Fei-Fei Li and Lisa Su in CoRL Conversation: The Public Sector Compute Gap and the Robotics Technology Roadmap
At CoRL 2026, Fei-Fei Li, who is about to become AMD's Chief Scientist, and AMD CEO Lisa Su appeared publicly on the same stage for the first time. Fei-Fei Li said that failing to allocate sufficient compute to the public sector is one of the biggest mistakes of this historic moment. Lisa Su responded that AMD is building compute capacity measured in gigawatts, and that leaving a portion of it for long-term research and universities is the responsibility of industry leaders. She also believes that robotics will not be solved by a single end-to-end large model alone, and that models of multiple scales must be co-optimized with software and hardware systems. The two also discussed that the value of pursuing a PhD lies in curiosity, whether AGI is near, and their advice to students: go solve the hardest problems.
Why read this
AI score 51
Fei-Fei Li and Lisa Su shared the stage at CoRL, offering their respective judgments on the public sector compute gap and whether robotics will be swept up wholesale by foundation models.
Topics
Body · AI translation
English text is being prepared.
English text is being prepared.
In November 2026, in Austin, the audience at CoRL 2026 was packed with robotics researchers.
Three chairs on stage, two guests: on the left, Li Fei-Fei, soon to become AMD's Chief Scientist; on the right, AMD CEO Lisa Su. This was the two women's first public appearance together since news broke of AMD's acquisition of World Labs, and the first time Li Fei-Fei faced the entire robotics academic community in her capacity as "AMD Chief Scientist."
The conversation did not linger on pleasantries for long. Li Fei-Fei steered the topic straight into sensitive territory: the compute "hunger" of the public sector is "one of the biggest mistakes at this most important moment in history." When universities and nonprofits cannot get GPUs, how are the next generation of researchers supposed to train? How is blue-sky, curiosity-driven research supposed to happen? She looked at Lisa Su—AMD is building compute "measured in gigawatts"—could academia really get a share of that pie?
In Lisa Su's view: "Reserving a portion specifically for long-term research and for universities is absolutely part of what we do." But she also drew boundaries: robotics cannot be cracked by a single end-to-end large model alone; the future will require compute of "all forms and all scales," with hardware, software, and systems co-optimized.
Beyond that, the conversation covered many other topics, such as: will robotics be swallowed whole by foundation models? Is AGI really just "a minute" away? In an era where even PhDs are being redefined by AI, is there still meaning in doing academia?
Li Fei-Fei, Lisa Su, and moderator Peter Stone (from left to right)
01
Peter Stone: Thank you both, Lisa and Fei-Fei, for coming to CoRL. Great talks—that video (Atlas's latest results) we're all waiting to see tomorrow; everyone's on the edge of their seats now. For the next half hour or so, we'll take questions from Slido (the on-site Q&A platform), and I'll also mix in some prepared questions.
Lisa, Fei-Fei, you're now on stage with the robot learning community. We've already heard about AMD's acquisition of World Labs (in progress), and Fei-Fei just talked about world models and the enormous opportunity ahead. Lisa, let me start with you: AMD has been investing across the entire AI stack, from data centers to edge and embedded systems.
Peter Stone: Looking ahead to robotics and physical AI, where do you see AMD's biggest opportunity? How does World Labs fit into this larger strategy?
Lisa Su: Thank you, Peter, and thank you everyone for having me. Fei-Fei invited me to "crash the conference," so I hope that's okay. Welcome to Austin.
I think this is an extremely exciting moment for AI. Looking back at all the progress we've made so far, we're actually still in the very early innings of what's possible. Physical AI, and robotics in particular, is a huge opportunity—for us, a multi-hundred-billion-dollar opportunity over the coming years. The key is that you have to truly understand physics, models, and the entire environment. So we're very excited to have Fei-Fei and the World Labs team join AMD.
Lisa Su: Besides the fact that she's a superstar—can we just say she's a superstar? (audience laughs) In our industry, hardware optimization takes many years; the more you understand the workload and the clearer you are about where it's heading, the better. The vision I've always talked about—co-optimizing hardware, software, and systems—applies especially well to robotics. Over the next 5 to 10 years, this will be an extremely important field. More importantly, for the entire AMD ecosystem, we very much want to extend AMD's open ecosystem to the whole robotics community here in this room. Thank you for giving me a few minutes.
Peter Stone: Great. We can see she's a superstar, because she was the PhD and postdoc advisor to Yuke and Roberto, who helped make this conference happen, and she has many other successful students besides.
Fei-Fei, over to you. From your perspective, how does joining AMD expand what World Labs can do? You just laid out the vision for world models—what does this combination let you build next? And what does it mean for the robotics community?
Li Fei-Fei: First of all, it's so great to be in Austin. I was telling a friend offstage that I know I'll be coming to Austin often from now on.
I've known Lisa for many years—speaking of superstars, we met about ten years ago. Thinking about everything Lisa has done for AMD and for the world of computing over the past decade, and now in this AI era, it's truly breathtaking. The relationship between modeling and computing has fundamentally changed in AI.
Now I travel all over the world, and ordinary people—not the people in this room—ask me: what exactly is AI? My answer is: AI is computing, and computing is AI. Whether it's world models or language models, frankly, we're seeing a great convergence of unimodal and multimodal models. As Lisa said, this requires co-design and co-evolution between hardware and software. This is an extremely interesting opportunity.
I also want to say to the doctoral students, master's students, and undergraduates in the audience: the way we do science is also evolving. AI is not only a problem to be solved, but also a tool. We are in the exciting era of "team science." Combining the strength of World Labs with the strength of AMD to advance computing and AI through team science is exciting for everyone.
Peter Stone: Indeed, they are intertwined. Robots are changing too, everything is changing. I said to the two of you offstage: we have a "late-breaking papers" session this time, and the contrast is extremely stark—papers submitted to the conference six months ago versus those submitted a month ago; even in such a short time, the technology is already completely different.
This question is for both of you. We see LLM foundation models working well on robots, and the recent batch of new models is stunning; we also see video generation models being used on robots. What do you think will drive robots in the future? Will it still be layered components each doing their own part, or will foundation models be able to solve the entire robotics problem end-to-end on their own?
Fei-Fei Li: We have to be intellectually honest. The results coming out of frontier labs, including some announced at this very conference, are truly breathtaking. We are in the early days of this turning point, and much remains to be explored. I think VLM/LLM-driven robots will play an increasingly large role.
But I don't think we have solved this systemic problem. There are still many problems in robotics and in physical AI, and it is not clear whether we already have fundamental solutions. So change will continue, and next year's CoRL will be very different.
For the past two weeks we have been reviewing the results of the BEHAVIOR Challenge, and seeing the speed of change is truly wonderful, and I don't think we have seen the end of it. As a researcher, I personally am on my tiptoes: which direction to bet on is somewhat nerve-wracking; but at the same time I am incredibly excited, because we are now seeing signals from multiple directions, and we have the opportunity to combine these potential approaches.
Lisa Su: The only thing I would add is this: in many vertical applications, the trend we see is that there will not be just one model, nor just one direction. For the wide variety of real-world applications, you actually need a very broad set of use cases and different types of models. Frontier LLMs will perform well in some places, but I envision that in other places, more specialized models will be more cost-effective, more power-efficient, and ultimately deliver better latency for many applications.
Peter Stone: Great. At least robots are indeed one of the factors driving demand for chips, computing, and tokens. Many students and scholars here are extremely hungry for compute, for the reasons Fei-Fei just mentioned. How does AMD view the chips and compute that robots require? Training models requires enormous compute; and once robots are deployed, they also need real-time perception, reasoning, and action, often under severe power constraints. Does that mean robots are essentially a systems problem? How do you design computing, memory, software, and the entire system together?
Lisa Su: The best thing I can say is: does anyone here need compute? (audience laughs)
We are at a point where compute is foundational, and in many ways even equivalent to intelligence. We need compute in all forms, and that is what excites me. I have worked in computing for 30 years, which makes me old compared to you. But the fact is: there has never been a better time than now, and never a more important moment than now.
Compute is needed at every layer: in the cloud, for training those ultra-large models; in enterprise and edge environments, where robots become so important; and in our physical world. My view is: there will be many sizes of models, and there must also be many sizes of compute. But fundamentally, this is a solutions problem—hardware, software, and systems optimized together to solve the problem as efficiently as possible.
Fei-Fei Li: I really like that framing. Looking at progress in robotics: it used to be slow because it was hard, and because it connects to the physical world. But the turning point we are seeing now is real. I don't have a precise prediction for the slope, but we are certainly on an upward trend. This will greatly drive demand for compute in all its forms, not just for training.
Peter Stone: Including at the edge. Let me follow up with Fei-Fei. You look at this problem from a very rare combined perspective: professor, advocate for academic computing resources, founder of a funded startup, and soon-to-be chief scientist at AMD. We said that one of the challenges is getting the compute needed to do frontier research. What do you think needs to happen to make advanced robotics development more accessible to researchers, startups, and students? What role can companies like AMD play?
Fei-Fei Li: Peter, you and I have long been colleagues and partners in pushing for public-sector access to compute. Back in those years, many of us pushed for NAIRR, the National AI Research Resource.
Let me put it bluntly: failing to adequately resource the public sector is one of the biggest mistakes of this most important moment in history.
In other words, we must provide resources for the public sector, including university researchers and nonprofit institutions, because this is the first stage of innovation. This country's innovation engine begins with blue-sky, curiosity-driven research in the public sector. We are also training students, the next generation of inventors and the workforce. And today in AI, the public sector is facing severe compute "hunger."
I think we must address this, because it is an investment in the future. We are also seeing the need for responsible AI—whether we are talking about safety or security, the public sector must play a key role in understanding AI's behavior and establishing benchmarks for it, explaining to the public and advising the government on the good, the bad, and the ugly of these models. So the public sector is absolutely critical. Yes, we need more compute.
Peter Stone: I completely agree, and clearly the whole room agrees. Lisa, what role can AMD play in this to help make it happen?
Lisa Su: I have a great idea—you should ask AMD's chief scientist. (Laughter)
First, I 100% agree with what she said. Our opportunity is this: we are building massive amounts of compute, compute measured in gigawatts. Frankly, demand far exceeds supply. But dedicating a portion of it specifically to long-term research, to universities, and ensuring the next generation of innovators can actually get access to compute is absolutely part of our job and the responsibility of industry leaders. We very much want to make sure this happens through our engagement with this community.
I wasn't entirely joking earlier—there is a reason I was so excited to invite Fei-Fei to join us as a scientist. AI has come a long way, LLMs have achieved what they have, but we are still very early. The question now is: how do we ensure all these technologies truly make the world better? Robotics is one of the places where I think we are "close but haven't really demonstrated the returns." I would be delighted—and I believe others in the industry share the same motivation—to help turn this promise into reality.
Peter Stone: We very much welcome that statement. Next, let's look at a few highly upvoted questions on Slido. Many people are curious: will AMD form a robotics team? We're assuming for now that AMD doesn't have one yet. But it seems it may be less unknown than outsiders think?
Lisa Su: We actually already have a fairly sizable physical AI/robotics research team, still very early. Look at the things in the computing stack that will go into robots: we have many embedded components—FPGAs, embedded processors, AI engines—that control many of today's robots, such as those you see in industrial settings.
But when it comes to what all of you in this room are doing, I think we are still at a very early stage. And this is exactly an important part of what Fei-Fei will do in her new role at AMD: invest heavily in robotics research and in relationships across the ecosystem. I firmly believe no single company can do everything alone. Ultimately, this is exactly why startups today are doing so much. You have industrial labs, universities, and some large companies, and what you really want is to draw on the strengths of many. In such a world, an open ecosystem, strong hardware, software that converges at different layers, and different labs each doing their own thing—this is how I think we can accelerate the pace and rhythm of robotics development.
05
Should you still get a PhD: A PhD is for curiosity
Peter Stone: Next question. Many people here are at the beginning of their careers: as PhD students, which problems are most important and most interesting? Conversely, which problems would you advise not to touch, not to use as a dissertation topic? We want to hear both sides. Also, many people at this conference have asked me, and I think everyone will be interested: what is the role of a PhD in the coming years? Is it still the right path to research?
Fei-Fei Li: Peter and I talked about this a bit offstage. This is a time of transformation. I won't lie: the three of us on stage are all a bit envious of most of you—you are so young, you were born into the AI revolution, while we entered this revolution as adults.
As for whether to pursue a PhD and what to do during it, my own answer is changing, and I'll be honest with you. I don't think a PhD is for everyone, especially in this AI era: learning tools and software engineering are changing so much that there is more than one path to becoming a researcher, an innovator, and having an impact. This doesn't rule out a PhD—a PhD is still great, I love it, and I believe Peter and Lisa love it too. But there are more paths now, and you need to do some soul-searching.
So what is a PhD for? I still fundamentally believe: a PhD is for curiosity. It is the best-protected period of your life's golden years, when you can pursue the boldest, most curious, most interesting, and hopefully most impactful questions.
One thing I firmly believe: no matter how advanced AI becomes, human intelligence remains precious—it may even be becoming scarcer—but it is precious. There are indeed massive numbers of tokens competing with our brainwaves now, but I think the civilization we are building needs human intelligence, needs our fearless curiosity. So for those who have a burning desire and curiosity, I do think a PhD is still a great path, but you must have that conviction to pursue questions in the most fearless way.
Peter Stone: I think that's really a great answer. (Leiphone)
Su Zifeng: Let me offer my perspective. When I was doing my PhD, I actually never really felt I wanted a doctorate. My parents said, "Oh, what can you do? You're too young, just stay in school." (Audience laughs)
And it turned out that was literally the best four years of my life, precisely for the reason she just mentioned: it's not about the specific problem you work on. Specific problems change, and today they change especially fast. It's about having the chance to truly become the best, or one of the best, in some field, because you have the ability to explore.
So I think if you have the opportunity, doing a PhD is a wonderful thing. Life is long, and no one can take your PhD away from you. You will always have that awareness: you once grabbed hold of a problem and solved it to the top of the field.
Su Zifeng: As for your question of "what everyone should research," I have strong views on this. There's a whole debate about AGI right now: how close are we to AGI? Sometimes I hear colleagues and people in industry say, "We're so close, just a minute away." I actually don't believe it.
The reason is: look at the amazing progress we've made in AI—it really is amazing. But how good it is depends on what problems we set for it and how we guide it. So when you treat AI as a tool, of course you should use the best tools in the world; but ultimately, you still benefit from your own expertise and your own intuition.
As I like to say, when you hire people, you can tell who has true intuition for their field and who just studies mechanically, or in this era, just uses AI to be clever. You need to find problems you're genuinely interested in and cultivate that intuition, so that no matter how the tools change, it doesn't matter, because you have deep intuition for your field. That's what lets you keep pushing the field forward, and that's hard for industry to do, because industry has no patience. I mean, none of us have patience. We want products to market faster, we want the next model out faster. And that's the difference with the university environment. Yes, I'm speaking up for you. Yes, I'm just saying it another way to the students here: pay attention, Lisa just said these were the best four years of her life—she didn't say six or seven years.
Peter Stone: So... that was well said. Actually, I want to follow up with you, Feifei. You said human intelligence is still very special and inspiring. Sometimes people ask me this question, and one of my answers is: right now AI is tied to data centers in many ways, meaning lots of compute, lots of energy, lots of data. But I often say the human brain consumes far less power and learns from far less data. So there's still something missing between how AI models are trained now and what we know is possible. So I often answer: try to do this with fewer resources, the same performance but lower power consumption—I think that's an important research direction, though it may conflict with the goal of selling more chips. But I want to know what you two think: should this be part of the research agenda for today's students?
Su Zifeng: It's absolutely very important. Going back to what was said earlier, there's no one-size-fits-all for computing. The physical world isn't one-size-fits-all, right? It varies. And having that capability is why I think we're still at a very early stage, because you're right, there are these large data centers, there are these scaling laws. But ultimately, computing will be everywhere, and it will need stronger performance, higher efficiency, longer lifespan, lower latency—all of these must be optimized.
Li Feifei: Also, I think we must be careful. Don't directly compare the 20 watts of the human brain with the enormous energy needed to train models. Evolution is expensive. Our brains are the result of billions of years of evolution, and I don't yet know how to calculate the total energy the entire animal kingdom consumed for that. So we shouldn't conflate these two factors. Moreover, I firmly believe brain science is a science—understanding the brain, recording the brain, replicating the brain helps solve brain diseases. But AI science is also a science, and it doesn't have to just replicate the brain, because the computational problems we're trying to solve—from writing code to generating simulations to empowering robots—have in some ways already surpassed human capabilities.
07
Disagreements, Anxiety, and Parting Words: Go Solve the Hardest Problems
Peter Stone: Okay, we're almost out of time. There's a question on Slido that no one has upvoted at all, but I'll still try it. Given that you two are still in your honeymoon phase, asking this might be a bit offensive: do you have any disagreements? Is there anything you...
Li Feifei: Climate. But I also love Austin. No. Okay, jokes aside, do we have any disagreements? You can see we haven't found any. We haven't argued yet. We haven't had our first... we haven't had our first argument yet.
Peter Stone: Okay, next I'll give each of you a chance for closing remarks. From these questions you can see that today's students are both anxious and excited about AI. Drawing on your own experiences, if you were in their position, how would you view this moment? What parting words do you want to leave with everyone here?
Li Feifei: Now, for many students, this is the most turbulent of times, because it is a time of transformation. You know your life is about to change, because you are in the middle of your studies, and your field is changing too. This whole thing is a turning point in technological civilization. I actually want to invite all students: in times of change, seek out the enduring values. As Lisa said, life is long. Many things will change—the problems will change, the challenges will change, the opportunities will change. But how to live the life you want to live does not depend on change itself, but on the things that do not change, those enduring values. For me personally, this journey has taught me that the enduring values are curiosity, integrity, and the fearless pursuit of truth, the search for truth, taking intellectual risks, and staying honest.
Su Zifeng: I want to say, I encourage you to be ambitious, to be very optimistic, really, because you are in the best of times. You have incredible tools, and so many opportunities before you. Many people ask me for career advice, asking what they should do. This is advice someone gave me back when I was a young engineer at IBM, about two years into the job. Someone asked, "Hey, listen, what do you want to be when you grow up?" I said I wasn't sure. At the time there was a position called IBM Fellow, which was deeply research-oriented, or IBM vice president. In my heart I thought, I probably can't be an IBM Fellow because I'm not smart enough, but maybe I could be a vice president. I decided to go ask, and then they said, "Okay, then the advice I'll give you is: always run toward the hardest problems." I didn't quite understand what that meant at the time, but that sentence has stayed with me for many years. This is also the advice I give others, because in the end, as engineers, scientists, researchers,
Su Zifeng: our job is to solve problems. To solve very interesting problems, very important problems, very exciting problems. We gain joy from it, we achieve success, and we push the world forward. So if you're going to do something, then go do the hardest problems. And have confidence that you can make a huge impact in them. I think that's why the present is so interesting, because the tools you have now can solve the most interesting problems that perhaps haven't been solved for decades. Isn't that fun? So, I think this is an extraordinary time to start a career. I hope that at some point in the future you will consider AMD. By the way, she (referring to Li Feifei) will be the hiring manager. But thank you CoRL, thank you for the invitation.
Peter Stone: Okay, thank you everyone. (Leiphone)
The top international robotics conference CoRL 2026 is about to be held. Researchers, students, and readers interested in the frontiers of AI are welcome to join our exchange group.
Here, you can meet peers, exchange research, share admissions information, scientific research collaborations, meal invitations, submissions, teaming up to attend the conference, and more, and find potential partners.
During the conference, we will also continue to bring on-site reporting updates, interpretations of guest speeches, summaries of frontier directions, and forum highlights.
See you at the venue, and see you in the group.
? Group entry portal: Scan the QR code to join the group or add WeChat szhsn610, with the note: CoRL + unit/school + name + direction.