While language models may help generate new ideas, they cannot attack the hard part of science, which is simulating the necessary physics," says AI professor Anima Anandkumar. She explains how her team developed neural operators — AI trained on the finest details of the real world — to bridge this gap, sharing recent projects ranging from improved weather forecasting to cutting-edge medical device design that demonstrate the power of AI with universal physical understanding.
Learn more about our flagship conference happening this April at attend.ted.com/podcast
Hosted on Acast. See acast.com/privacy for more information.
[00:00:00] TED Audio Collective
[00:00:13] The data that undergirds what we call artificial intelligence
[00:00:17] can sometimes be limited, static, biased, and even inaccurate.
[00:00:24] When we apply these flawed sets to large language models,
[00:00:28] it creates the risk for what scientists and engineers call hallucinations.
[00:00:32] When a generative AI model produces false or misleading information
[00:00:37] that it presents as true, hallucinations range from embarrassing at best
[00:00:41] to inaccurate, drastically harmful, and deeply offensive at worst.
[00:00:47] This is a puzzle that many experts are working on.
[00:00:50] How can we ground the simulation in more reliable data?
[00:00:54] And what would that unlock for the future of AI technology?
[00:01:00] This is TED Tech, a podcast from the TED Audio Collective.
[00:01:04] I'm your host, Cheryl Dorsey.
[00:01:07] Today we have the privilege of hearing from Dr. Anima Anandkumar.
[00:01:11] She is a professor at the California Institute of Technology
[00:01:14] and a former senior director of AI research at NVIDIA.
[00:01:19] Her work explores the limitations of traditional large language models
[00:01:22] and envisions a future where AI is grounded in the physical world.
[00:01:27] This approach promises more accurate utility and modeling,
[00:01:30] which can accelerate the design of life-saving devices, tools, and more.
[00:01:35] But before we dive in to Dr. Anumkumar's talk,
[00:01:38] a quick break to hear from our sponsors.
[00:01:51] Sometimes things in the world of technology are complicated
[00:01:54] and need careful explaining.
[00:01:57] Sometimes they just need a little hard truth.
[00:01:59] I don't think anyone is going to buy a banana with crypto
[00:02:02] at any point in the foreseeable future.
[00:02:05] I'm Lizzie O'Leary, the host of Slate's What Next? DVD,
[00:02:08] your clear-eyed guide to technology, power, and the future.
[00:02:12] Friday and Sunday, wherever you get your podcasts.
[00:02:22] I grew up with parents who are engineers.
[00:02:26] They were among the first to bring computerized manufacturing
[00:02:29] to my hometown in India.
[00:02:32] Growing up as a young girl, I remember being fascinated
[00:02:35] how these computer programs didn't just reside within a computer,
[00:02:40] but touched the physical world
[00:02:42] and produced these beautiful and precise metal parts.
[00:02:47] Over the last two decades, as I pursued AI research,
[00:02:52] this memory continued to inspire me
[00:02:54] to connect the physical and digital worlds together.
[00:02:59] I am working on AI that transforms the way we do science and engineering.
[00:03:05] Scientific research and engineering design
[00:03:08] currently involves a lot of trial and error.
[00:03:11] Many long hours are spent in the lab doing experiments.
[00:03:16] So it's not just the great ideas that propel science forward.
[00:03:20] You need these experiments to validate findings
[00:03:24] and spark new ideas.
[00:03:27] How can language models help here?
[00:03:30] What if I ask Chad GPT to come up with a better design
[00:03:34] of an aircraft wing
[00:03:35] or a drone that flies under turbulent winds?
[00:03:38] It may suggest something.
[00:03:41] It may even draw something.
[00:03:43] But how do we know this is any good?
[00:03:46] We don't.
[00:03:47] Language models hallucinate
[00:03:49] because they have no physical grounding.
[00:03:53] While language models may help generate new ideas,
[00:03:57] they cannot attack the hard part of science,
[00:03:59] which is simulating the necessary physics
[00:04:04] to replace the lab experiments.
[00:04:07] In order to model scientific and physical phenomena,
[00:04:11] text alone is not sufficient.
[00:04:14] To get to AI with universal physical understanding,
[00:04:19] we need to train it on the data of the world we observe.
[00:04:24] And not just that, also its hidden details.
[00:04:29] From the intricacies of quantum chemistry
[00:04:31] that happen at the smallest level
[00:04:34] to molecules and proteins
[00:04:36] that influence how all biological processes were
[00:04:40] to ocean currents and clouds
[00:04:43] that happen at planetary scales and beyond.
[00:04:46] We need AI that can capture
[00:04:49] these whole range of physical phenomena.
[00:04:53] We need AI that can really zoom into the fine details
[00:04:57] in order to simulate these phenomena accurately.
[00:05:01] To capture the cloud movements
[00:05:04] and predicts how clouds move and change in our atmosphere,
[00:05:09] we need to be able to zoom into the fine details
[00:05:12] of the turbulent fluid flow.
[00:05:15] Standard deep learning uses a fixed number of pixels.
[00:05:19] So if you zoom in, it gets blurry
[00:05:22] and not all the details are captured.
[00:05:25] We invented an AI technology called neural operators
[00:05:29] that represents the data as continuous functions or shapes
[00:05:34] and allows us to zoom in indefinitely
[00:05:37] to any resolution or scale.
[00:05:41] Neural operators allow us to train on data
[00:05:45] at multiple scales or resolutions.
[00:05:48] And also allows us to incorporate the knowledge
[00:05:51] of mathematical equations to fill in the finer details
[00:05:55] when only limited resolution data is available.
[00:05:59] Such learning at multiple scales
[00:06:02] is essential for scientific understanding,
[00:06:05] and neural operators enable this.
[00:06:09] With neural operators,
[00:06:11] we can simulate physical phenomena such as fluid dynamics
[00:06:15] as much as a million times faster than traditional simulations.
[00:06:21] Last year, we used neural operators
[00:06:24] to invent a better medical catheter.
[00:06:27] A medical catheter is a tube that draws fluids out of the human body.
[00:06:32] Unfortunately, the bacteria tend to swim upstream
[00:06:35] against the fluid flow and infect the human.
[00:06:39] In fact, annually there's more than half a million cases
[00:06:43] of such health care related infections,
[00:06:46] and this is one of the leading causes.
[00:06:49] Last year, we used neural operators
[00:06:51] to change the inside of the catheter
[00:06:54] from smooth to rich.
[00:06:57] With ridges, now we have vortices created as the fluid flows.
[00:07:03] And we can hope to stop the bacteria from swimming upstream
[00:07:07] because of these vortices.
[00:07:10] But to get this correct,
[00:07:12] we need the shape of the ridges to be exactly right.
[00:07:16] In the past, this would have been done by trial and error.
[00:07:20] Design a version of the catheter, build it out,
[00:07:25] take it to the lab,
[00:07:26] observe an hypothesis if something went wrong,
[00:07:30] rinse and repeat and redesign again.
[00:07:33] But instead, we thought AI,
[00:07:36] the behavior of the fluid flow inside the tube.
[00:07:40] And with it, our neural operator model
[00:07:42] was able to directly propose an optimized design.
[00:07:46] With 3D printed, the design only wants to verify that it worked.
[00:07:52] The bacteria are not able to swim upstream
[00:07:54] are instead being pushed out with the fluid flow.
[00:07:58] In fact, we measured the reduction in bacterial contamination
[00:08:03] by more than 100 fold.
[00:08:06] So in this case, the neural operators
[00:08:08] were specialized to understand fluid flow in a tube.
[00:08:12] What other applications can AI tackle
[00:08:16] and help us solve such pressing problems?
[00:08:20] Can deep learning beat numerical weather models?
[00:08:24] A group of leading weather scientists
[00:08:27] asked this question in February 2021
[00:08:30] in a Royal Society publication.
[00:08:34] They felt that AI was still in its infancy
[00:08:37] and that a number of fundamental breakthroughs
[00:08:40] would be needed for AI to become competitive
[00:08:43] with traditional weather models.
[00:08:45] And that could take years or even decades.
[00:08:48] Exactly a year later, we released 4CastNet.
[00:08:53] Using neural operators,
[00:08:55] we built the first fully AI-based weather model
[00:08:59] that is high resolution
[00:09:01] and is tens of thousands of times faster
[00:09:04] than traditional weather models.
[00:09:06] What used to take a big supercomputer
[00:09:09] can now run on a gaming PC that you may have at home.
[00:09:15] This model is also running at the European Center
[00:09:18] for Medium Range Weather Forecasting,
[00:09:20] one of the premier weather agencies of the world.
[00:09:24] And our AI model is not just tens of thousands of times faster
[00:09:28] than traditional models.
[00:09:30] It's also more accurate in many cases.
[00:09:34] On September 16th last year,
[00:09:36] Hurricane Lee hit the coast of Nova Scotia, Canada.
[00:09:41] A full 10 days earlier,
[00:09:43] our 4CastNet model correctly predicted
[00:09:46] that the hurricane would make landform.
[00:09:49] But the traditional weather model
[00:09:51] predicted the hurricane would skip the coast.
[00:09:53] Only five days later on September 11th
[00:09:56] did the traditional weather model
[00:09:58] correct its forecast to predict landform.
[00:10:01] Extreme weather events such as Hurricane Lee
[00:10:04] will only increase further
[00:10:06] unless we take action on climate change,
[00:10:10] such as finding new clean sources of energy.
[00:10:14] Nuclear fusion is one of them.
[00:10:17] But unfortunately there are still big challenges with it.
[00:10:21] The fusion reactor heats up the plasma
[00:10:24] to extremely high temperatures to get fusion started.
[00:10:28] And sometimes this hot plasma can escape confinement
[00:10:32] and can damage the reactor.
[00:10:35] We train neural operators to simulate
[00:10:37] and predict the evolution of plasma inside the reactor.
[00:10:42] And with it, we can use this to predict disruptions
[00:10:46] before they occur and take corrective action
[00:10:49] in the real world.
[00:10:51] We are enabling the possibility of nuclear fusion
[00:10:55] becoming a reality.
[00:10:57] So neural operators and AI broadly
[00:11:01] are enabling us to tackle hard scientific challenges
[00:11:05] such as climate change and nuclear fusion.
[00:11:09] To me, this is just the beginning.
[00:11:12] So far these AI models are limited
[00:11:16] to the narrow domains they're trained on.
[00:11:19] What if you had an AI model
[00:11:22] that could solve all and any scientific problem
[00:11:26] from designing better drones, aircrafts, rockets
[00:11:30] and even better drugs and medical devices?
[00:11:34] Such an AI model would greatly benefit humanity.
[00:11:39] This is what we are working on.
[00:11:41] We are building a generalist AI model
[00:11:44] with emergent capabilities that can simulate
[00:11:47] any physical phenomena and generate
[00:11:50] novel designs that were previously out of reach.
[00:11:54] This is how we scale up neural operators
[00:11:57] to enable general intelligence
[00:11:59] with universal physical understanding.
[00:12:03] Thank you.
[00:12:07] That was Dr. Anima Anan Kumar at TED 2024.
[00:12:17] And that's it for today.
[00:12:19] TED Tech is part of the TED Audio Collective.
[00:12:21] This episode was produced by Nina Bird Lawrence,
[00:12:24] edited by Alejandra Salazar
[00:12:26] and fact-checked by Julia Dickerson.
[00:12:29] Special thanks to Maria Ladius, Farah DeGrunge,
[00:12:32] Daniela Belarezzo and Raksen Highlesh.
[00:12:35] I'm Sherelle Dorsey. Thanks for listening in.