All right, we've spent the day on hard questions about AI. Who govern it governs it? How fast it's moving and how we design it around people. So for our closing keynote, we'll turn to one of the most hopeful promises: using AI to accelerate scientific discovery itself. Jeff Dean is co-founder and CEO of Discovery Loop, a new company focused on AI for science and engineering. And before that he spent 27 years at Google where he helped build systems like MapReduce, Bigtable, and create TensorFlow, and served as chief scientist and co-lead of Gemini. His talk is titled Discovery Loop: Accelerating Science and Engineering, and afterwards he'll be joined on stage by Professor Fei-Fei for a closing conversation. Fei-Fei is the founding director and senior fellow at Stanford HAI, the inaugural Sequoia Professor of Computer Science here at Stanford, and co-founder and CEO of World Labs. She recently took on the role as special adviser on AI to President John Levinson and as the co-chair of HAI's advisory council. HAI would not be here today if it weren't for her vision and commitment to putting people at the center of AI. And that founding vision remains at the heart of what we do here at the institute. So now please join me in welcoming Jeff Dean.
All right. Wonderful. It's delightful to be here with all of you today. I'm going to talk to you about Discovery Loop. Our mission is that we are trying to build AI systems that can automatically solve important problems in machine learning, science, and engineering. And we think this will ultimately lead to more discoveries in the world. And that's what we're excited about.
So there's this common cycle across many, many disciplines of kind of the universal loop of science and engineering where you propose experiments, you implement and run those experiments, you look at the results, you examine whatever measurements you can make about them, and you decide what to do next, which often means doing another experiment.
So our approach is we want to automate these kinds of experimental loops. We think if you can drive down iteration time, the time to run one time through these loops, and you can propose and automatically run many experiments, you can learn from each evaluation so that the subsequent experiments you propose are better and are informed by the previous experiments you've run, and scale both the quality and the volume of experiments, you're going to get breakthrough results. So that's in essence what we're out to do. If you squint at this nice list of US National Academy of Engineering Grand Challenge problems, these were problems proposed as really important for us to solve in the 21st century. And we're a quarter of the way through roughly. If you squint at kind of the sub-problems that you can imagine needing to be solved for some of these areas, in particular the ones in red, we think these kinds of automated discovery loops can actually assist and help with all of these different kinds of areas.
Now, one of the things that's important to us and partly why we formed this as a separate company is we believe technology can have an amazingly positive impact in the world if we do it right, but that we all have to work to make sure that it has this positive impact. And so we established our company as a public benefit corporation, and our public mission is to develop technology that addresses important societal problems and to provide broad access to its benefits to improve people's lives.
So a little bit about the four of us. I have three amazing co-founders: Sanjay Ghemawat, Quoc Le, and Oriol Vinyals. We've worked together for, depending on the pairwise relation, somewhere between 14 and 30 years. We really enjoy working together on difficult problems, and we've got a lot of different kinds of experience. So we've helped build Google products. We all came from Google, so these are all Google-centric things we've worked on. Lots of different products. You know, we've worked together on many different generations: Google Search, the initial ad system, content ads, Gmail, News, Translate. Infrastructure systems in the second column, in particular, focus initially at the top part of that column are things that help you organize, manage large amounts of data, run large-scale computations over data to extract interesting insights. The bottom four things there are really multiple generations of large-scale machine learning infrastructure to enable training more and more capable models, and TPUs are machine learning-focused hardware that we were involved in. Then we've worked on a bunch of different AI research problems, including some things that are kind of at the core of lots of different sort of highly capable models today. Things like model distillation or mixture of experts models. Neural architecture search is kind of an early version of these kind of automated loops that are trying to identify capable, you know, neural model architectures for solving problems, and then applications of AI to different domains, including some in science. So I was, Quoc and I were both involved in AlphaChip, which is reinforcement learning for chip design and placement and routing and chip design. Oriol was involved in AlphaFold and neural weather forecasting. So good mix of experience. We've also managed pretty large teams. So I managed the Google Brain team up to 600 people and the Google Research and AI teams up to about 4,400 people before I decided I was spending a lot of time in meetings about Q2 OPEX spending. And I pressed send on an email that took me from 4,400 to four reports. And then Geoff Hinton retired the next week, so then I had three. And then Oriol and I were both then co-technical leads of the Gemini project, and Quoc was our post-training lead. We're also quite proud of many alumni of our teams who've gone off and done lots of amazing things, including founders of companies like Anthropic, OpenAI, Physical Intelligence, which is probably known here in Stanford, various other things.
And we think we have a nice blend that mixes both AI research skills and distributed systems skills. So, these are the most cited AI researchers and distributed systems researchers per Google Scholar. And there's us in the red boxes. Also, the people with stars have all worked in our teams, which is nice.
We also have fun together. So, we've traveled in many different continents. There we are in London in the upper right, Patagonia in the upper left, Vietnam in the bottom left, Quoc and I are coming home from dinner on Uber-like moped services, and California.
So, one of the things we've been focused a lot on is how do we assemble an amazing founding team? And so that's what we've been up to for kind of the first, you know, period of existence of the company is hiring amazing people. And they've come from lots of different places. You see all the logos. Many of the people on the team have experience founding various companies or various important projects. We have a lot of great team experience in, you know, leading the Go programming language project, you know, key contributors to lots of different models: Gemini, DeepMind, Kimi K3. The co-author of the Kimi K3 paper and author of the tension residuals work joined us recently. We have a great mix of kind of very experienced people and people who are rising amazing stars earlier in their career. And roughly 10% of the company has the ACM Prize in Computing and 10% have IOI medals. So, we're pretty excited about the team we've hired. We've also really focused on hiring for kind of amazing ability to work together, which was great fun. So, why now?
You know, we now have these really capable frontier models that can do lots of amazing things. And partly that came about because there's been innovation over the last say 10 to 15 years in the AI space of, you know, lots of different kinds of things. The hardware layer has improved tremendously in performance, both TPUs and GPUs and sort of more specialized things. Software abstractions have gotten better so that people can implement things they care about more easily and readily, and it can be mapped onto the hardware more efficiently. Model architectures have improved. You know, word2vec and then sequence-to-sequence and then transformers have all been important innovations in what kind of models we should train. Training algorithms have gotten a lot better. Things like distillation enable very high quality but smaller models, and inference algorithms have actually improved quite a lot. So you can actually apply more compute at inference time and get better results with things like chain of thought or, you know, inference-time compute scaling where you try many ways of solving the problem and then kind of examine what's happened.
But a great base model isn't sufficient for solving some of the problems that we want to solve. So in particular, we think you start with great base models, but then you want agents plus models to be able to really effectively use tools to accomplish really complicated tasks. We want the agents and models to be especially good at creativity. Like coming up with amazing scientific or engineering breakthroughs is not just kind of doing things that have already been done in a lot of cases. It's actually putting together lots of insights, coming up with something genuinely new, and seeing how well that works. And we also think agents and models should learn from the previous experiments and previous experience. Both in trying to solve this particular problem, but also over time you will see systems that maybe have solved 5,000 problems, and the 5,001st one they can learn from the problems that are sort of most closely related to that particular new problem they're trying to solve.
So at this traditional experimental loop you often have a lot of manual effort at each stage. And when you try to solve a larger problem that becomes even more complicated because you have to typically decompose those larger problems into a bunch of sub-problems or sub-pieces. You end up with actually a difficult decomposition problem in and of itself. That's the left-hand side. And then each one of these sub-problems or sub-areas has lots of people running experiments, figuring out what worked, what didn't work, trying different approaches. You end up with a lot of coordination overhead, and resource conflicts can occur, large system effects. So this decomposition that you have to do is actually quite hard. And it's hard to combine the results across the sub-problems you set out. And perhaps even worse, the way you've decided to decompose things in a very large-scale project is actually pretty hard to change once you've kind of got the project up and running and going. And so you can't really revisit those early decisions very effectively when you have a, you know, a large-scale engineering effort like building an airplane or training the best frontier model or whatever.
So what we think is possible is you can take a relatively small number of domain experts, use automated systems that can ideally automatically decompose into interesting sub-problems and then automatically solve those sub-problems. But it may be that human guidance helps the decomposition, and then humans sort of direct what are the important sub-problems to solve, and really enable a small number of people with domain expertise to really tackle large impactful problems. And we think this is possible if you build the right infrastructure. You know, you want to have tens of thousands of automated experiments. You want a system that can actually handle that, can understand what is the value of running this experiment versus running that one, and decide what to do next. Automated decision-making with reinforcement learning or evolutionary algorithms. These earlier decisions can then be revisited more fluidly. And every earlier experiment should inform every future experiment.
So, what have we done so far? First three weeks we did 400 interviews mostly at Cafe Venetia on University Avenue to find amazing people, which I talked about some of them. Finding office space, we're in the old Zibb restaurant, so we have a funky, funky space for us. We've built our own agent harness that we're using for all of our own development and for automation. That harness has some really nice collaboration features, so you can think of it as a system that you interact with a model to have conversations and ask it to solve things, but you can also invite your teammate in to collaborate. And so then the two of you can talk to a model and an agent. And you can also have other people add comments to the thing you're trying to do or propose a design to the agent before you propose that. We've built really agent-first infrastructure so that it will scale to having tens, hundreds of thousands of agents doing things rather than sort of hundreds of people doing things. We think that's really important, and we've got probably on the order of 20 automated improvement loops running in the way in which we imagine that we want to solve problems. And we can see that they're running and improving, evaluating methods for making these experimental iterations more effective and faster. How do you learn from the previous experimental results and the previous measurements you're able to take in order to learn faster?
So this is kind of where we are: is building this software loop to automate machine learning tasks in particular, because we think that unlocks actually a lot of autonomous science. But initially we're focused on how do you make automated systems that can solve machine learning research and machine learning engineering problems. And our goal is really to make domain experts 20x to 50x more productive. You know, we're not trying to take people out of what we're doing. We're just trying to really empower amazing people or domain experts who want to solve particular problems and enable them to solve those problems with less effort on their part.
So we think 15 years of breakthroughs in machine learning have led to very capable models setting the stage for automating ML science and engineering to dramatically accelerate discovery. Thank you. Oh, and we're hiring, apply at discoveryloop.com. Thank you.
Oh, this one. Okay. Sorry, I took your water then.
All right, that was a great talk. Thanks, Jeff. And I think this is the last session or... Okay, so we're between the audience and snacks or...
Never a good place to be.
Yeah. So I was hoping that we can be a little bit nerdier for this session and get into a little bit of the tech, but we're supposed to have a mutual discussion. So I'm going to wear the dual hat of asking you questions but participate in the discussion. Thanks for that talk because I was going to ask you why as my first question for Discovery Loop. Congratulations.
And pretty much all four of you are more or less Google lifers, especially you joined Google 1999, right? Yeah. So it is very clear what the macro incredible technological progress we have seen the past few years, but I'm still curious: was there a moment in the past, I would have guessed a year or two years, that the technology has moved into a state that you and Quoc and Sanjay and Oriol are like, we've got to do something different and we cannot do it at Google?
Yeah. Yeah, I mean I think we all have been excited about seeing automated systems that can actually sort of help people develop software systems in a more automated way. Or, you know, Quoc did a lot of work on neural architecture search even a decade ago, which you can think of as an early version of that kind of recursive sort of self-improving system where you say, please find me great model architectures for computer vision problems or, you know, language problems. And so we knew that kind of approach worked. But in that day you were sort of specifying the search space as like a janky kind of pile of Python code rather than natural language as you can today. You can sort of describe to a, you know, a capable model plus agent system, you know, please go off and explore model architectures that might help interpret MRI images or whatever, and it will go off and do that. And, you know, I think that's just a really profound shift in what we imagine computer systems and AI systems to be able to do, right? And if you can get those things working reliably, that enables a lot of stuff. And in particular for science and engineering, we think that's a really ripe set of domains for how can you really advance positive impact in the world.
I see. So it was really the loops, the self-improvement that got you guys feel that you want to...
Some might say we're a bit loopy, but yes.
It is, and also I would also say that it was important to us to, you know, we've enjoyed many times starting very small projects.
And that feeling of magic when you have five people working on something and you add a sixth and then you add a seventh and 10th and so on. That really is fun. And you can do that in large companies, but we also wanted to do this as a public benefit corporation. And we felt like that was an important reflection of our own values and, you know, taking a bigger plunge of now you have to figure out real estate and all the other things that involve setting up a company. Yeah.
Even for myself when we started World Labs it was a moment of seeing the technology, it's different from you guys, seeing computer vision and generative models have converged at the moment where we can create generative models beyond just words was the moment for us to start World Labs.
Yeah, I mean like the generative models you all produced and that are now possible are pretty amazing.
Yeah. Okay. So let's talk about this self-improvement and recursive self-improvement. Let's nerd out a little bit. We're contemporaries in terms of our education through the field of machine learning. I was just thinking in my head how many names this concept has had in its lifetime from me being a grad student. We had incremental learning, then we had continuous learning, then we had meta-learning, then we had lifelong learning and neural architecture search. You mentioned that, now we have self-improvement and then RSI. So do you want to just treat the audience, especially the technical part of the audience, with some more rigorous definition of what is self-improvement versus what is recursive self-improvement? And my goal here is actually to tease apart the fiction from the facts.
Sure. Yeah. I mean, I think if you think about those loops that I talked about, those loops can be applied to many kinds of problems. And I think of like self-improving systems as one thing where you've given the system a goal to go off and improve, you typically some measurable objective like accuracy of an ML model, and, you know, it's trying different kinds of model architectures to improve that accuracy.
So pause there. A lot of machine learning algorithms are capable of doing that.
That itself is not recursive, correct?
So like self-improving system might be...
We point our system at a bunch of MRI data and MRI outcomes and we say I want to develop a model to help interpret MRI images. Right? That would be a self-improving system where you can develop that model automatically and can then, you know, see the accuracy of the different approaches that are being tried hopefully increase as the system runs experiments, but that's not improving the base system that is proposing what to try next, right?
So recursive self-improvement is how do you actually get that kind of self-improving approach to improve the base system that you're trying to get to propose, improve, if that makes sense.
Yes, it makes sense to me. But I think it's a very confusing concept for people, and now especially we're in the Silicon Valley, people are just using this in a very loose way, and as a scientist actually it bothers me. So I just want to be on this stage to make sure we define RSI really well. So let's talk about the physical limit of RSI though, because RSI is being hyped in a way it's almost like it's almost a system that has no end, we can just accelerate and to the infinity. What is the hype here? What is the fiction from...
Well, I mean, I think if you think about how humans improve the capabilities of models, which is what has been happening over the past 10 or 15 years, you know, you get really capable researchers and ML engineers together. They run lots of experiments at small scale. They say, 'Oh, these ideas didn't work. These ideas did work.' Then we're going to incorporate those into the next generation of the model. That's how Gemini 2 comes to be after we had Gemini 1, is we ran lots of experiments and improved things. So that is kind of a normal process of research and engineering. And I think when you start to apply automation to that, yes, you can get some improvements in how fast that can happen. You know, people will start off an experiment, the experiment will finish at 2 a.m., they're not at the desk until 8 a.m. to look at the results. So if you have an automated system, it can look at the results, figure out what to do next in a much more automated way, can maybe make a lot better decisions across an entire large project than each individual making decisions about what to do next. So, but I think if you think about what needs to be involved in creating the base model and making it better, there's lots of actual training and experimentation that needs to be done, and that's not going to happen infinitely fast, right? Like there's just actual physics involved. You need to run some computations, you need to look at the results, the computers need to like do their thing, and you need to save the data, have another model look at the results in order to decide what to do next. So I think some people have this illusion that you're going to be able to run this loop infinitely fast, and I just find that kind of implausible.
And it's good to be grounded in reality. Having automated systems that can try to improve things is definitely going to improve the rate at which we can improve those systems, but it's not going to make them go infinitely fast.
I think this is a really important point. Actually, I was not applying for a job at Discovery Loop, but I was at the Cafe Venetia with Jeff.
We have coffee for many reasons because we like coffee.
And we have been talking about this RSI, you know, hype and facts. We were actually prepping for this panel, apparently. But I was thinking, okay, this is something I haven't run by you, but tell me if you agree or you can disagree with me, that's totally fine. I was thinking about humans. Are we a self-improving system or are we an RSI system? And I think we're definitely a self-improving system. I suck at cooking. And if I were to try to cook an omelette, the first time I will really burn everything, and second I would have learned, and the second time I would do it a little better. Maybe after 10 times I'll have a reasonable omelette that I will eat, my kids will not eat. And that is self-improving.
But the RSI part has to do with my brain. In this process of learning how to cook an omelette, first of all, my brain has to be somehow altered. Some neuronal connectivity needs to be altered enough that I have a better system to cook an omelette. Maybe I can grow the sixth finger if we're talking about hardware improvement. And then by cooking more...
Maybe you're AI generated.
Exactly. And then by cooking more omelettes my brain will be further improved. And if that loop happens, humans is an RSI system. Is that the right analogy?
Yeah. I mean, I think the answer is humans are both, right?
I mean, I think through your life you learn a lot of things which is self-improving, like you learn how to cook an omelette, you learn how to do calculus problems, but you also learn how you learn.
Yeah. And that learning how to learn actually is what makes you self-improving and recursively self-improving because now you can learn to do more things more quickly based on you understanding what kinds of knowledge and the ways to get information presented to you actually makes you most effectively pick it up.
And possibly recursive self-improvement is possibly an evolutionary algorithm, right? So anyway, I don't want to go down the rabbit hole, but I think it's just a rare opportunity to have this discussion in public because we have so much hype right now about...
concept. I mean, there's a lot of hype about like, oh, it's going to go infinitely fast and it's scary. I mean, I think if you look at it as a pragmatic engineering system doing things, it becomes less scary and more understandable.