Into the Impossible With Brian Keating - The Computer Expert That Just Solved AI’s TOUGHEST Challenge (ft. Rose Yu)

Episode Date: August 26, 2025

Please join my mailing list here 👉 https://briankeating.com/list to win a meteorite 💥 Could an AI physicist soon out-innovate Einstein?  In this exciting interdisciplinary exploration, UC San Diego’s Rose Yu dismantles the romantic myth of genius-driven science and instead offers a thrilling look at how AI could become not just a computational tool, but a genuine partner in discovery. Rose draws from her pioneering work in traffic forecasting, pandemic modeling, and high-energy physics to show how machine learning is evolving and what it can offer to the scientific community. She walks us through the fascinating ways AI is already helping us rethink what’s possible in science: from generating novel hypotheses to rediscovering fundamental symmetries, and even beating traditional simulations in both speed and precision. With her characteristic clarity, Rose explains how she’s built data-driven surrogates for complex phenomena like the spread of epidemics and the turbulence of fluids. We also tackle some of the more existential questions: Can an AI ever experience “happy thoughts”? How close are we to AGI? And should we be worried about AI replacing us?  If science is humanity’s greatest invention, this episode is a fascinating look at its future. — Key Takeaways:  00:00 Intro  01:04 Can an AI physicist out-innovate Einstein?  02:07 Why are GPUs so good for AI? 12:24 Traffic modeling and AI 16:08 Is AI just imitating physics? 24:53 Epidemiology modeling and AI  30:49 Should we be worried about AGI? 34:32 The benefits and dangers of AI 39:53 AI scientists and scientific productivity  42:19 The impact of AI on academia and education  52:34 Outro  — Additional resources:  💻 Rose Yu’s website: https://roseyu.com/  — ➡️ Follow me on your fav platforms: ✖️ Twitter: https://twitter.com/DrBrianKeating  🔔 YouTube: https://www.youtube.com/DrBrianKeating?sub_confirmation=1  📝 Join my mailing list: https://briankeating.com/list  ✍️ Check out my blog: https://briankeating.com/cosmic-musings/  🎙️ Follow my podcast: https://briankeating.com/podcast  — Into the Impossible with Brian Keating is a podcast dedicated to all those who want to explore the universe within and beyond the known. Make sure to follow/subscribe so you never miss an episode! Learn more about your ad choices. Visit megaphone.fm/adchoices

Study and play. Come together on a Windows 11 PC. And for a limited time, college students get the best of both worlds. Get the Unreal College deal, everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft 365 premium

and a year of Xbox GamePass Ultimate with a custom color Xbox wireless controller. Learn more at Windows.com slash student offer. While supplies last, ends June 30th, turns at AKA.m.m.S. College PC. Yamava Resort and Casino at San Manuel is California's number one entertainment destination for today's superstars. Catch the Jonas Brothers return to the Yamava Theater stage on April 30th,

the powerful vocals of Demi Lovato on May 17th, and the signature Southern Country Rock of Eric Church on July 19th. Tickets on sale now at Yamavat Theater.com, only at Yamava Resort and Casino, celebrating its 40th anniversary. You in? Must be 21 to enter. This professor says AI will never feel shivers down at spline. AI systems can now automatically discover fundamental symmetries and new physical laws without being taught the underlying theories. Essentially, they're finding the hidden mathematical patterns that govern our universe through pure data analysis.

If machines can independently discover new laws like Lorentz and variants from particle physics data, without knowing Einstein's theories, it suggests AI might find entirely new physical principles we've never, ever, conceived of. Professor Rose-U's team trained deep learning models on data from the large Hadron Collider that automatically recognize symmetry patterns and high-energy particle interactions. The same symmetries that took Einstein and other geniuses decades to understand through pure theoretical insight. Professor Rose-U is a computational physicist at UC San Diego, whose AI models have been deployed by Google Maps for Traffic Predictions and ranked number one among 40 national teams for pandemic forecasting during COVID-19. Now let's meet

this brilliant natural genius who's taking artificial intelligence to the next level. Let's go. Professor Rose, you. So nice to have you here at UCSD's Arthur C. Clark Center for imagination into the Impossible Podcast. Great to meet you. It's a pleasure to be here. You've done so much wonderful stuff and it's wonderful and gratifying to know that you're a colleague here at UCSD, but you're also getting the recognition that you deserve. I want to start off with a little bit of a provocative question, which is fascinating to me, which is the following question. Can an AI physicist ever do what Albert Einstein did. So famously in 1907, Albert Einstein said he had a dream, a thought experiment, that if he was in free fall, like if he was like this and he fell down, that he would feel

no gravitational force field. So you could actually get rid of gravity. And that's part of what's called the Einstein equivalence principle. I want to ask you, he called that notion, that thought, the happiest thought that he ever had. He said it made shivers go down his spine. Can an artificial intelligence ever have a happy thought? Can it ever feel shivers down its spine or its CPU or GPU? And can it really create new laws of physics? That's a great question. So I think in general, AI is a bunch of machines that can think. So if you ask whether these machines can have emotions, I'll say most likely no, just by how they're constructed. I don't know whether it's possible even to build machines that have emotions.

However, the second question is whether these machines can create. And I think my answer is definitely yes. Because these models are built with a lot of data, and they're still with the world's knowledge. And then once they are imbune with knowledge, they can start creating new content. And you're already seeing these examples now with large language models. They create new mathematics theorems. They can also generate new high. processes. Recently, they have shown these models can create new molecules. So as we are seeing

more and more examples, it just vindicated that it's definitely possible for these models to create new theories like what Einstein has come up with. Obviously, everyone out there is familiar with LLMs, like chat GPT, Claude. I use them all. I pay for them all. There's probably a bunch of paying for that I don't even know about. My credit card gets deducted. I use them for everything. They read my kids' bedtime story. Originally, I would, was kind of embarrassed and I was ashamed. I'm using a AI to read a bedtime story or come up with the bedtime story and then it reads to that. But I mean, a book is kind of like that. It's just a natural intelligence and I'm just reading or listening to an audiobook. So I've gotten over my

guilt on that. But there seems to be something different about physics and that's the only reason I'll bring that up is because that's what I do. I'm a physicist in that it's an empirical science. Mathematics I could understand and I might have predicted that you could make new mathematical theorems and test things because proof is possible in mathematics, but it's not possible in physics. I can't prove the earth is round, you know, for example. It's not possible to prove that statement. You can exclude and falsify other statements. When I think about LLMs, I think about, you know, these original devices that were made on GPU systems. So we can talk about the computer architecture later on. But as I understand it, you know, the GPUs were invented to, you know, make Grand Theft Auto 6 and Minecraft

and run really fast and optimize for, you know, my, my, my, my, you know, my, you know, my, kids to kill their friends at one millisecond before their friends kill them on the game. They weren't designed for, certainly for AI, that just turned out that they were good for it. Why are GPUs so good for it, good for AI, and are they good for non-LLM-type AI systems? So the GPUs are particularly good because they allocate a lot of those resources to arithmetic computing, right? Because in general, when we talk about computer chips, they're mainly designed for two types of tasks. One that is responsible for arithmetic computing, like doing matrix multiplication, and the other one is handling all this logistical instruction processing

and then handling the scheduling of different jobs. So for AI, I think when I talk about AI, it's more naturally referred to the set of algorithms called deep learning, right? And these deep learning algorithms are essentially a lot of matrix multiplication and matrix computation. So by leveraging in GPUs that was specifically designed to put a lot more weight on processing, matrix, multiplication operations is much more efficient. Usually, like 10 times, if not 100 times more efficient. Of course, like these type of algorithms are not only limited to applications in LLMs. In my class, in deep learning, we talk a lot about other applications in deep learning.

For example, generate videos with diffusion models, generate trajectories for autonomous driving cars, or, you know, generating music even. And sometimes in our own research work, we use it to generate new symmetries, all the universe, and new molecules in chemistry. So, yeah, they're all based on the algorithm of deep learning. At some level, the GPU structure is naturally useful for doing linear algebra, matrix multiplication and doing all sorts of inversions and so forth at extremely high speed. And that's because, you know, computer screens originally aren't, you know,

or just two by two matrices and some level with different third dimension representations, brightness, color, saturation. But physics is different, right? Physics is continuous. So my example that I keep going back to is Einstein in 1907 knew that the planet Mercury had this weird behavior, that it was orbiting the sun in an ellipse,

which is normal. But the place at closest approach to the ellipse, the line connecting what's called the perihelian to the sun, was moving a little bit each year, which doesn't happen for any of the other planets. It only happens in a strong gravitational field. And he knew this, and Newton even knew this, or people shortly after Newton knew this. And the only way to correctly account for that is to include the fact that gravity affects time as well as curved space.

So if you just put a bowling ball on a tarp or some trampoline, it'll bend it, right? But you won't get the procession of the percellion of the planet or the orbit of the marble around the bowling ball. Unless you include the curvature of time, Einstein realized that and made space time this 40s, dimensional thing. When I tried to do this with my student Evan Wont over the summer, or last summer, and your student helped out. I forget your student's name. Thank you. It helped the very, very graciously of you. But we couldn't get it to work. We had to put in artificial, now maybe you could get it to work, you know, with your more advanced knowledge and understanding than I have. But we had to put in, like, putting charge on

mercury and put in all these cluges because we fundamentally had to discretized space time into quanta of boxels. We couldn't do that without, we couldn't solve that without doing so. Have you seen or encountered an example where an AI will create, you know, kind of that new leap that you see, well, no, no, you could discretize it in time all you want or in space all you want, but until you include time curvature, nothing happens? Are there truly new things that you couldn't understand, not beating humans in chess or folding a protein faster than we can understand? Are there new, like, creations, like four-dimensional, five-dimensional, you know, is it going to come up a string theory? How can it with GPU limitation do something that's

just not just faster, but truly novel? Can it really do something? I think we need to like kind of a disentangle between like the underlying, you know, algorithm or underlying hardware infrastructure with the type of models that are basically used for creating new things. So there are different kind of concept because we have GPUs, right, that was originally designed for video gaming. And these are sort of hardware that are particularly good.

with arithmetic heavy operations in computation. And then you use GPUs, you can say, okay, I will use that to design a set of algorithms, the models, in particularly generative AI, that can generate new content. And that's what we call creation, right? So in some of our work, you know, we have shown you can build generative models

that can generate new type of hypothesis in science, especially in physical science. One of the key example we showed in our work is just by looking at, you know, high-energy particle physics data from large hydrogen collider, the model can automatically recognize there is laurent symmetry from data, without knowing the knowledge of, you know, general relativity from Einstein. So that's something that the model can do, right, and it doesn't necessarily require discretization of space and time. And then I would say in general, the scientific discovery goes through this kind of iterative.

loop where you have some observation and then that's your data set. Once you have the observation, you will create a new hypothesis. And then a good scientist will try to, you know, shrink down these hypotheses to a few of these possibilities. And then like the example, you mentioned in trying to find another dimension of time and to better fit the data, right? And then once you have a better fitted hypothesis, you can verify them either through theory or through experimentation.

And then using this kind of new observation of the experimentation, you can refine your hypothesis and then, you know, go back to this iterative loop. And I think it's throughout this iterative process, like AI can play a role in every single step. Because AI is fundamentally designed to process, huge amount of data, generate a large number of hypotheses. It can help scientists sift through, like, all these possibilities really quickly, given this massive power to process, you know, internet knowledge. And then you can also create design experiments to help verify them. I will say the current limitation is that a scientist can go to the lab and then do the experiment manually, right? Sometimes they have help with the robots.

But AI algorithms are largely still limited to computers. And if they want to go out and verify the hypothesis they have, they still have to rely on humans to do this experimentation and collect the data. But a lot of these processes in scientific discovery are already been accelerated by AI. Perhaps, yeah, the understanding of the fundamental symmetries and underlie physical law, which Einstein and Eme Noether and many other people contributed to have to do with conservation laws that underlie physical principles, including independence of position and time and so forth. That's really exciting.

I wasn't familiar with that, so I'd love to learn more about the Lawrence and Variance Violation from CERN. But on a more practical level, people talk about, well, what is A. I good for, you know, making new chatbots or whatever. And everyone's played around with that or making some, you know, logos or, you know, music. My daughter's made a song up about herself and her friends and what they do and stuff. And that was kind of cute. And it's brilliant.

And they sound, you know, they're listenable. They're not horrible. And now with Google V-O-3 that came out very recently, you know, people are having fun and getting scared about what it will mean for, but I feel good because the, this format where you and I are talking and we're not AI's, you know, we can, we can, we can vouch for each other at least. But a normal like talking head video or you don't see the person, these are just computer generated podcasts. Those are, there's many of those, like a notebook L.M came out with this audio format thing. But one of the most practical implications, applications of your research,

how to do with traffic modeling. And this you came up when you were a postdoc at Caltech, a little known technical college north of San Diego. And that's because, you know, everyone wants to get out of L.A. and come down to San Diego. Talk about the traffic. and the applications of AI that you invented as a postdoc that were later, as I understand, taken over by Google and now used part of Google Maps or will be? Yeah, so talk about the traffic modeling. Sure. So the idea was also, you know, there's a lot of data collected by, you know, road network sensors, right? These sensors are recording the traffic volume and the speed at very high frequency.

So before, you know, this data would just like kind of looked at like with simple models and people will use very simplified assumptions to, to, to, you know, to try to capture the pattern in the traffic. And then with the deep learning rising, I think around 2012, right? And then we saw like, okay, what about we look at deep learning models for traffic forecasting. So for traffic forecasting in general, it's not a typical image data because the sensors are distributed as a non-Euclidean space on the graph. And then you cannot use any technique that were designed for videos or images directly apply them to traffic forecasting.

So we have to come up with our own deep learning model. In particular, we were inspired by this paper from the physicist, and he described how traffic flows related to fluid flows, kind of lattice. And he described the Navistoc's equation, based on Navistoc's equation, to predict the traffic volume. But that was also based on these kind of assumptions that you have. But this connection he pointed out in the paper kind of inspired us. And then we actually designed a deep network called diffusion convolutional neural networks.

So specifically, we put this concept of fluid diffusion and to model traffic flows on sensor network. And that model was very successful because before deep learning, people were only able to forecast accurately for about 10 to 15 minutes. Then after deep learning was taking over and we were able to train this more complex models with very little assumptions, right, with a large amount of real-time sensing data. So then we were able to accurately forecast the road network traffic for up to one hour. And that's why it was a big deal and it was deployed by Google Maps. So does it take into account statistics? Like on average, you know, 10 people per hour will be using their cell phone to text and it'll crash in?

Or is it actually not using statistical averages or past behavior in an ensemble of past histories and just statistically forecasting traffic, it actually knows what the traffic will be, more or less. I mean, there's some variables like a meteor could hit LA or something. But what levels of actual simulation versus prediction are occurring with these types of convolutional neural network? Right. So these type of diffusion convolution networks takes the historical traffic patterns.

Like, for example, the velocity, the speed, the volume of the traffic for different locations. And we also know that have data about, you know, latitude, longitude information. the traffic asset and information from police reports, right? All these data are fed into the model. And then the model implicitly learned the statistics from the traffic pattern and trying to project how the traffic pattern is going to change in the future. So because it's generating the future traffic pattern, so you can also think of as a simulator, right, for future traffic.

But it's not like a simulator based on differential equations that people have done before. You rather it's a simulator trained with data. So sometimes I call this type of. model data-driven simulator. Data-driven semantics. So when we see things in the physics space that now some models probably you were involved with

can accurately simulate what, you know, smoke looks like or water in a turbulent medium, are they actually simulating the individual particle trajectories? Or are they just like representative like, oh, a smoke will rise vertically and will have this kind of

density? Or is it actually solving physics or is it really just imitating what physics looks like and you could do it with a good artist. Yeah, that's a very good question. So actually when you try to simulate the fluid, right, there are two type of representation for fluids. One is called the Eulerian representation where you basically just like sit at a particular location and watch how the flow is going to change.

So that you don't need particle-based representation. The other one is based on the Lagrangian formulation where you actually think of think of fluid as a collection of particles and you trace the movement of individual particles and integrate them, right? that give you the overall movement of the fluid flow. So actually for both representations of fluid, there has been AI models that were designed to forecast or to simulate the movement. And I have seen success for both cases.

In terms of whether they are actually solving the equation, I think most of the time they're not solving equation in the traditional sense. Because when you solve a equation with numerical methods like finite element or finite difference, and that's what kind of consider the traditional solving the equation. But these models, they're based on deep networks, right? So the fundamental algorithm is back propagation. They don't solve the equation in terms of finite difference or finite element, but they give you a solution, which evolves over space and time.

So using that solution, you can predict the future evolution of the fluid, and then it's there much faster than traditional numerical methods. What are the biggest limitations for what you do? What gives both the largest systematic errors and the largest bottleneck maybe to what you do? Is it a GPU count and you just need bigger computers, faster computers, or is it some other rate limiting process that restricts how much progress you can make in a given application at a given moment? For now, we have a project called the multimodal.

When you need to build up your team to handle the growing chaos at work, use Indeed-sponsor jobs. It gives your job posts the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this shelf will get a $75-sponsored job credit at Indeed.com slash podcast. That's Indeed.com slash podcast. Terms and conditions apply.

Need a hiring hero? This is a job for Indeed sponsored jobs. Foundation model for automatic hypothesis generation. What's the acronym for that? Let me say if I can figure. But we have an acronym called Genie, Generative Hypothesis, Generative Hypothesis model. But I forgot what the rest of the letters represent. But it's a genie.

And then what we hope it can do is can generate hypothesis like a scientist. And then we wanted to synthesize multimodal information from both the textbooks, research articles, and large-scale numerical simulations in climate. One of the bottlenecks that we're facing is that a lot of the existing foundation models like chaty-b-t or cloud, they just fundamentally cannot handle large-scale high-dimensional simulations. So we have to come up with our own model, but then again, we were also limited by the data, so generating this type of data to pre-training a foundation model extremely expensive. And then there, you know, we weren't able to train really big models, like 70B models.

We were only able to fine-tune a 7 billion model. With 7 billion models, still we need a lot of data to train, and then this data have to be curated in a way that makes sense, right? Because we want to match the textual information. from research articles and textbooks with a simulation data. Is that what I think about supervised learning? So you're like pre-trained, like, you know, self-supervised pre-training. But for self-supervised pre-training, you need to like carefully cure the data set.

Otherwise, you're just adding noise to the model, right? And then the other challenge is how do you even define the evaluation metrics? Because in, in general, when you train like an LM, right, people will use, you know, actually users as a labeler and what they call the reinforcement learning with human feedback. And based on human feedback, they iteratively improve the model. But then for our model, because we have to rely on our domain scientists, in this case, the climate scientist. We have one climate scientist in a team and his student. It's definitely not very scalable if we just rely on our user to give our feedback. So we need to come up with a much more scalable way to collect the supervision and to evaluate the model, right?

To tell us whether the model is generated right hypothesis or the wrong hypothesis. So the types of data that you're talking about here, are they time ordered? know, time series data? Are they weather, you know, images from satellites? Obviously, you talked about traffic data. Let's go back to the traffic because that's easier for me to understand. It's sort of one-dimensional in some ways, but it's not in Euclidean in other ways. So what would be the kind of rate limit or the bottleneck there? Is it that, you know, some of the data, as you say, is noisy. By the way, I found out the root of the word in Latin for noisy, or noise comes from nausea, like to be sick, which is kind of interesting because I'm sure you get

sick when you see bad data, noisy data. Talk about what does the process look like. Is it some person in Nigeria looking at, you know, LA traffic on Interstate 5 one day and then someone in, you know, Bosnia looks at whatever. What does its supervision look like? What's the input to the model? You said it's police and this and that. But there has to be some human element that curates, if it's human reinforced, non-self-supervised, I guess, supervised. So this, actually this problem, This project was, I think, it was 2018, so it's like seven years ago. And at the time, I think, it was built before the LLM era, right? And then the data that we used was this, like, loop sensor that was already installed in, like, every highway in California.

And there's a particular system that was built many years ago, trying to just collect all the loop-tonsored data. It was time series data. The time series data. Yeah, it's a velocity volume. and then it gets sent over the internet to some central traffic bureau, and then the system collect the data. You can download them. Anybody can download them.

And then how do you convert that condition into a trained data set for your purposes? Oh, yeah. So for ours, you know, once our task, we define the task at the forecasting, right, which your input is, you know, historical features of the traffic. It could be like, you know, historical volume, speed, and potential, like, accident patterns, and your target, your prediction goal is the future traffic. And in our case, we were looking for like 15 minutes as an interval, but predicting roll out for one hour, right?

So that's once we define the prediction test, then we can essentially extract subsequences from the time series to curate the data set. Are you still involved in that project, or is that that's, you did it as a postdoc, but are you still involved at all? Oh, not anymore, yeah. Actually, I spent like a year and a half at Google and trying to look scale up, for this project and it was done. So when we look at the apps now,

I can put in the app and say, you know, how close can I cut it to, you know, go and meet my wife somewhere? And it will say like, well, if you leave now, it'll be this much time to get there. But then it actually allows me to put in, no, I have to be there at 5.30. So when should I leave?

It's a different question. Is this type of research that you applied? Is that now being used in these apps? And actually for forecasting? Because I wasn't aware about the features they developed after our forecasting model, right? Because our model was a forecasting model. It tells you how the traffic is going to evolve.

Of course, like the situation, you mentioned, use cases of, like, you know, ETAs or, you know, planning. And they're all dependent on forecasting. So I'm sure that Google and other companies have built on top of this technology, but I would not involve. What are the alternatives to GPUs? Are there other types of hardware technologies that could be useful for any type of AIML, you know, whatever deep learning? what are the alternatives to GPU-based, GPU plus LLM or chat in video, as I think about it.

Are there alternatives to GPU on the hardware side? As far as I know, I think, like, TPU is one option, right, tensor processing unit. And then there's also these essentially onboard computers, like FPGAs or, you know, adjacent, and these are specifically designed for edge devices. But I know there's also people thinking about quantum computers, Essentially trying to look at beyond the classic computers and think about using a completely different paradigm to compute and that thing that also have implications on a machine learning.

How would you use a quantum if I gave you a quantum computer, Willow, whatever IBM or Google has? I mean, how would you use it right now? Would it be of use to you or not yet? Or the algorithm's not? Because we're the users of computers, right? I imagine if the quantum computer is becoming a reality, then, you know, the quantum computer is becoming a reality, then And yes, we're going to have a massive acceleration in how we can calculate things that we care about now. Okay. The next question I have is involving what we just went through, and I can't say the name of this particular pandemic,

because YouTube's AI is so sophisticated that if I say the words, you know, C, whatever ends with a D, it will put a warning label on the video, you know, that the United Health, World Health Organization, anyway, it kind of like does something and influences, and they're very, sophisticated. They look at every frame of this video. They listen to every, you know, a bite of data in the audio and they can tell what we're talking about. But talk about, you know, for general epidemiological studies that you were involved with for a recent event that took over the world, you know, five years ago from now, earlier than now in 2020, talk about what your work involved in and how AI was useful in determining both forecast and

predictions and even implications and getting stuff prepared, you know, perhaps for the next occurrence because we're just a matter of time away from the next event happening. So what did you do for the previous, you know, concern that the planet had this worldwide issue? But also, what can AI help us in your work in particular, help us to avoid or prevent or minimize the effect of the next one? Yeah. So like basically in epidemiology modeling, they have very similar issues as traffic forecasting, right? Basically, you have this like, It's peak pollination season, and my business is scaling fast. To keep the nectar flowing, I need a phone plan with top priority data speed.

That's why I chose GoogleFi Wireless. My connections stay strong even when the hive is buzzing. Plus, unlimited plans start at $35 a month. Now that's a deal that doesn't stay. Explore GoogleFi Wireless plans today. Plus taxes and government fees. GoogleFi Wireless is not subject to data traffic deprioritization during times of high network usage. Very complex numerical simulator. In this case, it's an agent-based multi-agent simulator that is also stochastic.

So we need to solve this complex numerical problem and typically it takes about a week to generate a what-if scenario. Like what if we change the school lockdown policy, which, what if the virus characteristics changes? And you want to understand the potential impact of these changes. So to get that result in a week, it's just too slow, right? during the wartime of epidemic modeling and control. So we designed physics-guided deep learning method that are hybrid of using deep learning inspired by these principles in physics to forecast

the progression of pandemic up to four weeks so that we were actually ranked number one among 40 national teams in that competition. And then, you know, another thing we contributed is to build a very fast surrogate model. So this surrogate model is, is also based on deep learning. So it's taking data from both the simulator that we have

and also the report data from CDC and slash Sean Hopkins Physics Lab. So then our simulator or our emulator and was able to mimic the behavior of simulator, not only in terms of the average prediction, but also the confidence intervals of these predictions. So this confidence interval are particularly important for risk assessment and the decision making. And then our model was able to reduce the turnaround

time originally for one week to one day. How much commonality? I mean, if you ask the normal person, you know, does the traffic in LA, you have anything to do with, you know, forecasting and pandemic? I don't think the average person would see a connection. But you do a lot of what's known as cross-disciplinary and interdisciplinary research with a variety of people ranging from epidemiologists to climate scientists to physicists and at both the theoretical end of the application observation layer.

Can you talk about, you know, what motivates you? drives you? How did you get, you know, to a point where you're as conversion is talking about traffic as virus? And how can our listeners who are brilliant, you know, how can they sort of benefit or what lessons can they learn from your career that they could apply in their lives to see connections between different fields that, to me, seem completely foreign and unrelated to each other. But to you, obviously they had some relationships. Yeah, I think sometimes, I think the biggest drive for me is just to make an impact to this world. And, you know, when I started working on traffic problem just because I was stuck in LA all the time in the traffic. And I was actually

commuting between downtown LA and Pasadena. So I was really frustrated by the situation. I thought I should do something to improve the status quo. And it just happened that I reached out to relevant people, got the data. And then, you know, as a computer scientist, we are trained to make obstruction about the problem and trying to find solutions. to this abstracted problem. And then once you have the data, you have the model, you have the right abstraction, then you can easily come up, not easily,

but you can come up with the solution. And similarly for the pandemic, right? And I was actually stuck at home during the pandemic, and then I was moving to, supposed to move to San Diego, but I couldn't for a year because of the pandemic. So I also wanted to solve this problem to make an impact, to not myself,

not only, you know, myself, and people around me and the community. So I just, you know, I have collaborators who worked on the space and I reached out to them. And then, you know, I worked with them closely. So I think the commonality I have seen from this past collaboration is one, find a problem that you feel you are, you have the eager, you know, how you're urged to you to solve them. Like you feel like you, you wanted to solve them and it's so painful to you. And I'm sure it's painful to other people as well. And second, reach to the domain expert and try to understand what is the pain point.

And then especially to understand how your expertise in a particular field can help address their pain point. And then after you understand the pain point, right, and then you can formulate a problem, and then take the right abstraction, and then you will think about how to solve them. Do you think we've passed the Turing test? I think in a lot of the benchmarks that they have shown yes. How far away do you think AGI or ASI artificial superintelligence is, where ARIs are just don't need us anymore to do anything? As I understand it, they'll just kind of iteratively improve forever without much human input.

I think that's a very hard question. As a computer scientist and especially somebody works a lot on forecasting, and I know most of the forecasts are very uncertain, so I don't want to make a certain forecast about when we're going to have an AGI. Yeah, actually, I have written a paper about the future of trend of AI scientists. And then, like, we're trying, we actually collected a lot of the research papers in our archive and trying to understand how papers are being formed and how the ideas is evolving. How quickly ideas are evolving.

Oh, wow. And so that's like a forecast of future trajectories of science, but then also has implications on AI. But I do see, like, right now, it is very astonishing. like how quickly AI is progressing, like every single day. Like even for me working in the field for more than a decade, I have been quite impressed by the progress and the speed. But in terms of whether AI is going to replace humans, and I don't think so because eventually the thing,

the goal is to have a partnering situation, right? Like AI will work side by side with a human, and you know, they provide the type of efficient, and diagnosis, knowledge, comprehension, and just to accelerate the productivity. Yeah, and be an augment to our creativity, which is somewhat unique to humans, I think it is. So as you know, this podcast is named after Sir Arthur C. Clark's famous statement that the only way of knowing the limits of the possible is to go beyond them into the impossible. So that was Sir Arthur C. Clark. So we'll get to zoom in my producer, Carlo, zoom in as close as we can. So you are actually the recipient of the very first into the impossible medal of excellence,

which has on the front Sir Arthur C. Clark's picture on the back it has a monolith. And the reason for that, I don't know if you knew that, but the word podcast comes from Arthur C. Clark because it comes from the movie 2001 of Space Odyssey when Dave is talking to how the artificial intelligence that was thought to be super intelligent back in 1968 when the movie came out or around then, the book was written. He asked Hal, open the pod bay doors. And the pod then became the iPod, Steve Jobs liked that name. One of his engineers came over there. And now, so we wouldn't have the word podcast without AI and without Arthur C. Clark. Oh, that's a beautiful. Sorry. So yeah. So on the back, it says the catchphrasing, the only way of determining limits of the possible or to the impossible. And then on the rim, you know, all coins have three sides. People think coins have

two sides. But they have three sides. The coin could land like that. Actually, this one's pretty good. Alvin. Shout to my undergraduate assistant Alvin. And it has your name. and the date of the podcast. Thank you so much. Thank you so much. It's not a Nobel Prize, but you know, it's as close as I can get given all my writing about it. But when we think about AI, the kind of the trope in the popular culture is that it would be dangerous, malicious, super intelligence. I talk with Nick Bostrom and coined that term, you know, super intelligence. The notion that it's going to be running away and then turning us all into paper clips or maximize, you know, paper clips, you're shaking your head. So I guess you might be a skeptic of that.

malicious, malevolent AI, should we have guardrails on it or should it be open? We don't know what ChatTPT really does under the hood. We know a little bit more about Lama and the open source ones. Which ones will win and which ones are more dangerous and what should we be worried about? I think these worries or concerns are definitely valid. And I think there's a lot of research on AI safety. You said this place was steps from the water. We just haven't found the steps yet.

How much did we save? Enough. Enough to get lost. Or you could book a stay with Hilton. Welcome to your oceanfront room. Just steps from the water. The Hilton sale is on now. Book on Hilton.com or the Hilton app

and save up to 20% to get the stay you expected. When you want savings, not surprises. It matters where you stay. Hilton for the stay. And it is a big deal, right? Any technology has both sides. Like even in the old times, like the new. technology. It has both sides. So, so AI and obviously as a technology, as a very powerful

emerging technology that has also, he says, positive and negative side. So yeah, so we definitely should have Garbriols. And then one of the, you know, kind of focus in our group is also trying to build these trustworthy AI scientists, right? Because a lot of times when we produce tools for our scientists and these scientists don't trust them, just because the models are predicting things that don't satisfy loss of conservation. And these are relatively easy to check. So using our first principle knowledge as Godrail for AI scientists is something relatively easy.

But then there are other situations that are much harder to put guard rails on these technologies. With that said, I would say, still, it should not stop us from progressing the technology. What I'm worried about is people are too terrified by the negative. impact, the potential negative impact of AI, and then it's actually like kind of slowing down the necessary technology for us to programs and eventually overcome this negative impact. So I think I'm a big advocate for let's try to disentangle or separate the societal impact with the technological advances and let the people who are passionate about the technological advances work on the technology, right?

And then let's separate teams who have right expertise on this to design the guard rails. And obviously, there should be communication between teams, but then let's not stop the development of AI. Try to have on all the different perspectives from Max Tagmark to Jan LeCoon,

who's on in December. I'll put links to those episodes up here in the video. They couldn't be more diametrically opposed. The AI safety, you know, pure safety, safety at all costs. And once we release this, it's like, you know, creating our own destruction. You seem uniquely positioned in that you've worked on the things that connect humanity across a global scale. There aren't that many of them. Pandemic, something propagated, you know,

virally through the atmosphere. Nuclear weapons, obviously, they affect things in the very thin shell, the atmosphere that surrounds this fragile planet. And then you work on that research as well, and fusion as well. And then AI is, of course, you know, now everybody's, you know, kind of has an opinion about it. Which of these, if you could only work on one, fusion, climate's the other one. The climate is also in our atmosphere. So you think about all the three things in the atmosphere, nuclear weapons, you know, pandemics, diseases, and climate change is in the climate. And then you think about the other type of cloud, yeah, the internet and propagating systems. How would you rank those, you know, kind of concerns of existential risk and worried about the future of humanity perhaps?

Where do you rank the different levels, you know, whether it be climate change that you're involved with, whether it be looking for nuclear, you know, fusion as a reliable energy source to alleviate the climate issue and power AIs perhaps and then epidemiological issues like like we discussed earlier and then the threat of AI like where do you rank like which ones are most important to you to kind of create not guardrails but just to be aware of safety implications and then the benefit like you keep you're very optimistic person so yeah so what you're most optimistic about which you most scared about yeah because I think these like projects that I am involved right they're actually going to point out to the same theme because I you know I care a lot about

for sustainable development. So when we think about our society progressing at a whole, we need to be thinking about the impact of different things happening in the earth to our children, our children's children. And it's important that we think for them and then trying to find solutions now rather than later. So the climate project that I was involved is trying to look at the long-term impact of these different policies and hopefully provide suggestions for policymakers.

and then more actionable kind of project is nuclear fusion. What if we can create a new type of energy sources that is cleaner and hopefully that's more sustainable, right? And that's a more actionable solution to sustainable development. And epidemiology is undoubtedly a critical part of this because if we cannot be healthy, right, in general, then we cannot sustain. Shut down research. For long term.

So in terms of poking garbage on different. things, I think, yes, like definitely we should put God Reels on things that are more tangible and becoming faster becoming reality. But again, I think it has to be a dialogue between policymakers and technological developers, and we should have the right people doing their job, like, you know, the expert are doing their, according to their expertise, to do their job, right? We should not force technologists to come up with policies. We should not force policies. We should not force policymakers to focus on technological development. Right.

People should be experts in their domain and then, you know, more or less, it's important to be cross-disciplinary. I mean, you're doing that uniquely. But even for me, like, you know, I am a computer scientist and I am involved with a lot of interdictional project, but I always feel like I'm never a climate scientist. I'm never going to be a physicist. And because I have really good collaborators in my team, and that's how we were able to make the progress in projects.

But you're obviously, so learning in all these fields. So, you know, not too many computer scientists can talk about Lorenz and variance violation, for example. As we wrap up, I got two major topics that are super important to me and my audience as well. We have a lot of listeners out there in academia and science, obviously. I've had 21 Nobel Prize winners on the podcast. But I want to ask you about the AI scientist concept. First of all, if you can define it, what led you to this concept and what they can do for us, but start off with the genesis of it. Yeah.

So, like, it actually just like naturally grow out of the project that we have. been working on so far, right? Because we thought, like, okay, what if we can discover symmetry from data? And later we thought, what if we can discover equations from data? And we have shown examples that AI algorithms were able to successfully do that. So then we were discussing with a lot of scientists in different fields from biology to chemistry to physics. And they all seem to be passionate about improving their own productivity. Right. So then that's, you know, naturally comes the idea of AI scientists. Again, you know, the goal is not. not to replace scientists, but rather have more like a scientific assistant, right?

So these, you know, AI scientists will work alongside with our scientists, and then they will accelerate each step of the scientific discovery process. As I actually mentioned in the beginning of this podcast, you start with the data collection or observation, generate a hypothesis, and then you analyze and then test the hypothesis experimentation, and then you go back to generate more hypotheses, right? this feedback loop can be accelerated by AI in every single step. Like example is, okay, before we have to do literature search, right?

We will have our student or ourselves, you know, read a book or going on the internet. But these days with AI, you can read 3,000 books within a lunch break. Like, I don't think any human is capable to do that. So this kind of massive improvement in productivity, right, is just like a very strong indication that AI scientists is going to happen and it's going to. you be a big, you know, productive boost for all the scientists? How can we implement it now? I mean, can I heard Google has a project called AI scientists, something like that,

but I can't get access to it even though I'm a scientist. I got access to making videos. I applied, but I haven't got access faster to V-O-3 than I got to that. So V-O-3 I just used for making, you know, interest to the podcast. But how can our listeners and viewers out there, how can they actually get access to it and put it on their phones or put it on their computers and actually start, you know, doing research with it rather than, you know, hopefully in the, the future of being able to do it. How can we do it now? Or is that possible?

So, yeah, I think these concepts are still pretty much in a lab type of concept, like things that we are working on in our group is we open source a lot of our code and paper, right? You can all find them on my website. Then you can try it out. One of the web app that we developed is basically showing all the molecules that we generated with our algorithm. And then if you are a medicinal chemist, and you can go on this web app and you can test whether these, you know, molecules are up to you are standard. And if you really like these molecules, you can even take them to synthesize. That's kind of the feedback loop that we hope to build. And now we're also talking with people in, for example, in materials. And there are some professors here at UCST, they have robotic labs.

And these robotic labs can quickly, you know, test different combinations of materials and test their properties. And you can imagine there's the, you know, AI brain that drives all these experimentation. So that you can really connecting the AI in the computer with the AI in the real world. Yeah, I love to do a tour and figure out what those are like. Yeah, I was going to ask you about the next frontier that you're excited about. Is it robotics? Is it, you know, embedded AI systems? We're supposedly going to have optimists from Tesla coming out not too long from now, or XAI, I guess it's called.

What do you see as the benefit to, you know, sign? Let's just stick the sign. I mean, I know it'll be good at unloading my groceries from my car or something. But what can a robot like optimist, can it do anything for me as a scientist? Yeah, so I feel like the biggest frontier right now is what people call the foundation model, right? So before, like, people would have to build separate models for individual tasks. Like, even in my own lab, you know, we build models for causal discovery, for forecasting, for hypothesis generation. But it's just becoming more obvious now, like a lot of these models share the similar backbone architecture.

and they are also like transferring knowledge from each other. So the concept of foundation model is to have a single model that, you know, is pre-trained with common knowledge and then fine-tune for different tasks. So essentially I'm trying to automate myself out of this loop. That was my last question. Yeah. So in 1600 or so Galileo is my other favorite avatar, my favorite scientist in history. He actually came up with the initial laws of what we now call Werenzenberryance and Relativity.

but he was also a professor and he made his living not only being a teacher, but he also had to have his students live in his house. So can you imagine you're going to have 12 graduate students and sister? I don't even have a big of a house. Yeah, you don't have a house yet. I don't even have that big of a house. Okay, okay. Well, I'll have to get you a bigger house.

UCSD should do everything they can possibly to make you happy because you're so valuable to our campus. But the future of academia and in particular of our profession as professors, who you're an educator as well. as it comes through so clearly in your passion, enthusiasm. And the word education comes from Latin to bring out of, which is that you bring things out of your students and you do so, so well. And I just love this article in Quantum Magazine that was about your research and just your story. It's just so fascinating.

We'll have a link to that in the article and the video and the show notes and the podcast notes as well. I want to ask just a selfish question. Is our profession? Is it under threat? Because why learn, you know, Galileo came up with the, with the idea that you should test gravity in the lab by rolling things on an inclined plane, so it was much slower. They didn't even have clocks back then.

They had pendulum, you know, they had their pulse. That was it. That was basically all they had. Your summer starts now with Memorial Day deals at the Home Depot. It's time to fire up summer cookouts with the next grill, four-burner gas grill, on special buy for only $199. And entertain all season with the Hampton Bay West Grove seven-piece out. door dining set for only $499.

This Memorial Day, get low prices guaranteed at the Home Depot. While supplies last, price invalid May 14th of May 27th, U.S. only exclusions apply. See Home Depot.com slash price match for details. So to measure something falling in free fall was really hard to do that accurately, and that's why he took 1600 years from Aristotle to him to prove that heavier things fall at the same rate as light things with the leaning tower Pisa, whether he did that, we don't know.

But certainly his story's great. And then the question I have is, as a professor, why learn, you know, physics 1A from Brian Keating? Why learn, you know, about a ball rolling down an inclined plane, which from me, when you could have an AI avatar, you could have a robot, you could have all of his words are digitized. We did the first ever audiobook, which took 21 hours to read. All this exists. Or Einstein learn relativity from Einstein, not from Brian Keating. What are the risks from me to me, to you perhaps? although I think you're more impervious than anybody I've ever met.

You're like the MacGyver of science and technology. But tell me, what do you see is the future of education? Are we going to have jobs? Or are our students going to have professorships or maybe, you know, sooner than that? Is a professor still going to look the same in 20 years? Yeah, I think that's a very intriguing question. Like, you know, obviously academia is facing, you know, an identity crisis now, and especially in computer science, right?

because I think the most exciting application for LM is for coding, for writing code. I love it. Even I'm horrible. I never learned to code and I feel like, oh, it was good for me to wait because I could use that energy to build stuff in the lab. And now I have cursor do it and, you know, does it much better than I can do. Exactly. So, yeah, like I think back to your question, like we do have to rethink like our role as an educator, as a professor in university. Right.

So knowledge, you know, knowledge sharing or kind of just like educating the next generation will look very different right now. Because I always feel like, well, are we going to be out of jobs? Maybe. But then, you know, as I think about it, a lot of times when I interact with my student, it's actually not about the knowledge I share with them or like a specific technical concept I teach them. Right. It's really just about, you know, this personal interaction and how they were able to. learn from this very nuanced details of how I teach and how I think and how I try to solve a problem. I feel these kind of nuanced details or interpersonal interactions are always going to be

there and that's actually very important for education, right? And I think, I don't think these type of nuanced interactions is going to be replaced by AI, right? Because AI is just going to be a much more efficient tool out there for us to use, but it's not going to replace these type of very important roles. I feel like AI is almost training us in some ways. And the education, the reason I brought up the Latin of that, which, you know, education, educare, it means to pour out of, not to pour into, which we don't want to think, oh, I'm just spraying information in my students. Yeah. Oh, I love your information. No, but they don't care. But really the job is to bring stuff out. And I feel like that's what a prompt is. You know, so we're teaching people, you know, we should be teaching

people had a prompt. You know, the joke 10 years ago was learn how to code if you want the job. Now, I'm proof you don't need to really know how to code. It helps to know a little bit. But teaching, you know, my kids had a prompt. I mentioned this, this Sora, Sona or sorry, I forget what it is. You can make, well, you can make videos, but you can make these beautiful songs. And she was prompting it. And she was like, oh, I really want, you know, do a leapas, you know, kind of style. And then the AI comes back and says, you used a forbidden word. You can't use anyone's actual style. So she learned. Oh, okay, so I have to say like, describe what Doolipa is like, you know, what her sound is like, her style is like.

So she learned how to prompt better from the AI, even though the AI, you know, kind of rejected this prompt. So I find it very, very fascinating. And so I want to just conclude maybe with if personal story about how you came to be, again, the name of the podcast into the impossible, I like to conclude a lot of the interviews that I do with this question, you know, if the only way to know the limits of the possible is to go into the impossible, what kind of lessons. or teachings, would you give to a younger rose you, a 20-year-old, Rose you, or 15, whatever,

as a kid, maybe a little bit of your personal story, your backstory. If you had 10 minutes to talk to her as a little girl or whatever, what would you say to her? What would you do to give her the courage and the intensity, the passion, the caring and the intellect that you have that allows you to go into the impossible? What would you say to her? Yeah, I think, you know, if I were going to talk to my younger self and I would say, like, you know, just take a little bit more risk. I think because as an immigrant, you know, I came here after my undergrad and got into a PhD program in the U.S.

So that was a big deal to my family. But then, you know, because I'm an immigrant and I always have this kind of anxiety that if I don't follow things, you know, exactly if I don't follow the dogma and that I won't be able to stay in this country. But it does seem now it's even getting harder to stay in this country. But then if I had an opportunity, right, then I would say, well, why don't you take a little bit more risk and then just trying to do things that you feel you can make an impact on? Right. So like, you know, but when I started out as a PhD student, I was just simply following whatever my professor gave to me at the research project. So I spent like two years working on project that I weren't really passionate about. And then later on, right, as I started to become more confident in things, and I know how to pick projects myself, even those projects were a little bit far away from, you know, my advisor's research direction.

And later on, I expanded to domains that are not even in computer science, like outside of computer science, right? And you always get questions from, you know, 60 or 80 years old. the professors in this domain, like whatever you're doing is, it's garbage. And I don't think, you know, people like in the audience, and, you know, if you wanted to do something, you're passionate about, just go for it.

And obviously, you're not going to please everybody, right? But if you feel you're solving a real problem, and I'm sure there will be people who are appreciating and will give you support and encouragement that you need to make progress. That's wonderful. Well, if they try to come for you, they're going to have to go through me, Rose. I won't let you go. You're too valuable to us that you see San Diego. Thank you so much for everything you do. It's a lot of fun. It's so much fun to have you here. I will put a link to the article about you, which I think is just the first of many. You're so impressive. And I'm so proud to have colleagues like you at UCSD. It's what makes being here so wonderful. Thank you, the honor. Thank you very much. Thank you. I'm so proud to have colleagues like Rose, at UC San Diego. She's exactly the kind of fearless, cross-disaster.

disciplinary maverick thinker that we need, as AI reshapes everything from traffic patterns to the fundamental laws of physics itself. Her optimism about human AI collaboration combined with her track record of solving real-world problems gives me hope that we're heading towards an augmented intelligence rather than a replacement of our human intelligence. And if you enjoyed this deep dive into AI's role in scientific discovery, you'll love my episode with Jan Lecun, where we explore whether or not artificial intelligence can truly understand the world, or whether or not it's just a very sophisticated pattern matcher. Two brilliant minds, two very different perspectives on the future of machine intelligence. Don't forget to like, comment, and subscribe. So you don't miss our next

breakthrough episode. Ambition comes in all shapes and sizes. At First Citizens Bank, we roll with your goals because we're built for what you're building. Fit for your ambition. Enjoy more ways to save at Ralph's, like low prices in every aisle. And when you download the Ralph's app, you can clip and save more with digital coupons every week. Plus, you can earn fuel points to save up to $1 per gallon at the pump. At Ralph's, you can enjoy more ways to save and more rewards every time you shop. So it's always easy to save big every day with savings and rewards.

Ralph's, SoCal for over 150 years, savings may vary by state. Fuel restrictions apply. C-Sight for details.