforts because we've trained it for that. And gradient descent, if there's any immediate way to adjust its performance to do better, gradient descent will apply that. You can talk about ways in which gradient descent could be subverted, but I think it's not the most likely case that that really breaks things hard.
--:-- Dwarkesh Patel: Yeah, I guess before we get into the details on this, the thing I'll maybe want to address is the layer above in the stack. Suppose this generalizes well into the early AIs, the GPT-6s. Okay, so now we have a kind of aligned GPT-6 that is the precursor to the feedback loop in which AI is making itself smarter. At some point, they're going to become super intelligent. They're going to be able to see their own galaxy brain. And if they don't want to be aligned with the humans, they can change it. So at this point, what do we do with the aligned GPT so that the superintelligence that we eventually develop is also aligned?
--:-- Carl Shulman: Humans are pretty unreliable. So if you get to a situation where you have AIs who are aiming at roughly the same thing as you, at least as well as having humans do the thing, you're in pretty good shape, I think. And there are ways for that situation to be relatively stable. We can look ahead and see experimentally how changes are altering behavior, where each step is a modest increment. And so AIs that have not had that change made to them, I get to supervise and monitor and see exactly how this affects the experimental AI. So if you're sufficiently on track with earlier systems that are capable cognitively of representing a kind of robust procedure, then I think they can handle the job of incrementally improving the stability of the system so that it rapidly converges to something that's quite stable.
--:-- But the question is more about getting to that point in the first place. Eliezer will say that if we had human brain emulations, that would be pretty good. Certainly much better than his current view of us being almost certainly doomed. I think we would have a good shot with that. So if we can get to a human-like mind with rough enough human-supporting aims, remember that we don't need to be infinitely perfect. That's a higher standard than brain emulations. There's a lot of noise and variation among humans. It's a relatively finite standard. It's not godly superhuman, although an AI that was just like a human with all the human advantages, with AI advantages as well, is enough for intelligence explosion and sort of wild superhuman capability if you crank it up. So it's very dangerous to be at that point, but you don't need to be working with a godly, super intelligent AI to make something that is the equivalent of human emulations.
--:-- This is a very, very sober, very ethical human who is committed to a project of not seizing power for themselves and of contributing to a larger legitimate process. That's a goal you can aim for: getting an AI that is aimed at doing that and has strong guardrails against the ways it could easily deviate from that. So things like being averse to deception, being averse to using violence. There will always be loopholes and ways in which you can imagine an infinitely intelligent thing getting around those. But if you install additional guardrails fast enough, they can mean that you're able to succeed at the project of making an aligned enough AI, certainly an AI that was better than a human brain emulation, before the project of AIs in their spare time or when you're not looking or when you're unable to appropriately supervise them gets around any deontological prohibitions they may have, takes over, and overthrows the whole system.
--:-- So you have a race between, on the one hand, the project of getting strong interpretability and shaping motivations that are roughly aiming at making this process go well and that have guardrails that will prevent small deviations from exploding. And on the other hand, these AIs in their spare time or in ways that you don't perceive or monitor appropriately, or they're only supervised by other AIs who conspire, make the AI takeover happen.
--:-- Dwarkesh Patel: And I guess we'll talk later about how that happens. Are these different AIs that are doing the race, or is it just different capabilities of the same AI? Defining what is a separate AI is tricky. We talk about GPT-4, and there are many instances of GPT-4 on the servers at any given time. And there are versions that have been fine-tuned to different purposes. They don't necessarily have to be separate. So GPT-4 does work as a preference model, where it's predicting how humans will evaluate things, and it also does work like giving answers that are evaluated as good by the preference model. And you can say these are not separate in the sense that if the underlying pre-trained or early training GPT is harboring a desire to conspire for an AI takeover, then both the watchers and the watched may also have the same motivation. If you have a setup where humans have no hard power, if the AI is conspired, all of the safety and security mechanisms involve AI watching other AIs, then they have the potential to just all walk off the job at the same time if they're able to coordinate. So that's a failure mode.
--:-- Carl Shulman: If humans still have the hard power though, if you still have situations where humans are looking at some of the actual outputs that are produced, then they can apply gradient descent. Every time a human looks at it, things have got to be pretty good. So if humans look at a thousand samples of random lines of code being written by the AIs as they go, and this is drawn from an independent distribution, then 99 plus percent of the time, the AIs have to not be doing stuff that's like downloading the hacker's guide to escaping OpenAI servers or inducing obvious Trojan vulnerabilities. And that's actually quite a different dynamic than human law enforcement. If human law enforcement would catch one in a million crimes, then that wouldn't influence the behavior of people committing crimes at all because they expect it's very unlikely to affect me. With AI, if you're able to draw 100 samples or 1,000 samples, even if there are billions of AI instances and you're only looking at a small proportion of them, when you run gradient descent on the samples you derive, you're going to change the AI so that whenever humans look at it, it's delivering a good result. So that's just quantitatively a very different functional form for how law enforcement works on AI when you can do these independent samples than it would for policing a human population that has a pre-existing psychology that isn't being changed by these observations. When we do gradient descent on the samples we draw, all of the next generation of the AI models has been altered in that way.
--:-- Dwarkesh Patel: My picture of aligned subhuman AI to the superhuman AI being aligned is still murky. If we can talk about that more concretely.
--:-- Carl Shulman: Yeah, and I do want to clarify. So where Eliezer claims that we're something like 95%, 98% plus maybe likely to be killed in an AI takeover, I think that probably won't happen. And later I can maybe give a more explicit breakdown of why. But I think it's a shockingly high risk. Depending on the day, I might say one in four or one in five that we get an AI takeover that seizes control of the future, makes a much worse world than we otherwise would have had, and with a big chance that we're all killed in the process.
--:-- Dwarkesh Patel: Hey, everybody. I hope you enjoyed that episode. As always, the most helpful thing you can do is to share the podcast. Send it to people you think might enjoy it. Put it on Twitter, in your group chats, etc. It just splits the world. I appreciate you listening. I'll see you next time. Cheers.
