
Loading summary
Host
Karthik, welcome to the show.
Karthik
Happy to be here.
Host
Yeah, thanks for coming on. I think it's going to be fun. You guys just announced a bunch of things at the University of Michigan. We're going to talk about that and then also talk a little bit about how AI is changing. Research, education, I think, like finding a job, the economy a little bit too. But to kick things off, what did you guys just announce at the university?
Karthik
So, first of all, I think it's an incredibly exciting time to be alive given all these things that are happening not just around AI, but in science, research and all of these things you just spoke about. So at the clock on, we announced the Institute for Agentic Computing, which is a partnership between the OpenCLA foundation and the University of Michigan. So the goal of the Institute will be to develop responsibly powerful agentic frameworks that people can use for a wide range of things. Again, we can talk more about what those things are. So there'll be a core development of agentic infrastructures and then a large number of people working with those developers to apply those agentic frameworks to very many different fields. Any field that humans have ever touched, I think will be identified.
Host
And how did this come about, the creation and the thinking around starting this?
Karthik
Yeah, so first of all, I think Open CLI is one of the most popular software frameworks that is out there. Right.
Host
And what is openclaw for someone who doesn't know?
Karthik
So Peter Steinberger and his team introduced OpenClaw in November. So what it does, really, at its core, it basically reduces friction. It basically gives you a very strong personal AI assistant that you can use to automate a wide range of tasks. Right. And many of the initial applications are, you know, people were using OpenClaw on their laptops to automate many of the manual things they used to do with emails and communications and files and things like that. Right. But that's just scratching the surface, you know, Then people started using it as a social agent. So I use open cloud to train my agent and now my agent can talk to your agent and they can start doing interesting things that part of which are controlled by us and part of which some of these, the agents themselves do. So the way I think about it is you had chatbots which came on the scene about three years ago, and over the last year they've become extremely powerful. But those are just giving you ideas. You type something, it gives you an idea, you use that and then do something.
Host
It's kind of like better. Google is sort of A way to think about it.
Karthik
Yeah, much better. Google that can also give you cognitively powerful things, can put ideas together. I don't want to minimize any of those things. Those are chatbots. And then you have agents that basically act. So it's not just passive that you're giving it some, that it is giving you some information and then that's it, you go do something else. So agents act upon information.
Host
So it's like saying, hey, go order me a pizza. And it goes and calls Domino's on the website and places your order and you get a pizza.
Karthik
Exactly. Maybe it'll deliver the pizza too soon.
Host
The robots. Yeah, the autonomous robots.
Karthik
Yeah, actually that's another thing. So it is not just restricted to software. Agents actually act in the physical world. Also, people control their own personal robots with openclaw.
Host
Oh really? I didn't know that was happening already.
Karthik
Yeah, Agents are not just for software. So anyway, I mean, if you think about we have personal chatbots, then we have personal agents. I think the next evolution is social. I think these things are going to the social and economic infrastructure around which our society is organized. So all of those may now be identified. So you could have. I could have a few agents of my own that encode some of my skills and my expertise. You could have a few of your own with your skills, your expertise. They can talk to each other, they can collaborate in a certain way and it can be very decentralized. So openclaw, if you want that one sentence headline, it is like an operating system for the agentic world, if you will.
Host
And OpenClaw is open source. Correct. And there's a foundation that is attached to openclaw.
Karthik
Correct.
Host
How does this all relate to the institute and the university?
Karthik
Yeah, so you're right. I think OpenClaw is an open source project and that is why it basically really caught fire. I think there may be more than 3 million users right now. Really? In almost no time. Yeah. The foundation was formed to ensure that the future of OpenClaw is open source. And many of the people who are co developers of OpenCL, like Peter Steinberger's team, so they'll be in the institute as well. So think of that as the core layer and then there's a layer around them which is you can think of people at the University of Michigan or collaborators, pretty much anyone around the world who's interested in taking these frameworks and then adapting them to their specific domains and using them.
Host
And I know we got connected from Dan in the investment office at University of Michigan. They're kind of involved in this a little bit. They funded something, I think they funded a company that is sort of third party helping also. Can you explain what's going on there, just so people know?
Karthik
Right. So first of all, the University of Michigan's fund is heavily involved in the OpenCLA Foundation. Then they also formed a new company called the Lobster Compute Company.
Host
Lobster Compute. That's awesome.
Karthik
Right? And basically so there, I mean, the OpenCloud foundation is purely doing open source and it is basically helping the developers. And then the Lobster Company, Lobster Compute Company is basically, you can think of it as the investment wing. So that's where maybe money goes in. For more startups and many other things,
Sponsor/Announcer
this episode is brought to you by numerl. Numerl is the fastest, easiest way to stay compliant with US Sales tax and global vat. It's easy to set up and they automatically handle all registrations, ongoing filings, and their API provides sales tax rates wherever you need them with all the integrations you need. Numeral sports over 2,000 customers in both the US and globally and they pride themselves on White Glove High Touch customer service service. Plus they guarantee their work and they'll cover the difference if they mess anything up. They're fresh off a Fundraise closing a 35 million dollar Series B from Mayfield which they're going to reinvest into building an even better product. If you want to put your sales
Host
tax on autopilot, check out Numero at
Sponsor/Announcer
their new domain numeral.com that's n u m e R-A-L.com for the end to end platform for sales tax and VAT compliance. This episode is brought to you by Flex. It's the AI native private bank for business owners. I use Flex personally and I love it because they use AI to underwrite the cash flow of your business, giving you a real credit line. The best part is 60 days afloat, double the industry standard. Flex has all the features you'd expect from a modern financial platform like unlimited cards, expense management, bill pay that syncs
Host
with your credit line and and their
Sponsor/Announcer
new consumer card, flexelite. Flexelite is a brand new ramp like experience for your personal life. A credit card with points, premium perks, concierge services, personal banking, cars and expense management for your family net worth tracking across public and private assets and a whole lot more fully integrated with your business spend. One card for your businesses, one card for your personal life. One card for everything. To skip the waitlist, head to Flex 1 and use My Code Turner to get an additional 100,000 points worth $1,000 after spending your first $10,000 with Flex Elite. That's Flex Dot 1 and Code Turner for $1,000 on your first $10,000 of spend. Thank you, Flex. And now let's jump in.
Host
And it might be interesting then for people to understand how universities work. I mean, we can maybe talk about that, like how the University of Michigan kind of intersects with this institute. I know we talked about it the other day, but I had no idea this is how it works. So can you kind of explain what this whole setup is with a university? Institutes, departments, all this kind of stuff?
Karthik
Correct. Yeah. So I think even without the Institute for Agentic Computing, I think universities are very fascinatingly organized places because there is, first of all, universities do many different things.
Host
Yeah. It's not just teaching classes.
Karthik
It's not just teaching classes.
Host
It's like a small percentage.
Karthik
Yeah. Advancing research, innovation, startups.
Host
You've got sports teams.
Karthik
Sports teams.
Host
You've got hospitals, sometimes hospitals.
Karthik
Yeah. Especially the University of Michigan. So these are very complex organization, but if you want to break it down, fundamentally the core of the university lives in departments. Right. The department of Physics, department of Aerospace Engineering. Right. Department of Sociology. So this is where students get admitted to and they take courses, they get their degrees, et cetera, et cetera. Right. Mainly from a teaching perspective, but also from a research perspective. But as you can imagine, especially these days, research is not siloed. Right. If you're in aerospace engineering, it's like you don't just work on aerospace engineering. If you're in sociology, you don't just work on sociology. A lot of the interesting problems, research problems come when areas intersect.
Host
So it could be the intersection of. You maybe do ethics in aerospace or something like that.
Karthik
Correct. In fact, we have a research area in space Ethics.
Host
Really?
Karthik
Yeah. So you were not far off there. So. Yeah, you can think of that. For instance, the engineering school and the business school have a joint program,
Host
so
Karthik
there are many, many other collaborations. So think of the departments as some kind of low level units where faculty are hired, tenure is given, students are educated, classes happen, et cetera. And on top of departments we have so called institutes. So these bring different departments and different researchers and different kind of students together. For instance, if you take something like computational science, which is the institute that I direct, it's called miechd Michigan Institute for Computational Discovery and Engineering. We have faculty and students from, I don't know, 40 different departments across campus, all exploring different aspects of computing for science. Right. Because you're developing a certain kind of scientific solver infrastructure that you can use to solve materials problems, chemistry problems, aerospace problems, I don't know, robotics, all of these things. So institutes sit on top of departments and bring people together to do interdisciplinary research.
Host
So maybe how many departments then? How many institutes are in the University of Michigan?
Karthik
That's a hard question. I don't think anyone knows. Okay, I'm just kidding. The rough number is about 200 departments. And one of the amazing things about Michigan, IMSA is pretty much all of the 200 would be in the top 10 of any ranking you can imagine. So that's pretty special. And it covers all areas of human activity. I would say institutes would number in, I would say maybe dozens of institutes. Examples would be, I spoke about mine at the institute that I direct, Institute for Agentic Computing that's coming up that we just announced yesterday. So that is one institute. We have Institute for say Firearm Safety, their Institute for Social Research. In fact, that's the largest social science organization in the entire world.
Host
Yeah, my mother in law actually worked in the institute for probably about a decade.
Karthik
Okay, wonderful. Yeah. So there are institutes of various sizes and various scopes that bring together faculty from many different disciplines together to go after some grand challenge problems.
Host
So you just announced something, some research that you had done and you did it at this thing called Clockon, which I'm not sure what day this recording is going to come out, but by the time someone's listening to this, it's already happened. It's this open clock kind of conference. What is Clock On? For someone who's never heard of Clockon,
Karthik
Clock on is basically think of it as a gathering place where you have a spectrum of people who are, let's say core developers and core users of OpenCloud. Or you could have people who are just genuinely curious about what is happening. And typically when you bring together tech meetups, you're catering to a very specific specialized audience. But Clockon is much more democratic. It does not distinguish between an ordinary person who is interested in AI or agents versus somebody who is a big developer. Right. So it brings different kinds of people together and has something for everybody. Right. So you have very powerful keynote demos that show how these things can change science. You have more basic things like what is openclaw, how do I use it, how do I install it in my computer? So it basically runs the whole gamut. So it's a very interesting mix of people and expertise all centered around how personal AI can basically accelerate a lot of things.
Host
We do so this was kind of like a meetup conference thing. Just a bunch of enthusiasts and experts and heavy users coming together to spend time, do some demos, do some presentations.
Karthik
Correct. I think it's mainly focused around getting people to meet, getting people to talk, and then, you know, show some demos, educate them, and maybe get some ideas.
Host
I think there was a startup pitch competition too, that this may have not ended up happening. But I heard that Dusty May, the basketball coach, was presenting an award or something. I don't know. This might not end up happening. Maybe we have to cut this part,
Karthik
but it may be possible. I hear the Michigan cheer team is also going to be there.
Host
Oh, wow.
Karthik
Hail to the victim.
Host
That's a pretty broad spectrum.
Karthik
Yeah.
Sponsor/Announcer
When you can build anything, amplitude lets you know how to build the right thing. Use human language to get complex answers about your products. No more manually selecting events or building charts or dashboards.
Host
Just ask.
Sponsor/Announcer
Use agents to sense changes in customer behavior, decide what's causing them, and ask you if it's okay to fix it. Continuously in the background while you work. Get the answers you need while building directly in the tools you are already in. Like Claude Cursor, Lovable and more and more. And for the first time, understand if your agents actually work. Measure quality, debug failures, experiment and measure their ROI with agent analytics, amplitude with AI analytics. All you have to do is ask.
Host
And so you just announced some new research? I think so. What did you just announce?
Karthik
Yeah, so this is pretty amazing. So some of our colleagues at MIT and us, we've been working together on scientific discovery with some of these new AI agents. So MIT collaborators have developed something called Science Claw, which is basically a very interesting approach that the big question is, how can science change when you're combining humans, agents, and then powerful tools? And all of this is done completely decentralized. So there is no. It's not like one person is asking people to do this or that people are doing agents. So it's basically the next generation of collaborative science. So we showed a couple of examples yesterday at Clockon. So the first one is a very interesting example. So the question goes like this, right? What do cricket wings and Bach chorales, the basically music and composite materials have in common? That's the broad question. So basically we have agents in biology, agents that understand, say, material properties, and then agents that understand music, all collaborating, coming together in a decentralized way, and then finding a particular part of design space that was unexplored and ending up finding some incredibly useful material resonators there. So it's an example where you have three experts. Think of those agents as experts. They know a lot about their topic, but they don't quite know about synergies between different topics. But then when you bring them together, some very interesting things happen.
Host
And so what could you use it for? What's something that someone might create with this or make?
Karthik
What is resonance? Right, yeah. So if you have, you know, some kind of a material that is excited by a disturbance, Right. You want it to be oscillating at a certain frequency. Right. So the operational use for this is very, very broad. So it is not as if this was discovered today and I'm going to use it tomorrow. But now if you can design material resonators for any kind of property you want, but then how do you achieve that design? So it is that process of achieving that design that these agents found. And the second application that we showed is much more straightforward for people to understand. This is superconductivity. So if you can pass any kind of energy across a medium and you can achieve that with no loss, zero loss. Right. Currently, if you pass any kind of energy, electric current or any energy, there is some losses that happen. Superconductors give you no loss. And they can be used in many, many different applications. As you can imagine, superconductors have been discovered for many decades now. But most, I mean, pretty much all superconductors, for us to use them, they can only be used at extremely low temperatures, close to absolute zero. It's like minus 100, minus 200.
Host
So it's just not practical.
Karthik
It's just not practical in many applications. And you need low temperatures, high pressures and things like that. So we are using these agents to discover superconductors at much higher temperatures, at more realizable temperatures, like room temperature superconductors. Again, I don't want to give the impression that we used this agent and we discovered it, and hence this is a Nobel Prize winning discovery. Right. This is the first part of the long chain of discovery. So next we'll have to build the material, test it, get some information, come back. So it's the first part of the scientific process, but this part is the hard one because the possible space of superconducting materials and configurations is immense. So without some of these newer techniques, it would take a very, very long time to search that space.
Host
Yeah. So it might be interesting to talk about this new technique, how it works. But even before that, the old technique, how would you discover a new theory or possible option, like a new hypothesis in Science. And then how is it changing with AI?
Karthik
Yeah. So let's break down the scientific process. Right. So what is the scientific process? You make some observations in nature and then you basically write down a theoretical model that explains the observation and then you manipulate that model to get the property you want. Right. And these are all like fairly simple models because it's still the mental map.
Host
This is the scientific method that we probably learned about in high school.
Karthik
Exactly, yeah. Theory, observations, experiments. So that's the first part. And then you go to more detailed models. Right. These are the kind of models that a lot of people in the institute that I direct, they run. Right. This is very detailed models of a particular physical process. And then you gain some more insight and then you do an optimization. If I control all of these variables, I get the design I want. But again, a computation is different from reality. So then you go and build that whatever thing that you're designing or studying, do experimentation, do measurements, and then you iterate. Right. So all of these steps still have to be followed in the age of AI. But before, until recently, all of these steps used to be done in sequence and they used to be done by different people. Somebody is a theoretician and they take a number of years to figure out the first part. And then there is a specialist in computation. Maybe they run some computations, they run optimizations, and then there is a specialist in measuring things. And then you put things together, maybe you have some outcome in the end,
Host
are you kind of waiting for other people to be done with their phase? And you're maybe working on multiple at a time.
Karthik
Yeah. So generally these things were very sequential. Of course, some of that even before AI has been made more simultaneous and not completely sequential. But what AI has done or has the promise of doing in many. It has done in some applications and presents promise in many applications, is to make all of this simultaneous.
Host
And
Karthik
especially with specialized agents, you can do decentralized science and you don't need to know everything about every domain. But I still think expertise matters, but it accelerates the whole process.
Host
So how does that work? That it makes. So these can all run at the same time in a sort of decentralized. Like if I were to ask you just how does that actually work versus how does AI enable that to happen? What's the thing? Because did you need AI to do this? What does it unlock?
Karthik
See, that's the thing. So people think in most of these cases, AI is an enabler, Right? So think about AI having access to skills and tools. Okay? Skills is my scientific expertise. I can encode as a set of rules, literally in a text file.
Host
Yeah. So you are an expert on aerospace physics. You know every single possible thing to know about how aerospace intersects with the world or something like that, and that's all you know and you don't know anything else?
Karthik
Yeah. So that plus I also have a certain kind of expertise and a certain kind of insight that I can also build into this database. So think of those as skills. My flavor of the scientific method might be different from your flavor of the scientific method because we have different tastes, different perceptions, and we work in different fields. So all of that can be encoded. Right. Think of those as skills. So agents have access to skills and then think of tools. Tools would be like, I have a simulation software that can take a real world problem, turn it into a computation and help understand and manipulate the real world virtually. Tools could also be experimental facilities. I could be in a lab setting, I could have a bunch of robots that do repeated experiments and they can also be connected and coordinated by AI. So I think human expertise and AI supplemented by tools is what makes this happen. Right. It's not AI alone. Right. But sometimes I think the reasoning capabilities of these AI models are so powerful, especially the last six months or so, that maybe many of the human skills are actually being identified by AI on its own. But I still think human expertise matters. So a short answer to your question is AI with access to tools and skills is what makes all of this possible.
Host
Okay. And this intersects with the Institute. You are essentially going to be doing a bunch of these experiments through the Institute, or what is the kind of the Institute going to be doing practically on a daily basis?
Karthik
Yeah. So again, I think, at least at the beginning, the core job of the Institute is to make sure the open source project of developing OpenClaw is done and it's maintained in the best possible way. Right. Because there are so many users and users need confidence that this is going to be open source and it's going to be developed, et cetera, et cetera.
Host
And there's a couple, I think there's two other people that are running the institute with you.
Karthik
Yeah. I mean, for now we have a leadership team. Myself, there is Professor Brad Orr from the University of Michigan and Kurt, whom you may know, Scott Skiffstadt from Engineering. So, I mean, we are the initial leadership team. Right. So that's the leadership team. I think if you think about the core piece, there are the core developers and maintainers, as we call them, and then there's the Next layer of people, like people say my students, for instance, or my colleague students, who want to apply this to their specific problems, not just taking the software and just using it, but also adapting that for specific use cases. And the reason why it's logical to center this in a university is because we have people in every possible discipline. Economics and sociology and medicine and engineering.
Host
So you can kind of just say, like, hey, we need a legal expert. It's like, oh, we've got a guy. Like, there's just a guy who works for the school that's the expert on this thing.
Karthik
Probably in many fields, the world expert sitting right here. But the other important point to think about is it is not just about developing and applying software. Right. It's not about software alone. There is also the physical piece I mentioned. Robotics is an important piece, and we have amazing robotics department. But also to develop this in a very responsible way and in a sense even to prepare society for what is coming. Because a lot of the apprehensions people have about AI is not even about the technology. It's about how fast it is moving and how it disrupts many existing structures. So can we design institutions that are ready for this and can we design this responsibly? And I want to make it very clear this is not just the OpenCLA foundation and the University of Michigan. This is the core. Think of it as a central node. But for this to be a success, we basically need pretty much representation from everywhere in the world, various different groups. So think of it as a meeting place for that.
Host
So could I get involved just as like a person? That is not. I mean, I don't work at the university or openclaw in any way.
Karthik
Absolutely. I feel like I said, there are 3 million users of OpenClaw right now, probably like 500,000 GitHub repositories. In a sense, they're already part of the ecosystem. Right. So this centralizes some of the most important effort. But certainly this is pretty much a window to developers and various kinds of contributors around the world. And the other thing that I want to emphasize is nobody can predict with a great degree of confidence how the technology is going to evolve in the next year or two years. So we'll be very adaptive and dynamic to, you know, the changing scenario. But the goals are very clear.
Host
Yeah, this might be completely outdated in
Karthik
six months, I don't think, you know, the core methods will be outdated in six months. Maybe a particular piece of software that you're using may be updated, might be outdated. But, yeah, I think things do last for longer than that. But I think many other things will change in unpredictable ways.
Host
You mentioned that these institutes are almost like a VC investor, like VC fund in a way. Can you just explain that? Because I thought it was kind of interesting.
Karthik
Yeah, so I mean again, when I explain, when people ask what do you do as a professor? People think and it's fine. People think our entire job is teaching. It's certainly not, because that is teaching in the classroom. And then you're mentoring research students, you're mentoring your PhD students to do original research and then you do your own research, you lead a research group, you have a lab where you're doing some world leading research in your particular domain. So one way to think about almost every professor in a top research university is like a startup founder, because you're recruiting some of the best students in the world. I just completed my PhD admissions, I got 200 applications for PhD. Maybe I accepted two or something. And you're fighting with MIT and Stanford and all these. So just like you're a startup, you're fighting other places to recruit top talent. You do that. And then all of what I said costs money. So you raise money just like every startup founder does.
Host
It doesn't just show up.
Karthik
It does not show up, unfortunately.
Host
So how does that work at the University of Michigan, that funding aspect?
Karthik
Yeah, so normally the largest portion of funding that any faculty member gets or the startup, so to speak, gets, we write proposals to the federal government. Like there is NASA, there is National Science Foundation, Department of Energy, Department of Defense, National Health, Institute of nih, et cetera. So we write proposals on interesting research ideas. And that is also very competitive. I would say maybe one in five proposals get accepted on average. I think at the University of Michigan it's higher, but normally it's one in five.
Host
And this is basically saying like, hey, government, I have this idea, here's what it's going to, here's the impact it could have on the world. Give me $10 million to work on figuring this thing out.
Karthik
Yeah, pretty much. Right. So you present some evidence. Hey, in my past research I've done this. Here is some preliminary work that shows a new direction and it is promising. And yeah, here is my idea, here is my five year plan. Then you're asking for funding and it's very highly competitive, as you can imagine.
Host
What makes it so competitive?
Karthik
Well, there are so many excellent universities in the world. Right. So it's not just the University of Michigan that is good at so many things. There's so many other Research groups around the country who are also pushing the envelope in their own fields. So anyway, so it's again coming back to the original question. Being a professor is like a startup founder. Yeah, right. And in one sense, institutes that sit on top of departments and bring together different faculty can be thought of as an incubator or a vc, because many times in miechd, which is the institute that I direct, we identify an interesting direction of research that is not mainstream yet. So we bring together some faculty members and maybe students and postdocs and build some critical mass around that area. And then maybe we give some seed funds. We run something called the Catalyst Grants Program.
Host
This is without going to the government.
Karthik
Exactly. This is a smaller amount of money. We're not giving 10 million. We give $100,000, for instance. People can use that money to explore an idea and then the institute will actually help those faculty members, professors, to put together a proposal. We recently won $20 million center from the Department of Energy. We're very proud of.
Host
What was that?
Karthik
It's called the Predictive Science Program. So we started a new center called C Prime center for Prediction Reasoning and Intelligence for Multiphysics Explorations. Pretty much as I said, expertise, computations, AI coming together to address a very important problem. So, yeah, you can think of institutes as incubators in that sense.
Host
So what is a good kind of research proposal or idea? Look, like you talked about, it's hard to get them approved. So how do you know if something is a good thing to even spend some time on?
Karthik
I mean, first of all, we don't send proposals completely in a blue sky sense. There are some foundations that say give me your best idea, but that is much more rare. So normally the federal government will have requirements in a certain topic.
Host
So they kind of have a request for research that they put out.
Karthik
Yeah, it's called rfp Request for Proposals. So there somebody may be interested in, say, nuclear fusion, propose some ideas that can improve the efficiency of fusion.
Host
This is someone at the federal government level that is in charge of dishing out these grants.
Karthik
Correct.
Host
They may say we want some work to be done in nuclear fusion or fission or any category. We have 2 billion earmarked and we might fund 20 projects or something like that.
Karthik
Yeah, something like that. Okay. Generally it's not 2 billion, but in any case, as you can tell, I'm
Host
coming into this with no knowledge of how this works.
Karthik
Yeah, but anyway, I think the right order of magnitude for you to think is funding a PhD student for a year costs about $100,000, right. So in that sense funding a full PhD student's time at a university is like half a million dollars over five years. Okay. And normally faculty members have five to 10 students, so that's the order of magnitude. So now you can do a little bit of order of magnitude math.
Host
So each professor, each faculty member may be administering like a $500 million kind of budget per year.
Karthik
A median would be about 500. Again, to give you a sense, the University of Michigan is I think the third largest, has the third largest research program in the US measured in terms of dollars. So our total research activity is about $2.2 billion per year. And I also don't want you to think that it's all federal government. A majority of that is from the federal government, I would say 60%. Our university particularly puts in quite a bit of its own money. And there, I think no other university does so much for research than at least in terms of expenditures, I think about $700 million a year the university spends towards research. And then there is state funding, industry funding, which we can't forget. It's not as much as federal government, but it's an important piece. So it's a whole range of things.
Host
So this is basically like a corporation, big company that has. Are they sort of outsourcing their R and D to a university in a sense like saying, you do this for us and we want to make products from it.
Karthik
Sometimes they're looking for a specific solution, like an outsourced thing, but sometimes they're looking for good ideas and sometimes they're looking for due diligence. Right, because we are experts and we know how to judge things possibly better than most. So it's a combination of things.
Host
What's the benefit of that for the university? What do you get out of doing all this research that some other company benefits from?
Karthik
Well, first of all, they pay us to do it.
Host
Okay, fair.
Karthik
Well that's a start.
Host
But don't you want to capture that economic value for the university instead of a different company?
Karthik
And that happens too. Right. So again I said our annual research expenditure is $2.2 billion or so at the University of Michigan. But we are also, I think after Stanford and mit, the third largest producer of spin out startups out of university research. I think every year we do about 32 or 33 startups. So it's not like just money comes in and then we just produce research papers and graduate students. There are also innovation that is happening and some of our colleagues have done really well in that area. But again, just to make sure I'm answering your question, we don't take money just because it is money. Right. So every faculty member or research group is interested in furthering boundaries of knowledge or innovation on certain problems. And if the funding is aligned with that. So that's the way of having impact. So that's training students is impact. Training students to think, training students, a scientific method is impact solving problems that somebody cares about, like a company or the government. That's impact. Purely pushing the boundaries of research for the sake of pushing the boundary of research. That is impact too. Right. So that's what the investigates. And of course then you also have startups and other things.
Host
So with the case of a startup, when you're the university would own some of the equity in the startup and when there's some kind of an exit outcome, it goes back to the university, to the endowment to just dumped into the budget for a year. How does that usually work?
Karthik
Yeah, correct. If it is a startup that is spinning out of directly spinning out of research that happened at the University of Michigan, funded by say the government or whatever. Yes. Then the university would take some equity and some part of the IP royalty, et cetera. But I must say, I mean, I have a startup myself and the university is actually very fair. It's not like they're here to make money out of it. They truly want to help innovation grow. They want the faculty members impact and students impact to be higher. And of course there's certainly some economic benefit too.
Host
So I think it might be interesting now to kind of talk more about how the world is sort of changing because of AI. I think we're kind of like some people kind of see it, some people don't. I think an interesting way to start that is you mentioned that there's a lot of academics that almost don't believe in AI. It seems like you're all in on it. You've basically created initiatives to lean into it and you're using it to do research. But there's some people that almost think it's a fad. So do you know what is going on there?
Karthik
Yeah. First of all, I think especially with something like AI, I mean, it's general truth that we can hold in everyday life, but certainly applies more to AI. Right. Multiple things that are seemingly contradictory can actually be true at the same time.
Host
Okay, like how so you know, you
Karthik
talked about some very learned professors. Maybe they're not as much bought into AI. Right. They say, okay, I asked AI this question and it hallucinated a Nonsensical answer. Hence it doesn't work. Right. Hey, but you did that two years ago. So many things have happened since then. But I think the core thing remains right. We got to separate some of the marketing and publicity and certain kinds of people who run AI, AI companies, et cetera, from the actual model capabilities and the scientific value. And sometimes people mix these issues if somebody's overselling something and people fail to see what is actually the most important thing. So yeah, I mean the other way to think about it is in some domains AI is already incredibly impactful. And examples are certainly coding, I would say mathematics, maybe physical sciences like theoretical physics and things like that.
Host
So how is it good in those things? In your domain? Where are you seeing AI just being extremely useful?
Karthik
Yeah. So first of all, why are these particular domains more favorable for AI? It is because in coding and in math especially if the AI did something wrong, it is immediately, you can immediately know that it is wrong. And when that feedback goes back into the model, it can basically correct itself. These are called objective metrics because if it produced a little wrong piece of code, maybe it won't compile so you know it's wrong. Or maybe you run the code and you get the wrong answer. Then the feedback from the output of the model back to its thought generation, if that is very direct, then those are the domains where AI is already very powerful.
Host
So it's all rules. You can just say 2 plus 2 equals 5. You can say, nope, that's not correct because 12 plus 12 is 3, 4. And you just keep doing that over and over until it knows all the rules essentially correct.
Karthik
The thing is, it doesn't need to see all of the rules. That's the beauty of it. It sees enough rules that it is basically able to be intelligent in areas where it has not seen. So that's one thing. And the other thing which has been pretty remarkable in my opinion about AI, at least these latest frontier models, is they're able to compose ideas from different fields, bring it together in a very seamless way, or even within a field, half baked idea here, half baked idea there, they're able to see patterns and merge them. That is very powerful. And the other thing is many people just use AI just in chatbot mode. They don't take advantage of the many tools that can be built around AI models. And if you're working in mathematics, there are infrastructure is called interactive theorem provers. So AI can suggest something that can go into the theorem prover and the theorem prover can basically Use that information and feedback very important things back to the AI. So if people are not using this and they just use it as a chatbot, they're missing so many of these amazing things that can be. So again, we come back to human expertise plus AI plus tools. That is when you see much of the benefit to come out of AI.
Host
So it's not just you say AI cure cancer and it just goes out and does it. That's not possible, right?
Karthik
Yeah, that is not possible. In some domains it is proving mathematical theorems that people hadn't touched before, or people either for lack of attention or lack of patience or whatever. You tell AI, hey, solve this thing. It's actually doing it. But then of course, there are things like cancer and fusion energy and many other practical problems that have to go through the full scientific process, right? You have to, like, if you're discovering a drug and AI gives you something that has to be verified using computation, it has to be verified using trials. So the way to think about it is bottlenecks, right. Whenever the bottleneck is completely cognitive, like math. Math is all cognitive, right? AI can go all the way, I feel soon. But in many problems, the bottleneck is not just cognitive. It is the fact that physics doesn't agree. Or you have to go build this very expensive experiment to test your idea and then give feedback to the model. Right. So, yeah, in those problems anyway, always think of, again, human plus AI plus tools. And in some cases, AI can actually cover a lot of ground, like in math, in some places, pure cognitive intelligence Maybe can address 30% of the problem, but that 30% can be cleaned up by AI, but you still have the other 70%.
Host
So what are some of those other bigger bottlenecks then that we're maybe running into?
Karthik
Like I said, I mean, first of all, I don't want to give you the impression that everything cognitive has been solved. Right. Current AI models still have many limitations. They're. They're very good at taking language and reasoning. They're pretty good at taking math and reasoning computations and reasoning. They're not so good as good as humans at spatial reasoning or physical reasoning. That's where physical AI and robotics is lagging a little bit behind more language and math and these kinds of things. So I feel it's a matter of time before that gets conquered. But the bigger bottleneck in solving very hard problems, like truly discovering a superconductor that I can use tomorrow in this particular application, is when you interact with the real world. You need to actually build something and you have to test it. And that's not something that AI does on its own. Maybe in the future self improving AI can get smarter and cover more of the scientific process. But true discoveries in most fields require you to run complicated computational solvers or run complicated experiments. That's for the moment out of reach of AI. So AI has to work with all of these different things. So there's something called the Amdahl's law, right? So let's say there's 10 units you need to complete to complete your work, and AI completes eight of those units with no time, the other remaining units are going to choke you down. So just because you have identified 90% of the work, doesn't mean all of the work just happens on its own.
Host
So then what are some ways you think that the world is going to change as AI gets better? And maybe there's timelines associated with these, maybe there's not. But what are some of the big things that you're thinking about, especially as you are building out this institute and running experiments and doing research, what are you kind of expecting over time?
Karthik
Yeah, it's a very broad question, right? And like I said, if anyone answers that question with absolute certainty, I think they're probably not being honest, right? So you can only talk about likelihoods. There's a possibility of this happening. There's a possibility of this happening. So under those constraints, I think what is AI doing in the economy and in science and all of these things, right? So I feel like we've built the entire economy around scarcity and friction, right? Scarcity of resources. For instance, if you think of before the Internet, information was scarce, it was a commodity.
Host
You were an expert in how to do something in your town, and you're the only one that knows how to do it. So you make economic value from knowing this thing. No one else knows how to even
Karthik
think about and no one else has even heard about it. Information, it's one example of scarcity of a resource. Information, knowledge, data. And if you think about some of the major inflection points in intelligence in history, like you had language before language, there's no transmitting ideas. Then writing came about, that was a big thing. Printing came about, that was a big thing. Internet came about, that's a big thing, right? And each of these things, those information barriers broke down and certain kinds of things became less scarce, right? But until this large scale AI hit knowledge and cognition, those were still scarce. True intelligence, those were still scarce, right? You had to go to, you have to basically learn from age 5 to age 22 to understand something and to be competent in that field. And now some device that something that you can buy for the cost of a Netflix subscription is actually pretty much getting you not just the information, but the knowledge and intelligence. Right. So that scarcity of certain kinds of things is getting wiped out. We have more abundance of intelligence now. The other thing I mentioned is friction. Like I said, a lot of economy is built around the fact that moving this thing from here to here required somebody to pick it up and move. We created friction. If you wanted to buy a house, you had to go through a realtor. That's friction. You come to a university to learn things in a certain way, to be able to know some things. Sometimes friction is good, actually. Right. But sometimes it was slowing things down. Now AI is removing all of those. So, yeah, so the economy will change in many different ways. The value propositions that we used to place on different things is changing before our eyes. We used to place people who are, say, extremely good at remembering things or extremely good at doing math or extremely good at a particular type of intelligence or a particular type of skill were very valued. I still think there is value to many of these skills, but I think value propositions will change.
Host
And you gave a Code Red to your students. That's how you described it. What was the Code Red you gave to your students?
Karthik
Yeah, so this was about a month or so ago when I brought all of my PhD students in a room, a dozen of them, and said basically models, some of these reasoning models even. I'm not even talking about agents, but agents are helping. They're able to reason through things at the level of my expertise in areas that I am one of the world's experts in. And I had many weekend projects with some of these agents where I have a research idea. And I'm not even describing it completely fully. I'm describing it to some sense. I'm talking to it. And it is doing research. It is exploring all kinds of configurations, it's writing code, testing ideas, coming back with results, writing reports, things that would have taken me four or five months getting done in a weekend. I don't want to give the impression that hence all of my research can be put into a weekend. But many of these tasks are things that I could not give to a second year PhD student at the University of Michigan whom we select. Our acceptance rate is like 5% or something. This is among the best in the world. And if some of the tasks that I give to AI, I mean, AI does as well a job as a second year PhD student. It raises a question.
Host
So what's one of these tasks that AI is now doing at the same level or better than one of the PhD students? Just to give me an example, because I don't know if I know exactly what you're talking about.
Karthik
Okay, yeah, I'll give you a specific example. But before that, I want to say a PhD student is not just about completing tasks, and a PhD is not just a task completion. Right. So I'm not.
Host
PhDs will still exist. Right? Like you'll still exist.
Karthik
PhDs will still exist. Original ideas exist. There is value in training people, and people come up with ideas in a very different way. So there's all of that, Right?
Host
Yeah.
Karthik
So I don't want to equate a PhD student to just executing a set of tasks. But as a professor who wears many different hats. Right. I'm a researcher, I teach, I have a startup, I run an institute, I learn new things, et cetera, et cetera. So my time is very splintered. So generally, if I have a research idea, I would work on it over a weekend, for instance, or a few weekends. And then if it is, some of those will fail because they were bad ideas. Some of those are maybe reasonable. Then I say, hey, what? Why don't you look at this? This is how we used to do research until two, three years ago.
Host
So you would come across something that maybe is worth spending a ton of time on, and you'd suggest the idea to someone.
Karthik
Yeah. And sometimes the students themselves go through that process. Right. They work on something over a month, month and a half, whatever. And they come to me and say, hey, what do you think about this? And I find that interesting. Sometimes their idea is better than mine, but whatever, right? Regardless, going from ideation to actually having that research idea makes sense. So that time, especially in fields like mine, I'm a computational scientist, which means all of what I do, most of what I do, not all, involves basically wrangling equations and computing them and looking at physical phenomena and modeling them, et cetera. So many research ideas now in my field and many other fields, you can test out very quickly. And sometimes as the AI model or the agent is working through these, even if you're guiding these tools through our ideas, as they're working through the ideas, they can give you connections that steer your own thoughts. So, long story short, I called my students to say, until, say, December 2025, these tools were kind of good, and in some areas, not so good. But I always used to talk about These tools in terms of future potential. One year ago it was absolutely horrible. Six months ago it was less horrible. And now it's decent. That is how I used to say. But now they've reached a threshold where some of the ideas that are coming out of these models and the way they're doing reasoning and the way these agents are reasoning it is about as good as leading edge research. So if we are not using these tools at this point of time, then we may be missing out on something. But I also had another part of the code read. You develop a lot of intuition by doing it the slow way, the rigorous way students. There is value in having students not use AI and develop their own intuition and their own thinking. But then there is also a place to use these tools. So I don't know, you can't put things on stone. But I still want students to have original thinking, do things the hard way, make mistakes, learn from mistakes, move ahead. So that's the friction. If you completely remove that friction, then you completely lack intuition on how things work. But at the same time, if you completely ignore these tools, somebody's going to get there for faster than you. So it's a very hard problem and it's happening in all fields. Right. So basically Code Red was to show them how powerful these tools are and be aware of what is happening, but at the same time, not forget to do things thoroughly. So it's not an easy balance.
Host
Yeah. So it's basically telling them they need to completely master AI and also master not using AI at the same time.
Karthik
Yeah. So don't substitute AI. So, I mean, maybe we got to separate the learning phase from the creating phase. Okay. And sometimes there is overlap between these two things. Right. Like, the only way you learn physics is by doing problems. Right. By doing the physics, by working through the problems. Of course, AI knows the answer, but if you substitute. But in the process of trying to solve that problem, there is some friction that's created that basically gives you intuition. So don't skip that. But at the same time, don't be oblivious to these tools. Right. So they can be used in the right way.
Host
Yeah. It reminds me in a way of. I don't know if you've ever come across the CFA charter. For the cfa, it's kind of like a cpa, but for investments. It's like you learn a ton about all these different investment categories. I kind of. The way I kind of. So there's three different exams you have to pass. There's a level one, level two, level three. I Think the level one is an undergraduate degree in finance. Just everything you learn in a finance undergrad, it's just one test, and you got to know all of it. And then the second level is almost like a master's, so you get a master's in finance, and then the third is maybe it's like a PhD in finance, a little bit more theoretical. And at the end of it, you need to be able to say, answer a question. Paul and Linda are in Canada, and they have an investment portfolio in the United States, but they have a couple investments in Bangladesh and France, and it's in local currency, and they want to hedge their bets in Mexican pesos. And they have a kid that's graduating college in 18 years or going to college, and they need to be able to fund the college. And this is the amount in their portfolio, and they want it to grow to this. And you just have to be able to give them a recommendation on what they need to do. And that is the most. I just described an absolutely insane situation. And you need to be able to do all these hand calculations to calculate all these different. What kind of hedges do they need? They need to consider inflation, the different returns of all these different potential asset classes. And at the end of the day, you're never going to actually go through and do all those things by hand. And, like, you might just Google it, right?
Sponsor/Announcer
Or.
Host
Or you would pay an expert to, like, to figure these things out or something. And so. But it reminds me of, you know, I had to go through and learn all this stuff, and it sucked. And I was like, I'm never gonna actually do any of this. But it does give you some intuition on how to think about some of these things. And, you know, it's kind of a lot of the time in investing or finance or business, there's like a spreadsheet that someone's looking at to make a decision, but you need to know what goes into that spreadsheet. It's in the same way of a math problem. It's like you need to know in physics what is influencing that outcome at the end of the day, but you don't actually have to solve it, but you do need to know how it's solved.
Karthik
Yeah, right. So you're absolutely right. It's the same thing. And this complex scenario that you described, I think it's a good example, because by going through that process of doing it by hand and getting it wrong, and your professor telling you, or your tutor telling you, or your colleague telling you, you missed this. That is how you actually get that intuition to go see a problem? Because I think the value shifts from. I mean, intuition is always useful, but I think that judgment that you need to give can be informed differently.
Host
So how, in terms of. I know you said teaching classes is like a. It's a percentage, it's a smaller percentage of what you do. But how have you kind of seen the way that students are learning and how has that kind of changed over time? And then how do you think it's going to change going forward?
Karthik
Yeah, I mean, it's a pretty interesting scenario. Right. So we have had, I would say three shocks in the past few years. Right. One is Covid.
Host
Oh yeah.
Karthik
So that itself was big enough. Right.
Host
I forgot about COVID I know.
Karthik
People went online, maybe high schools were not as rigorous in making them do the math.
Host
And you saw that show up at the university? Oh yeah.
Karthik
I mean, I could see in certain parts of the undergraduate class, I could see Even right. In 2022, the incoming class, I could see some mathematical deficiencies. I would even call it, at the time, I used to call it persistence, like keep going at a problem.
Host
So the kids weren't as persistent.
Karthik
Yeah, I could see a bit of that. And you could say it's been happening over a longer period of time.
Host
When did it start, do you think?
Karthik
I think maybe Google and smartphones probably started that. I don't want you to think people are. Students aren't great, they are. But certain level of mathematical skills, rigor and persistence, especially among the undergraduate crowd, we could see it dropping a bit. And then Covid really was a shock. And just as we were recovering from COVID we got ChatGPT November 2022. But at the time, the models were pretty bad. They were hallucinating like nobody's business, but they could still do a bunch of things. And to me, the bigger one is the last six months where I think especially at the undergraduate level, however hard you set a question or a problem, I think the best AI models can actually do it pretty much in any field.
Host
So any kind of take home test or question situation, kids are just acing everything.
Karthik
No, I mean, I'm actually pretty. I shouldn't say surprised because we are leaders and best or whatever. Right? No, I think the honesty and integrity, I don't question as much. And we have a tradition. For instance, in the school of Engineering, for the past 150 years, no exam has ever been proctored. This professor gives the exam in class and then steps out of class, waits outside the classroom. We've done it for 150 years. And I still see much of that persisting. But without a doubt, students are using AI tools to study and then to replace some parts of their thinking. For sure. And I'm sure, I mean, I'm not naive, we know human nature. Maybe some students are completely relying on those tools, but it's a mix. But what we again end up missing is that persistence and that struggle. You ask how my own teaching has changed. I remember I teach one of a class that students find to be one of the more difficult ones in my department. Maybe the most difficult, but also the most enjoyable. Not just because of me. The subject itself is beautiful.
Host
What's the subject?
Karthik
It's aerodynamics. That's what I teach.
Host
So what's the point of the class? What am I getting out of taking that class?
Karthik
Well, you're going to learn how, for instance, wings generate lift or how to design wings for certain properties. How much power do you need to move an aircraft? And you learn it. You learn math, you learn physics, you learn engineering. And it all comes in a very special kind of way. So it's quite abstract. And some students struggle, but everyone enjoys it. And I teach it. Some of my colleagues who teach their also excellent. So it's a hard class, but students learn a lot and they enjoy it. But anyway, you asked how is changing teaching? I remember six years ago, I give a homework problem and half the class would not even know how to start the problem.
Host
That would be me.
Karthik
They're like, what is this? And I don't take pleasure in torturing students, but I knew it was excellent for learning.
Host
So what made this problem so hard? Can you give me an example of something I might be. I might be getting from you and just. I wouldn't know where to start.
Karthik
Yeah, maybe I teach how. So you've seen these aircraft fly and then sometimes you see these trailing vortices. When an aircraft flies through clouds, you see these vortices that roll up behind the wing. I teach that. I teach how that happens for an aircraft. And then I say in the homework, I would say, here are a flock of birds that are flying in a geese that are flying in a V formation. Tell me how efficient flying in this V formation would be. Without having mentioned many relevant details, many students would not know where to start. I guess with Googling you can get some amount of help. And there are some other problems that I Google and there is no way Google knows anything related. So then students take a few hours to even figure out how to start. But you Learn a whole lot in that struggle. The hour and a half that you put in where you went nowhere is actually where you learn. And now that is completely remote because
Host
you can just Type it into ChatGPT.
Karthik
Whatever question anyone gives in any course at any level, in any university in the world, including the most famous mathematician alive, AI, at least, either will either completely solve it or it can recommend six or seven directions to start, and then you follow. Right. So, yeah, you're suddenly losing something.
Host
So this geese question, I'm just curious now because I'm thinking, what would I even be trying to figure out? How do I solve this?
Karthik
Okay, so basically I told you how, you know, a single aircraft flies and how those trailing vortices look like.
Host
Do you give me the formulas of, like, how to calculate this stuff ahead
Karthik
of time for a single. The lecture would be, you know, going through the derivation for a very idealized airplane. Right. But then you basically use that particular information, and then you treat each of those geese as different aircraft, and they have to support a certain amount of mass to be able to fly. And then you basically create a bunch of these little airplanes and the vortices corresponding to those. And then you'd optimize over that.
Host
And you're giving me the size of the geese in the description of the
Karthik
brain sizing of the geese or something.
Host
Okay, yeah.
Karthik
I mean, maybe that's not the best example of problem that they don't know how to start. But. But there are other more complicated examples. But for some students, even that connection would be harder. But yeah, in general, I think maybe moving away from the specifics, it is in that friction and the struggle that you actually learn. And if that is being replaced, then you don't develop as much intuition or that judgment. Right. Like you described a scenario in your CFA where you have to think about 25 different things. But yes, doing it persistently, rigorously, by hand has a lot of value. And that's being replaced. Yeah.
Host
And then the interesting thing about where not even just AI, but just computers and software comes into play, it's all these little mini calculations you kind of have to make. And if you make a single mistake in any one of that chain, even if you knew it, if you're trying to do this all in your head, it would take so long versus even using a calculator, it's just chains that you know exactly the answer, even just software speeds that up so much. And then with AI, it's basically creating all those software calculator calculations, doing that all automatically for you so then the skill becomes, instead of how to use your calculator to solve a problem, the skill becomes knowing how to use AI and an agent to solve a problem, the meta.
Karthik
Right. But what if one of these agents is actually doing the wrong thing? You have a sequence of these and one of those is wrong.
Host
So you still need to be able to understand what the agents are doing.
Karthik
Yeah. The other way to think about it is if everything can be automated, then what is the value you bring to the problem? Anyone can do it.
Host
Yeah. So do you think we're going to have to start learning how to create and manage and run agents? Like, that's a lot of what education is going to be turning into?
Karthik
I actually don't think so. Like, did we need a lot of education to use ChatGPT 3.5 when it came? No. Right. But you needed a lot of common sense to be able to use it effectively, you know, so the other day I. This was two weekends ago, my student came with a brilliant, brilliant idea. It's going to be published sometime soon and I didn't quite get the intuition for it. He was certainly understood more about it than I did. And then over the weekend I was wrangling with that and I asked Codex, the OpenAI agent, to. I described a problem, I gave it all the context, it wrote code, and then I wanted it to create a visualization to help me understand what was going on. So it did everything beautifully. But then the visualization was kind of off. The angle at which it was showing me was off because it had no spatial reasoning. So I asked it a bunch of things. Hey, I don't do. This does not make sense. There is no more clarity and it would make it worse. So then I said, help me help you. I said, describe the view that you're showing with some numbers and help me and make the figure interactive so I can rotate it whichever way I want. And when I rotate it, those numbers would change. So let me tell you what the best angle is by taking those numbers and giving it back to you. So anyways, human and AI, basically, without getting into details, I think people say, oh, the great education is how to use AI. No, I mean you can use AI, these things. Anything that seems like a little bit of activation energy, like, oh, not everybody knows how to use an agent. I think in a couple of months it's just going to be clicking a button. So it is more like, how do you synthesize what you know and how do you use your common sense to wrangle with AI? So I think that's more important. Learning how to use AI, it's never going to be a big issue.
Host
Yeah, well, so how does the role of the university change then, with education becoming either less or more important? I don't actually know the answer to that. But how do you think the role of universities is going to change?
Karthik
Yeah, so I mean, it has been changing. Right. Over the past few decades even. Right. We used to be gatekeepers of information knowledge nobody else had. And then slowly, the Internet and large courses and all this content, we're eroding a little bit of that.
Host
So this would be like when MIT put the classes online.
Karthik
Yeah, So I would say, I mean, that is one of the most seminal moments. I think people should talk more about that moment. Right.
Host
So what was that? When Was this? About 25 years ago?
Karthik
Yeah, around the early 2000s, MIT took all of their class notes and class videos and lectures and homeworks and just put it on the Internet for free for anybody in any part of the world to access.
Host
So you could, in theory, anyone who's enrolled at MIT and they had to go to all these courses, anyone in the world could just get the same
Karthik
stuff, same material, same lectures from the expert in the world. The videos, homeworks, everything. You know, that's amazing if you think about it. Did it change the world enormously? I mean, for certain people it did. It's an amazing thing. But again, so that was passive, right? It was not teaching you. It was basically revealing information. So now it's basically giving you cognition and knowledge. So I think a lot of the things that universities used to do, I think will be less valuable than before. As keepers of knowledge. You had to go talk to this expert to know anything. And you had to go to this famous professor in Michigan medicine to learn about that particular way of thinking or that scientific method. And professors who had maybe your Nobel Prize had a different flavor to it. So a lot of these things are now pretty much encoded. I mean, not all of it. A lot of it is now encoded in US Skills MD file or identified in some way. The models know it. So we have keepers of knowledge for thousands of years, and now suddenly this box that you pay $20 a month for knows maybe not all of it, but most of it. Right. But still, universities have a big role to play because what are the other values we bring in? Right. So just having access to content doesn't mean people learn.
Host
I've never looked at any of those MIT videos. I mean, I think I've seen someone talk about it. And went to the YouTube page, but I've never actually watched any of them.
Karthik
I might still use them once in a while. And I put my notes on, some of my notes on the web and in YouTube. And I get a kick out of people watching it and learning and sending me emails. But in any case, so just because the materials are out there doesn't mean people learn. And just because knowledge is accessible for $20 doesn't mean people extract the most. So universities still add value in bringing, say, young people of a certain age together in a certain environment and say, putting 25 people in a class and introducing, I mean, maybe the right way to say it is introducing friction in a certain way, introducing deadlines in a certain way, exams making you think. So that's one thing. And of course people learn from people. And then of course it's not just about the lecture room. There is other kinds of mentoring, personal mentoring happening, especially at the graduate level. And you have access to very specific facilities, right? Like, I don't know if you know, the University of Michigan has the world's, at least the nation's most powerful laser is right here. It's called Zeus.
Host
What do you use it for?
Karthik
To study theoretical physics, you know, to study, say plasmas or to study atomic properties, material properties.
Host
Like are you cutting things with it or are you just pointing it?
Karthik
You point it towards, say a certain material and the impact of the laser on that material basically can change some properties. You can get the spectrogram and you can do a lot more with it. Right? I'm just giving you one example. And a lot of theoretical physics, you can do theoretical developments based on that. So that doesn't exist inside ChatGPT. Right. And again, my work is computational, but many of my colleagues have physical labs where they're building flexible aircraft. Aircraft that we fly on are not as flexible. We have one of the best battery design facilities among any US University right here. So these are specialized facilities. So anyway, to summarize my answer to your question, bring people together in a certain way under certain constraints that they cannot sit at home and just get these kind of things. So there are still a lot of. And then of course, not just learning, but creating new things. So universities, many of the discoveries and innovations directly come from university labs. And the president of Arizona State, he has a very nice one minute video where he picks up this iPhone and says There are like 800 technologies here that were developed at a university lab, right? And what Apple does, of course Apple does add value, but they put things together in A certain way. So I think learning, creating and introducing certain kind of friction, exposure to types of ideas. So I think all of those still belong in a university and will belong in a university. And then the last thing, which maybe now becomes even more important, is credentialing, filtering universities, for better or worse even to get into a place like Michigan, I don't know now what the acceptance rate is 10%. And then you come in. I mean, I'm not saying that's a good thing. Right. We should be more accessible. That's a different topic for a different day. But after you come here, you go through a program where you're credentialed, you're given grades. So I feel those will become more important now than before
Host
because they prove that you know the topic or that you have spent time running through the motions of learning.
Karthik
Yeah, first of all, it was. Yeah. Many times people recruit from top schools even for the reason that, oh, they got into the top school, so they must be good at something. So that is one. I would say implicit credentialing. Again, I didn't say it's good or bad. It is what it is. And then, of course, you go through the program at various levels of rigor, the university's credentialing. But I would say great inflation is probably going to stop now.
Host
I've seen those charts where it's like the average grade used to be like a 2.3 GPA and now it's like
Karthik
4 or something like 3.6 or, I don't know, something like that. Harvard had a big ruse about too many people are getting as. So I feel these kind of distinctions will grow and we will take evaluation and credentialing more seriously. Also because there are easy ways to complete work.
Host
Yeah. Do you know what was going on there with great inflation, from your perspective, is the one who's giving out these grades. Why was it happening?
Karthik
Many things. So again, this is one of those things where people usually have one answer that I don't know. Professors are catering to customers or whatever. If you are the students, they would say we are working harder. One of those scenarios where all of these things may be true. Certainly you can agree that students are better prepared to come into college now than they were, say, 20 years ago. That's one factor. And then, of course, I wouldn't say standards are dropping. But I think in most universities the expectation is if I work hard, I get an A. Right. And many times that is actually correlated with outcomes. You really work hard, you're actually learning. But it's not always. And then. Yeah, I mean, again, I don't want to be an idealist. I'm sure there is a little bit of. I'm paying $60,000 for a year for my degree. I'm not saying, hence people give better grades. But if you dismiss that completely, then you're being naive. So. So as with any complex thing, there are many things that come together. I mean, something can be explained in many different ways. Like I said, students do work hard. Not everybody. They do work hard. And it's not like every professor is just giving out A's for free. But yeah, all of these are conflicted for sure.
Host
And you think that this is going to. The inflation's going to stop. You think it will come back down or just level out?
Karthik
I'm seeing a movement where things are either leveling or beginning to come back down a bit.
Host
How do you pull that off? Like bringing it back down because you have to be stricter or something.
Karthik
Or like Harvard actually pulled it off. I mean, I think two years ago, three years ago, they actually showed the grades inflating. Like 70% of people who went to Harvard got an A or something like that. Some ridiculous number. And then the professors said something and then they brought it down a bit. They plateaued and then they brought it down a bit. But now students are complaining. Students are like organizing and saying, oh, it's mental stress and we are working so hard. Again, it'd be stupid to dismiss that claim. But it'll also be stupid to completely overrule what the professors are doing. So as with many things, maybe both are true.
Host
And so I feel like for a lot of people, the reason you go to college and go to university for most people do it because they want to get a job after. They think it'll increase their chances of success to get the credential, which makes it easier to get that interview, the stamp of approval. What do you think is going to happen to the job market over the next decade? Or maybe it's happening today? What are you seeing, especially as it relates to how AI is changing things?
Karthik
Yeah. So again, nobody can predict, but I think there are some things that are true. First of all, you are seeing in tech industry about 100k jobs, 100k layoffs or so in the last couple of years. And I think in 2026, we're already seeing 100k. Right. So that seems like a large number, but I think we have about 7 or 8 million tech workers.
Host
So in the grand scheme of things, it's like 3%, 2% or something like
Karthik
that, 1% or something. So I'm not dismissing, I mean it's creating a lot of pain, et cetera, et cetera. So I think there'll be, especially in tech areas, I feel there'll be more of these layoffs in the next few years. But I don't think it's going to be as dramatic or drastic as what some of the talking heads are saying
Host
because some people would say based on what they're saying, you'd think massive loss of jobs, no one's able to work, AI just replaces everything and everyone's unemployed.
Karthik
Yeah, I think that's certainly probably going too far, at least in the short term because the economy is built in a certain way that these kind of impacts will take longer time to penetrate. But if you want to ask me what's going to happen in the next two to three years of which I have a reasonable handle off, not a great handle, I don't think 20% of the population will be unemployed. Maybe it'll be, what is it, 4.6, 4.5 or some percent unemployment right now, maybe that'll go by a percent more, which is actually pretty bad.
Host
20% increase in.
Karthik
But I feel instead of having huge layoffs, I don't think there'll be massive layoffs. Maybe there'll be 100k ish here and there. But I find it more concerning for fresh graduates to get into jobs. And the code red that I described earlier has something to do with it because many of the entry level skills in certain domains that actually can be automated but again if you don't have the entry level skills you don't actually become an expert. So it's a little bit of a chicken egg situation. But I do believe the way the economy has been built there's going to be a slower impact. Plus people talk about all kinds of things like oh, everybody can have a billion dollar startup, where is the market? Who's going to be buying your thing?
Host
You're saying the one person billion dollar startup.
Karthik
I'm sure there'll be a few 1 billion, there'll be a few 1 person unicorns. I think there is already one to my knowledge.
Host
Did you see the telehealth company?
Karthik
Yeah, that's right.
Host
I mean there's some interesting stuff that's came out that they may not be fully compliant in all the things they're doing.
Karthik
Yeah, I have no doubt about it. But anyway, I don't rule out that there may be a few single person unicorns and One of my colleagues, Jerry Davis in the business school, he has very colorful remarks, thoughtful too on 0% unicorns, like completely an agent running the whole thing. And I don't think it is completely out of question, but those will be exceptions. I think the economy is set up in such a way to blunt huge disruptions and people, when they evaluate a technology, they always think linear, but things are very non linear, they can saturate. Then maybe they go. So yeah, I think in summary, I feel people who already are in good jobs or most jobs, AI will help them be more productive and maybe they will do more. But then I worry more about entry level jobs.
Host
It's just harder to get that first job because maybe instead of someone who's a manager hiring a new person on their team, you're just able to use AI and software.
Karthik
Again, not in every profession and not completely, but again, you have to be naive to think that it won't have an impact.
Host
Well, I think one of the recent guests on the show, we talked a lot about the US healthcare system and he just talks about how we've had almost these kind of job programs throughout the course of American history. Manufacturing was kind of built in jobs mechanism for employing people. And healthcare is kind of that way right now where there's a lot of different cities and regions and states where some sort of healthcare system is the largest employer. And it's just people who sit at a desk and you are doing things in the hospital or you're a nurse. There's a lot of people who help bring people around the hospital. They do some admin work. So there's a lot of these cases like that where those sort of exist to employ people. And even if AI makes their jobs more efficient, the government is kind of funding it. And the government's just not going to say let's not have this job anymore. So there's kind of like this multiple pieces of friction here.
Karthik
Yeah, certainly in the short term I agree 100%. Right. But I cannot look five beyond five years beyond the horizon and know how nonlinear the impact would be. So there are a lot of destabilizing things. But you're right, there are many jobs where I think the way job is structured and the way incentives are structured I think will protect people a lot. I mean there will still be layoffs, but I don't think it's going to be as catastrophic as Dario Amadeus talks about, at least in the near term. But as I said, I do worry about entry level jobs. I think that's the bigger Concern.
Host
So what advice are you giving your students when it's like, hey, if you want a job, here's what you have to do.
Karthik
It's very hard advice, but the advice that I give is whatever you study, just get more rigorous about it and add some value.
Host
So just like get really good.
Karthik
Get really good at what you do. I don't tell them, oh, use an AI tool. To me, this part is more important. Going through that process, building that intuition. And then I think using the tools can maybe just out of necessity or common sense. But then I'm of a certain age where many of my friends have now kids who are in like 10th grade and they're thinking about college and what should I do? People thought computer science was a sure shot to an amazing career.
Host
Yes. Getting. Becoming like a millionaire, Coasting forever. Heaven.
Karthik
Yeah. So anyway, so I tell them it doesn't matter what field you're in, even if you study computer science. Computer science is not just coding or software engineering. There is theoretical computer science. There are so many other things beyond coding and software in computer science also. But I think more fundamental advice, especially if they come to me with open advice, what should we study? Maybe I'm biased. I would say physics, chemistry, biology, mathematics,
Host
those are all things AI can kind of do really well right now. Right?
Karthik
AI can do very well to a certain degree. But most of the unsolved grand challenge problems in humanity involve those fundamental sciences. Right. And then there are also questions. What is the nature of life? Where did we come from? So nature of reality? So these are not things that we can answer right away. So there is, first of all, you study things for the intrinsic merit, but you have to do it really well and really build up your basics. And those problems are never going away. You're not going to be. Somehow tomorrow somebody has this unified theory of physics and that's it. It's not going away. But the second advice I give is, yeah, get really deep, but also know how to synthesize things in the right way. But don't skip. This part is the most important. You need to know something really well so that you actually have expertise and not everything is aiable. Sure, AI can do your math, physics and chemistry, biology homeworks, but yeah, go to those fundamentals. I feel those are hard skills that will matter as long as you're able to synthesize information across different disciplines. Yeah.
Host
Who? One advice I always give people too is you probably want to niche down a lot more than you think. So if somebody just says to me, I want to be a scientist or something. And I'm looking for a job as a scientist that could be anything. But if you tell me and you might think that this is bad, as I say, I want to be a scientist that works on paper cup strength, being more durable, holding water and the strongest paper cups ever in the world or something like that. That's just a super specific problem that you're really good at. If I do ever come across something that requires that skill, you are probably going to be the best person at that specific problem. So I may think of you. It's almost like when you're. We think about as a sports team, like if you're a soccer coach and you're trying to make your team better and you're looking for players and there's like all these people at the tryouts and you think about who to add on the team and there's just like this person that's really good at throw ins. Right. Like throwing the ball from the corner and like they may not be the best player at other aspects, but they have an insane ability to like I don't even know how you be good at a throw in. In soccer. They're just extremely good at it. You might make the team because you're just, you're good at that one skill. Or like corner kicks. You see a lot in soccer too where someone is just so good at the corner kick aspect and you get to. You made the team because that's what you're good at, right?
Karthik
Yeah. I think the way you're describing it, maybe the ideal profile is something like a V. So somebody who knows something about everything is just broad, it's like a rectangle. Right. And then somebody who knows only one thing is just like this very deep.
Host
I think it's like a T or. Yeah, that would be like an I.
Karthik
Yeah, maybe a T is you need to know obviously enough about how things connect, but you need to be really deep at some things. I don't think that value is going to go away. But all of that said, it also won't be simple. Like I said a few minutes ago, just because you're so good at something and you're very smart, that alone used to be enough to make it in the world. You were good at math or you knew this really well. I think those skills are still useful but not sufficient. So I feel as a scientist and in general also the value proposition will shift to people who solve real problems. So maybe instead of saying here is an idea that I generated, maybe the value of that Idea itself is probably going to go down to zero, in my opinion, because AI can generate ideas. Maybe not all of them are good, but some are good. Cost of generating ideas is almost zero now. But to be able to take that and then keep going and going and going and actually solve a problem that people care about, I think that's where the value proposition is going to be. Earlier, it used to be, he's smart, he knows math, he can do counting in his head. He knows this. He can remember things. He or she, they. And yeah, now it's like, okay, so what? What are they doing with that skill? I think, yeah, so going all the way. I think that's what would value propositions will go towards.
Host
And you've actually seen. This is like totally different topic. You've seen two different college basketball national championships as like, firsthand is member school. So what is that like?
Karthik
Oh, amazing. So, yeah, I'm a sports junkie and certainly enjoy college basketball more than most sports. Not all, but it's in my top two or so.
Host
So what's your top sport?
Karthik
Top sport is actually soccer.
Host
Oh, really?
Karthik
Yeah, I'm incredibly passionate about it.
Host
That was actually a lucky guess then when I gave the soccer.
Karthik
I see. Anyway, so when I was a graduate student at the University of Maryland, we won the national championship in men's basketball and women's basketball.
Host
Oh, in the same year.
Karthik
Not the same year, a couple of years apart. And we also had the best stretch in history in our basketball history. In women's and men's. The football team was good, but yeah, the distinct night in 2002, I remember when we won the championship, it was insane. It was insane. And I got to experience it again.
Host
At Michigan. Yeah.
Karthik
Yeah, at Michigan a few days ago. And you know, I actually. Couple of my professors, colleagues, we were all in University South University, along with all these writing kids. I'm sure they didn't know we were professors, but it was amazing to me. Of course, being in a school where you won something, it's a very hard accomplishment. Right. 64 teams, straight knockout. It's a big deal. But the other aspect is sports brings out certain kind of emotions that many other things don't. And what other event in life do we see? 10,000 people completely happy, out of touch with reality, deliriously happy, forgetting all those things that are happening in the world, all celebrating about a thing. Right. So, yeah, to be part of that is amazing. So, yeah, sports evoke certain kinds of wild emotions and sometimes I let them go wild.
Host
Yeah, it's kind of one of those things. I've seen those videos of robots playing sports and people are like, oh, soon they're going to be better than us at sports. I'm like, I don't really want to watch a robot play sports. I don't know if they're. They will ever fully replace humans playing sports, will they?
Karthik
Yeah, yeah, I think again. Well, no, I don't think. I mean, for the past 20, 30 years or so, computers are better than humans at chess. Doesn't mean, you know, we still.
Host
People, like, stream them live stream playing chess online.
Karthik
I think chess is more popular now than 30 years ago when they beat. When computers beat humans. I mean, there is some human pleasure involved to see people compete.
Host
Well, there's this concept. I think it's called the Bionic Games. Have you heard about this? It's basically how steroids and certain drugs are banned in the Olympics. It's like the Olympics, but you're allowed to cheat, you're allowed to take certain drugs and you're allowed to have. I'm assuming you'll be able to have a robotic arm that. Let's say you lost your arm in an accident, you get a surgery to replace it, and that arm is somehow stronger than an actual human arm and you're able to do better tasks or sports or games. So I think that's kind of a thing that's maybe come in.
Karthik
Yeah, there is a market for everything. One of the most popular sports in the world is motor racing. That's human and machine.
Host
So we'll probably get new sports. That's probably what's actually going to happen is there will be new sports that are created.
Karthik
Yeah, that's very possible for sure.
Host
And I think you actually told me one time you almost died once hiking in a national park. What happened?
Karthik
Yeah, that's one of our passion. My wife and I, we travel a lot and at least there was a time when we used to do some interesting trekking and hiking different parts of the world in Argentina and here and there. There was this one time in the Grand Teton national park when it was a snowy day and somehow they gave us the pass to climb and pitch camp. It was maybe an overnight camping trip and it was a beautiful setting. My wife and I were in the initial half an hour of the hike, a ranger was coming down and he said, I'm quite shocked that they gave you a permit because weather conditions are not so good. And then we said, yeah, I think we'll be okay. We've done a bunch of hikes and then he said, don't make me come and rescue you up top. Yeah, nothing will happen. But then he said after a while, you might not see the trail, so just keep to the right of it or follow the footsteps. I don't know. He said some vague things.
Host
So it was snowing while this was happening?
Karthik
It looked like it was going to snow, but I'm sure at the elevation it was snowing and he was coming down. So anyway, we kept going, and then there was a light snow, and we enjoyed it. And then the snow got heavier, and then the trails disappeared. And whatever instruction the ranger gave us, we were not able to follow.
Host
Oh, no.
Karthik
I think I fell into a hole. Was not too deep, but I fell into a hole. Then my wife fell into a hole,
Host
like, right next to each other.
Karthik
No, within maybe 20 minutes of each other.
Host
Oh, so you fell and you got out.
Karthik
Yeah, yeah, yeah.
Host
Okay.
Karthik
And I think some of our backup socks got wet, whatever. And then it snowed heavily. And then somehow we found something that looked like a trail, and we pitched camp, and it was pretty high up, and it snowed some more. And then because we had those falls, socks were wet, and the socks that were inside our backpacks were wet. And it was super freezing. And my wife's. Especially. My wife's hands were going numb. I think few more hours, she would have probably have lost some things. And then the coyotes start howling, and she's like, I don't want to be. I don't want to die being eaten by coyotes or something. And this was. I. When was this? This was probably like 2005 or 6. Cell phones were not very good, and we couldn't call anybody. And we kept trying and trying, and somehow I reached. This was in Wyoming, and somehow I was able to reach my friend in College Park, Maryland, and I think he was able to call the rangers, and then they came up with some very hot stuff, and they wound up our tent. And it was actually the same guy who said, don't make me come and rescue you.
Host
Oh, geez. How late? Like, was this the next day? Or.
Karthik
Like, it was probably midnight or 2am or something like that?
Host
Jeez.
Karthik
Actually, no, it was little. Yeah, it was around that time. He said when he got the call, he was having dinner with his wife. So anyway, so we've had some adventures. I would say.
Host
Geez. Any upcoming trips planned?
Karthik
Yeah, I mean, we have some. Nothing big on hiking. We have some trips to Japan and, you know, Europe and things like that,
Host
but more laid back.
Karthik
More laid back. Yes. Yeah.
Host
Well, this has been awesome. Thanks for taking the time to chat. It was a lot of fun.
Karthik
No, it's a pleasure. We touched upon so many different topics and you were certainly a good host.
Host
There'll be a lot of data out there for the LLMs to train on. Maybe we can teach the language model something.
Karthik
I don't know how much real insight I had, but I'm sure I compressed a few things. Yeah.
Host
Well, it's a lot of fun.
Karthik
Yeah. Thank you for having me.
Host
And thank you for listening.
Sponsor/Announcer
Thanks again to this episode's sponsors. Flex Upgrade to Flex Elite with the link in the Description and get $1,000 after your first 10,000 to spend numerl Put your sales tax on autopilot@numerl.com and amplitude for AI analytics. Just ask Amplitude if you enjoyed this conversation. Please like comment, subscribe and share this episode with a friend who still thinks AI is just a fad. Make sure to check out the back catalog of over 100 episodes with founders of companies like Robinhood, Sweetgreen and Mercury.
Host
Tune in over the next few weeks
Sponsor/Announcer
for guests like Eric Israel and Metropolis,
Host
the parking company that's quietly pulled off
Sponsor/Announcer
one of the most successful versions of the AI enabled roll up strategy that
Host
everyone was talking about last year.
Sponsor/Announcer
And Jim Balosic at Sendcut Send, making custom metal parts as one of the pioneers helping to reshore manufacturing back to the us if you don't want to miss any of these, subscribe to my newsletter. The Split linked in the description to get each episode plus a transcript emailed directly to your inbox every week. Thanks again for listening.
Host
See you next time.
Episode: How Agentic AI is Reshaping Science, Education, and the Economy | Karthik Duraisamy at the University of Michigan
Date: April 17, 2026
Guest: Karthik Duraisamy, University of Michigan
Main Theme:
Exploring how agentic AI and open-source agent frameworks are transforming scientific discovery, university research, education, and the broader economy, and what that disruption means for individuals and institutions alike.
This in-depth conversation with Professor Karthik Duraisamy dives into the latest developments at the University of Michigan’s newly announced Institute for Agentic Computing and the broader movement around open-source agentic AI frameworks like OpenClaw. The episode also explores how agentic AI is already transforming scientific research, the higher education landscape, and the job market, giving practical insights for students, researchers, entrepreneurs, and knowledge seekers in the age of intelligent agents.
[00:23–06:19]
"You had chatbots... and over the last year they've become extremely powerful... But then you have agents that basically act." —Karthik [02:47]
[08:18–14:25]
"[Michigan] covers all areas of human activity... IMSA is pretty much all of the 200 would be in the top 10 of any ranking you can imagine." —Karthik [11:23]
[15:23–22:53]
"What do cricket wings and Bach chorales... have in common?" Agents as domain experts found synergy leading to a new material discovery. —Karthik [15:29]
[25:31–29:31]
[29:55–36:17]
[40:41–55:59]
[63:21–83:20]
[82:11–95:45]
[92:55–99:13]
[99:13–106:45]
| Timestamp | Segment | |-----------|----------------------------------------------------------------| | 00:23–06:19 | Institute for Agentic Computing, OpenClaw, agentic frameworks | | 08:18–14:25 | University structure, role of institutes, interdisciplinary impact | | 15:23–22:53 | Science demo: agent-based material discovery, research process transformation | | 25:31–29:31 | Institute structure, open-source model, global involvement | | 29:55–36:17 | Professors as startup founders, institutes as VCs | | 40:41–55:59 | AI in science: skepticism, power, Code Red for students | | 63:21–83:20 | Education shifts; university value proposition in the AI age | | 82:11–95:45 | Grade inflation, credentialing, job market, skills advice | | 99:13–106:45| Human pursuits: sports, meaning, adventure |