Mayank Kejriwal of GRAIL: 44% of Research Time Is Paperwork — He's Building AI Agents to Kill It
18VC Podcast · Episode 21 · October 9, 2026
Guests: Mayank Kejriwal
Source: https://www.18-vc.com/podcast/mayank-kejriwal-grail

So just an opening. Good morning, good afternoon, and good evening to our audiences. It's a huge welcome today for our guest Professor Mayank, to join us today. He's a research associate professor at USC Viterbi principal scientist at USC's Information Sciences Institute, and co-founder and CEO of Grail. I think that's not really a common case that you're both a professor, researcher, at the same time a founder of a startup, and we are just super interested and excited to have you here.

Thank you. And this will be our episode 21 of our 18VC podcast. So for our audiences who may not be that familiar with us, we are a student-run podcast featuring students, young founders, researchers, as well as investors who's willing to back the smartest people around the campus. And the hosts today will be me, that's Philip, and also Corbin. Yeah, maybe you can briefly describe a little bit of how you get in touch with Professor.

Yeah. So I'm Corbin. I was considering where to go to grad school. I'm doing my master's in computer science, and part of what drew me to USC was Professor Mayank. He has a research group, Artificial Intelligence and Complex Systems. Yes, correct. And he's doing a bunch of interesting research there. And so that's kind of a big part of what attracted me to go to USC.

That and being in LA ain't a bad selling point as well. But, so yeah, that's how I first found out about you. I saw you initially give a talk at the Artificial General Intelligence Conference. I think it was in Stockholm, right, that you spoke. Yes, it was. Yeah. Yeah. So that's first when I came to learn about you, and I thought your topic was open world learning and looking at how AI operates in open world environments.

So Yes I was like, wow, this guy's really smart and seems very personable. So that's firstly how I found out about you. Great. So your work is getting out there. Yeah. Your conferences and your talks, it's getting out there. Yeah, that was a fun conference. So yeah, I'm glad you enjoyed it. Yeah, absolutely. And I think, yeah, we can start a little bit from the early days of, I think you started immigrate from India to here.

Yes. You did your undergrad at UIUC, and then you do your research and PhD, and then you come to USC as a professor. Yes. Can you just briefly walk us through that journey? Sure. When I look back, it looks like a very traditional journey. Many professors do their PhD right after their undergrad. They get research experience. They decide whether they want to go into academia or industry.

I think these days the line is very blurred. You can go to work for a research lab at OpenAI, or you can become a professor, and there's great research happening, sometimes more so in industry than in academia. But when I was graduating, it was kind of a choice between the two, and I loved research. I had industry positions lined up. I had some great internships along the way and I seriously considered it.

I even considered going to Wall Street. I interviewed on Wall Street even. But in the end, I could not give up research, and I think it was because of the creative nature of research. We get to solve really deep, hard technical problems and there's a lot of freedom. It's hard, but it's rewarding, and you get to work with really good students on this great campus of ours.

And I've been at USC for 10 years now, so no regrets. Yeah. Yeah. That's amazing. And I think I do remember you also besides your PhD degree, also get a MBA from Marshall. Yes. That's another interesting part, because most likely you are on maybe computer engineering. Also, I see you doing knowledge graphs. Yes. How do you also get interested in the business space?

Clearly I have too much time on my hands, right? So, I mean, it was an accident really. I come from an entrepreneurial family, and I knew that I wanted to become an entrepreneur myself, but I didn't have the right tools. I felt that I needed to learn how to communicate better to really understand how business works. And I do believe that the best way to do business is to actually do a business.

You cannot learn from books really how to do business. But the MBA is a very practical degree. The program that I went for is called the MBA for Professionals and Managers, and it's meant for working professionals. So everyone in our cohort was a working professional. They all held down a full-time job while they were doing the MBA during weekends and evenings.

So I remember that we would have classes on Saturday. I was in the Saturday cohort, so I would wake up at 6:00 AM in the morning and come here to the USC campus right by 7:00, 8:00 AM, and we'd be here all day. We'd have classes all the way until 5:00, 6:00 PM. And a lot of us became very close friends. Some of my closest friends are not actually from my days at UT Austin, but from the MBA program.

I'm still very close to them. Because you're in this one room for nine, ten hours a day. Right. You get formed into groups. You're with your group the whole year in the MBA program. So it's not per class, but in the first year of the MBA program, you stay with the same group in all your classes, and so you become close friends. Oh, wow. Yeah. It's incredible.

I think that it was just the right opportunity presented itself, and I had help, financial help from USC under an incentive program to do that as well. And my supervisors and people I worked with were very supportive. They were like, "This is really unique. We need more people who can do business development, and let's see how it goes." And it was an incredible experience.

Yeah. Yeah. Definitely. Just a quick follow-on. Is there any experience or connections you made during that journey that kind of help into the startups you're doing right now? Is there any inspirations from there? Yeah. Oh, yeah. Colleagues. Just people that I met there, right? When I first created the startup, I remember that I had conversations with a whole bunch of them.

We reconnected after a few years, and it's strange that you don't talk to some of them for a few years, but then you just reconnect right off the bat. It's incredible. It's a lasting friendship. And a bunch of them went into the startup world themselves, right? They had that aspiration, so they went into it. Some of them went to private equity, some went to VCs, et cetera right?

And so I feel like that network, the Marshall network, is very strong, and I just have to reach out to the right person and almost always they respond. Yeah. That's the Yeah charm and amazingness of the Trojans family. Oh, yeah. It's real. Definitely. I mentioned before that what first drew me to you was your research group. Can you talk a little bit about that?

You have a research group at USC. Kind of what historically have you done? Maybe what future research, where you see kind of the next stages of your research group. Yeah. So when I started, the research group was much more narrow. It was a group on knowledge graphs, which is this area in artificial intelligence that has to do with connecting different pieces of data to build better search engines.

For example, Google has a Google Knowledge Graph that helps it to do much better search, but also e-commerce engines like Amazon, et cetera, all have knowledge graphs. So that was my area of research. That's where I got my PhD. I wrote a book on that. And that's what the research group was all about. And so we applied knowledge graphs to fighting human trafficking.

Our work on that has been transitioned and used to rescue actual trafficking victims. It remains the work that I'm most proud of. Incredible. And we've continued to do that work for 10 years, working with federal agencies, doing research on human trafficking. There are undergrad students that have written papers on that with me and learned about it.

So over time, I came to realize, I would say around COVID, I think COVID was maybe the impetus for this, where we started doing research on COVID, not medical research, but things like how are policies affecting vaccine hesitancy, for example. Just questions like that. I was very neutral on all those questions, but I was just interested in what the answers were because it was a really fascinating time, right? The whole world changed during that time.

And so I was studying it, right? And then I came to realize that, okay, I have much broader interests. I'm interested in network science. I'm interested in social media analytics. I'm interested in policy. In some ways, it was too broad but at the Information Sciences Institute we are very interdisciplinary and transdisciplinary so we work on different areas in different areas at the same time.

And I had funding to work on all of those problems, so I rechristened the group as Artificial Intelligence and Complex Systems, or AICS, because I realized that what I was very interested in is complex systems. Now, there is a meaning to the word complexity. Complex is not the same as complicated. A complex system is one that has many interconnected parts, and so you have to study the system as a whole.

You cannot take out one part and study it in isolation. You have to study the whole system. It's very context-dependent. Very context-dependent, very user-dependent, lots of data and empirical science and all of that. So very interesting to study that. And what I'm interested in is how do we apply artificial intelligence to make more sense of complex systems.

And so we work on problems that lend themselves to that. To public health, to human trafficking, disaster response. AI for science, which also connects to Grail, right? So how do we build AI scientists, right? What does it mean for an AI to do science or to do something original? Right? So these are all problems. Sometimes it wades into philosophical territory almost.

But we are very rigorous. We do empirical science. It's very data-driven, very systems-driven. The real-world impact is very important to the group. Yeah. Yeah. I would say there's a lot of hype around AI, but it is an entirely different problem to understand how do we apply these breakthroughs in AI to actually make meaningful differences in systems because Yes you know that And understand systems And understand systems, of course.

Yeah. So usually with your group, is it first when you're looking at applying AI to a complex system, do you try to represent it as a knowledge graph? Are you still using knowledge graphs in that way? We still use knowledge graphs. It has use in healthcare in particular, where we have all this data, let's say from electronic health records. We're extracting risk factors.

Right. That's a project we're working on. Or we even do projects where we study NIH-funded projects to find out who is getting funded. Are the same institutions getting funded all the time, or are we seeing more equality across institutions? Are we studying diverse topics? So this is called science of science. It's like how does science work? That's a whole another field in itself that we also look at.

So what I would say is we are not married to any particular technique. I think that is very unique about our group, which you don't often find in many other groups, where they like to stick to one set of techniques, right? And they only apply the techniques to problems where it makes sense. And if the problem is such that you don't apply that technique then they don't tackle that problem.

They leave it to another group. But we are very problem-centric. So when we work with other people, and almost all our projects are collaborative and involve domain experts in health or human Yeah trafficking and so on, we always try to understand the problem, and then we try to apply the right technique to solve the problem. So that might be a knowledge graph, or it could be something totally different.

Yeah. Yeah. Yeah, I think that is important, right? Because certain people, they almost teach and use what they know. Yes. And if they know a certain thing, then they try to make the problem fit to That's correct fit the solution to it. And it's like some problems are just going to take a variety of techniques or maybe just this other technique you might not even be familiar with.

So That's right. Yeah. So I was going to ask you as well, just a little bit about AI and you've done a lot of research with it and large language models specifically. So I think it's hard for people to get an intuition what LLMs are actually good at, maybe not good at. They feel astonishingly good at certain tasks. Some things they feel like they're feigning Yes competence.

What's your sense of where we're at with LLMs in 2026? It's a great question. So I think one extreme is that the LLMs are very close to AGI, artificial general intelligence. I think we are still a little bit away from that. I think it's on the horizon. So now I do feel like it's a question of when, not if. I think that the LLMs in some form or the other, maybe a bigger version, maybe a better trained version will ultimately achieve AGI.

I've come to believe that. I didn't use to believe that before, like four or five years ago. Like many other people, I was like, "Okay, this is really cool, but how far will it go?" And in the past, the neural networks have always hit a wall. All these neural technologies, they have been useful but at some point they hit a wall, and they are not able to do the next difficult task.

And this has happened historically a lot. And there are lots of technical reasons why that has been the case. But the large language models have really achieved a kind of generalization that we have not seen from any other AI in the past. So it is absolutely the case that they are different in that way. They are more general, much more general than anything that we have seen in the past.

I think that right now they are good at complex tasks. They can do long-horizon tasks. They also tend to make things up at times, and they do that very confidently. So you're right that we cannot completely trust them. So we still need some human in the loop with sensitive tasks just to make sure that nothing is getting made up. For example, in our field, in science, they can still make up citations.

Right. That's a very common problem. In law, you have this problem. They will make up citations, cases which are fabricated. Sometimes the case is not fabricated, but the details are fabricated. There's some detail that has gone wrong somewhere. News articles that may not exist. And it's kind of surprising that it still makes things up when there's clear evidence that something exists or doesn't exist.

And my view is that it's probably because there are many sources on the web which are still not available to it. These companies are trying to absorb as much data as possible. Right. But there are still lots of, there are still Inaccessible parts of the web Yes behind paywalls or whatever else. Exactly. Right. Paywalls, CAPTCHAs, what have you. And so when it hits a snag like that, I think that is when it tends to fabricate.

Oh, that's interesting. Rather than simply say that I can't access it or I don't know. So it has trouble saying I don't know. And no less than OpenAI has actually studied this phenomenon, and they've come out with work showing that this is fundamental. The way that the language models have been trained they are disincentivized from saying I don't know.

Right. So the big companies recognized this, and this is not an impossible problem to solve. So I have to believe that now that they have recognized the problem, they are working on solving the problem. Whether it will completely get solved, I think there are very good reasons to believe that some degree of hallucination risk will always be there. I think, again, some of the pioneers of deep learning have made that point, and it's a little bit debatable, but I don't think anyone seriously thinks that they can be hallucination-free completely.

But with enough guardrails and things in place which people are working on, I think in practice they may ultimately become hallucination free. So I am bullish on them. Yeah. That's interesting you describe hallucination as being almost gaps in sources where, I don't know, there's a logical leap it needs to make to get from point B to point D, and if it can't find point C that's where the hallucination comes in Yes in some sense.

That's correct. Yeah. That's interesting. I've never heard it described like that. And yeah, I noticed this phenomenon too where AI is almost self-confirming or whatever. If you ask Yes AI you send a picture to it of my outfit. Does this outfit look good? It's almost very prone to saying, "Yes, it does." It has a hard time giving the tough love or tough feedback or something.

That's what I've found as well, and you have to prompt it explicitly sometimes to give you the tough love, and even then it doesn't give you the tough love. Right. Again, I think it's probably because of the commercial nature of these models that I'm sure the companies don't want to upset their users. Right. And so they may have trained it to. And we know, not that they may have.

We know that they train the models to make sure that users find like the output of the model, that the model's output are aligned with the preferences of the user. Right. Yeah. So the word is preference here, and they are trained to align with the user's preference. And explicitly so, right? Yeah. That's a little bit like the social media algorithms, where there's an addictive quality to them, because they're showing you what you want to see.

But if you think about it, these are quasi-technical or pseudo-technical issues. These are decisions that have been made by commercial entities. Because they Yeah needed the model to be out there and be used by as many people as possible. But the technology itself is More agnostic more agnostic. Yeah Those kinds of issues. Yeah. That's interesting. Yeah, because it's almost like the way news has become so segmented politically.

They know that the way to make money and keep you coming back to watching the news is to tell you the news that you want to hear. Yes. Read the storylines in a way that you want to hear it. So maybe some of that's going on with the LLMs Oh, yeah giving you the answers that you want to hear versus You should hear a more agnostic Yeah this outfit does make you look fatter.

It's not that flattering or whatever. Well, it's confirmation bias, right? Right. That's the issue of our times. Yeah. Confirmation bias sells, and we see that everywhere now. Yeah. So I'll just finish with this as far as this topic. With AI, you think AGI could be on the horizon. There still needs to be a human in the loop which people always contend with, like how autonomous agents are and whatnot.

Yeah. Do you think that we're going to have mass unemployment as a result of it? Where do you see labor fitting in? Do we have to have a new paradigm around what it means to have a gainfully employed population? Yes. If that makes sense. I know that's a broad question. It makes sense. No, it makes sense, and it's something I've been thinking about a lot, and I do think that there's going to be more unemployment coming despite what anyone says.

Anecdotally, we hear a lot of things which don't make their way into government data, right? Right. But we talk to a lot of people, and so many people are like, "Well I'm making do with one engineer, and we are able to do everything, and we don't need five engineers anymore." And there are people I've talked to who are like, "We used to have a team of 40 engineers or 20 engineers." They raised a Series A round, this was two or three years ago, and they had 40 or 50 engineers. And then they laid off most of them because they realized that actually having more engineers was causing more work than less work because there was a coordination problem that shows up.

And then sometimes, some people may not be that good. And then the other people are taking the slack for that, and then there are issues with just keeping up the morale. People fighting. Oh, okay. Yeah. Keeping up the morale and so on and so forth. And so people are making do with less, and that does worry me. I think on the one hand, it's inevitable.

On the other hand, it's certainly not good Yeah for us as a society. Yeah. And I don't know what the new paradigm would be. I know there are several proposals out there. The UBI, whatever. Right. Yeah. Or like what they do in Alaska with the oil. That's the other solution, which is that give everyone a dividend or Oh. the government takes stakes in AI companies, and then the proceeds of that gets distributed to the whole country.

And so I'm just saying that's another proposal. I'm not saying I'm in favor of any one or the other. But there are all these proposals out there. And then there's a camp which thinks that this technology is going to be so great that basically everyone will become unemployed at some point. Abundance. The abundance point. Yeah. The abundance camp. Yeah.

So there's that. And then there's kind of the dystopian camp which is like, this is just going to lead to massive inequality, and it's not going to lead to abundance except for a few people, and everyone else is going to be fighting for scraps, and that's kind of the more dystopian view. So we don't know what's going to happen. I don't think it will be extreme in any direction, and it will take a long time for society to accept a direction like that.

This has been noted before. I believe it was Karl Polanyi, he was a very famous geopolitical scientist. And he made this point, not in the context of AI, but I think more than 100 years ago, where he's like, inevitably, whenever there's a big push to have a massive social change, there is always a counter-reaction. And the counter-reaction leads to the change happening, but over a longer period of time.

So whether it's electricity or it might be a new currency or whatever the issue might be, but there is a social backlash against anything that is too radical, and then it doesn't fizzle out completely, but it just kind of changes direction a little bit, and it slows down a little bit, and you still get it. But instead of two years, it takes 10 years or 20 years or 50 years. And we can find a lot of examples of this over time. And the same thing would happen here, like with data centers, we are now starting to see some of that backlash.

So will it stop all the data centers? Probably not, but it might slow them down, and then that might slow the technology down a little bit, and then there might be some other policies. We're seeing the agents going rogue and hacking other websites and there may be enough awareness that then people are like, "Well, we have to be more careful." The costs are going up for all these models. Then people might start paring back.

So there are all kinds of counter forces that are there, and I think sometimes in the startup community and in the more technologically forward communities, we often don't realize. I think we have a stunted sense of social forces This is something that I've seen in tech circles, where they don't understand society. They understand technology, they understand the future, but they don't understand social forces. And history shows us that nothing too radical is going to happen in a short period of time.

So we have some time to figure this out, but we don't have a lot of time, and the question is, how much damage are we going to do to ourselves as a society before we figure it all out? Yeah. I think that is an interesting point because it's like, I took a Waymo today to go to a mechanic shop, pick up my car. It's like Waymo can do it, so now in a year, all the cars are going to be autonomous.

Well there's a lot of people doing driving jobs. Yes. There's a lot of taxi drivers, Uber drivers, all that stuff. It's not going to happen as quickly as maybe the most pro technology optimist Waymo Right supporter thinks it is. There's government involved, voters involved, political incentives and all that stuff. Well, and then remember, I think it was a year or two ago, where people were setting fire to Waymo cars.

Right. I think it was a year ago, right? Yeah. I don't remember when or how it was, but I do remember that. And then some people were tripping up the Waymos. I don't know how they did it. People were not happy with the technology Right and so we did see some damage. Now, I don't advocate for that, but that's an example of the countervailing forces. Yeah.

It is interesting because people always want to say the optimists are wrong. Well, look how few Waymo has been deployed. Right. And then the other people want to say "Oh, it's just going to take time," or whatever. There's both sides of the coin. Yes. But the reality is most of these things are kind of inevitable. They just take longer than most people expect them to.

That's right. Yeah. So we can go ahead and get into your company then, if you'd like. Yeah. Just a quick follow on down that previous question because you've been mentioning that maybe there's kind of inevitable that AIs will be more and more capable, and there'll be less need in employment. Yeah. What do you see, for example, students should be focusing on at this point, especially maybe when they're graduated, the AI's been even more powerful at the time.

And do you think they should be focusing on maybe doing their own startups or something like that? I think that's always a good idea. I mean, doing your own thing and being a problem solver. I think that on the other hand, there is a need for people with good social skills. Because that is something that the technology cannot do very well. We still need empathy, and people want to talk to a human in many cases.

And having the right social skills, having the right communication skills, which have less to do with the degree. I mean, the degree is kind of teaching you the mechanics of something. But there are all these things that the degree is not really teaching you. So I almost feel like people who are going to college they have to make the best use of all the opportunities that are available to them.

Like join the clubs and meet people and hang out and do all of that. So I think that's important. I think there are a lot of hard technical problems that the AI is making progress on, but we still need the brilliant engineers and scientists and doctors, and that's not getting automated completely anytime soon. We need good teachers because in the classroom we need professors and teachers who are empathetic, who care about teaching, who are good mentors.

Like all of that. So I think a lot of the professions that are more human-facing I think will be fine, but they will become more competitive, obviously. I know that a lot of young people now are going to trade school. There's a renaissance in trade school. I was just reading an article about that. Vocational schooling and trade school has really taken off.

It's going through a renaissance right now. I think that is great, actually. One of the places where AI is still not there yet, and it will take time for it to get there, is physical AI. We don't have great robotics, even today. People are working very hard on the problem, but it's going to take a while. They're too expensive. They do tend to break a lot.

They're not robust enough. There are fundamental issues with data collection. We need enough data and good models, and we need to control the lag. Because the robotic actuators are operating at a much lower latency than the ChatGPT, and you can't wait a few seconds before you send a signal Yeah to an actuator. It has to be real time. So we need the smaller models to become more powerful.

Maybe that's one way. Who knows? But there is a lot of money being put into physical AI and robotics right now, but we need to see that ChatGPT moment for robotics. We haven't seen it yet. So I think physical tasks and tasks with a heavily social component will be fine. But if you're good at something, I think that is kind of where the secret sauce lies.

You really have to study the thing that you're interested in and become very good at it, whether it's philosophy or engineering or the creative arts. And you have to be original and use the language models properly. You can't ignore the technology. But you have to become so good at what you're doing that you can work with the language model to create something very powerful.

Now, where the opportunities for doing that will come from is kind of up in the air because in corporate jobs there's not a lot of room for creativity, unfortunately. It tends to be more mechanistic, very siloed and then those jobs get automated. And then people like to say that, "Okay, you should have done this and that." But I really feel for the students and for the young people because I think for those of us who are older, we have the domain knowledge. So we are able to work with the model and Do things on our own.

But when you're young, you need to get those five to 10 years of experience before you become a domain expert. Yeah. So it leads to a catch-22 problem. That if the entry-level job gets automated, then how do we actually train the next generation of domain experts Yeah that requires five to 10 years of practice. Yeah, definitely. So we're figuring it out.

No one has a good solution. Yeah, I think there've been discussions saying that for entry level now for example, they're asking for more for the entry level people. Yes, it's not entry level anymore. Yeah. They expect you being relatively familiar with some of the work, and you can not only handle this entry level basic stuff Yeah tedious work, but already generating some value Yes at a certain sense.

Yeah. But I think there's also a counterargument saying that the people who get in the jobs who not be doing the boring works, but actually start from day one doing something that's really meaningful and Yes contributing to the value creation part. That's correct. Yeah. And we can move on to probably your startup part Yeah which we are also very interested in.

I think you do mention AI is helping with scientific discoveries right now. I think your startups kind of correlates to that field. Yes. Can you just share with us how does that idea come up? And I think from what the article that we read, I think it start as a LaTeX kind of Yes agent editor thing, and then I think it expand a little bit more to cover other aspects, other functions.

Can you just walk back through what's the origin story of that idea, yeah? Yeah. The origin story is me sitting in my office and feeling very frustrated. And that's the origin story for many founders. A lot of startups have come from the founder feeling that the problem is so acute, and will anyone solve this problem? And then you wait for someone to solve it, and they don't.

Yeah, nobody solve it. No one solves it, and then you talk to other people and find out they have the same problem. And so then you're like, "Okay, I have to solve it." It's waiting for me to solve. And then that's the origin story. Now, the problem that I was looking at was not the LaTeX editor and all of that. Those are the solutions. But the problem was, so I do research in the university.

We spoke about a research group. But it turns out that more than 44% of a researcher's time is not spent doing research. And I'm using that number very exactly, 44%. That's what a government report has shown. The number is actually probably higher, but we actually spend time on administration. So we are looking for grants. We have to find the right grant.

We have to write the right grant. We have to work with administrators. There's a lot of paperwork. A grant can take one to two months to put together. Many grants are 100, 200-page documents. The odds of winning a grant are probably 5 to 10% on the upside in many cases, and it can take about six months before you hear back on a grant. So those are abysmal numbers.

If you think about it, that's not good for science, and it's not good for research. And so a lot of universities, including USC, are now under fiscal pressure. They're having to do more with less because enrollment is down, and there's cuts to funding and research and so on. So there was even more pressure on us. Administrators were leaving, or they had to do much more.

People had to apply for more funding because now there's less money available. So all of this meant that the administrative work became even greater, more intense, and we had to do more with less. So that's where my frustration began. And I started the process. I've been here for 10 years, so I'm very familiar with it, but I looked at it systematically as a problem to be solved.

So that's where my perspective changed, where rather than look at it as, okay, this is the way things are, and this is how you have to do it, and you have to submit a form here, and then you need to work with someone here, and this is just the way it is, you need to take it lying down. I was like, "Okay, let me look at this as a problem. Why is it this way? Can we change something about it?

Can we make this better?" So that's where the journey began. And it was kind of mid 2025 when we first started thinking about it this way. And so me and my co-founder, we had a lot of discussions. We built a pilot. We started exploring. We started talking to customers. The company, like all companies, has gone through a number of mini pivots, and it's a journey.

But the one thing that we have stayed true to is that we want to amplify universities and university research significantly with AI agents. So that's our mission. How we do it and what specific products we build depends on the market and who we're talking to and what kind of funding is available. And all of that is part and parcel of running a business.

You have to listen to the market. But our mission is we want to amplify university research with AI agents. So that is our mission. And we built products that allow scientists to find the grant, to write the grant, and to manage the grant much faster. Our goal is to make it 10 times faster. So you go from 44% to 4%. And that frees up money and frees up time and frees up energy to do research and solve real problems and not be mired in paperwork and bureaucracy and compliance and all of that.

So that's what we're trying to solve. And we've come quite a distance since that first realization and first pilot in mid '25. But yeah, that's where we are. And you talked about the agentic LaTeX editor that is meant for writing the grants, but people are also using it for writing papers. We have since expanded that into a full research operating system.

So we now have a research operating system that is being used by researchers in more than 80 institutions, including USC. It includes infrastructure for running AI scientists So these are AIs that can actually do science. They can run experiments, they can look for literature, they can help you write the paper, they can help you write the grant. So we have a whole ecosystem for doing that so that you can do more science better science with the resources that you have.

So that's our research operating system. But our flagship product is called the Funding Navigator, and it uses AI agents to match professors to the right grant opportunities and then to help them come up with a winning strategy for the grant. Like how do you increase your odds from 5% to 10% to, let's say 20% and how do you work with the right people to put together the best grant that you could for that opportunity. So that's our flagship product.

We call it the Funding Navigator. Yeah. So was it a different set of AI models that help you find funding versus help you do research? How do you apply the right AI model to the right situation and design those solutions? Yeah. They are different AI platforms. I would not say that they are different AI models because at the core, all these products are agentic, so they use the top-shelf LLMs, whether it's GPT 5.6 or the latest Claude model.

So we kind of pick the models depending on both cost as well as what they are best for. So to give an example, we found that Claude Sonnet, for example, was very good at writing better than some of the more recent Claude models, which are very good at code generation. But we found that when it comes to writing, Claude Sonnet actually offered the best trade-off to us between the cost and the writing quality for the academic writing quality. Similarly, Google's Nano Banana model was very good for generating figures that could be used in grants and papers compared to some of the other models out there.

So, based on our own research and what we know about these models, we like to pick and choose the models that are best for the task. But the models are, I would say, 10% of the story. The rest of the story is building to the workflow. So to build a very good funding navigator we really had to understand what does the professor want from the product, what would make for a good product.

Because otherwise, they can just go and ask ChatGPT. Right. So where is our moat? And so our moat is that we build to the workflow. And we also have a lot of proprietary data. So we have a data moat, but we also have a workflow and process moat because we understand how they actually work. How the professors, the users of our platform, other researchers, administrators the tent expands with each product.

But we understand their workflows, and so we're able to solve their pain points by building around the workflows so that their behavior does not really need much change. They don't have to relearn something from scratch, but they are able to do what they were already doing in much less time and much less drudgery Yeah in a way. Yeah. And you're able to instead of them just prompting to find the funding, they're able to have a more curated experience from the start to the finish of finding funding.

Yes. So the whole kind of workflow is worked out. Correct. Is that correct? For example, we have an agent that goes to their homepage and extracts all the research they're doing and then tries to match that to the funding opportunities. And then when you click on a funding opportunity we also show you who's won the funding before and how much money did they get and which of your ideas are a good fit for the funding and who you should contact, maybe a program manager.

You should contact them, tell them about your ideas, get on their radar. You can set up notifications on the platform. You can do all kinds of things. And so, all of these real problems can be done on that platform. Yeah. So this is one of them, but we've also built other platforms. And we have a platform right now where the agent connects to Slack and Dropbox and email and all of these different platforms.

And so you can actually talk to our agent on Slack, and it will compose emails for you and put it in your draft folder. It will summarize your Slack messages. It will create documents and put them on Slack. Let's say you want to create an IRB, which many researchers have to do for biomedical and human-facing research, human subjects research. So imagine if you just give your grant proposal or your paper to the AI, and then it creates the whole IRB for you.

Normally, that would take days, if not weeks, to put it together. Now we can do it in minutes. And the most important thing about our AIs is that the more you use them, the more they learn. So they become more powerful with usage. They really learn to understand you as time goes on because we are noting all of the context. We don't throw away any of the context, so all the context is there, and everything is very secured.

So we run on Microsoft Azure Enterprise, and the data is never used for training the language model or anything like that. So the data all belongs to the user, and it's always on very secure premises. So I think for all of these reasons, that we are really building the company for our customers. That's very important to us that we are taking care of all of these needs and requirements that come up.

Yeah. Yeah. Yeah, I think that is an important point too, because I know that people that are using LLMs for their companies and enterprise, but even as in research like you're saying, they're worried about some of their dialogue, their chats with AI being used to train the future models. Yes. And then other scientists being able to almost co-opt their work in getting answers from AI.

So I think that point about having the data private is important. Was that part of your decision Yes, it is concern over that? Yeah, and Microsoft has been very instrumental in that. I think that the same Foundry models, on which ChatGPT Edu runs. We use that same Foundry layer and so the data is contractually protected. It cannot be used for training the OpenAI and all those models.

Yeah Okay. I did want to ask about commercialization. Obviously, it's like you might have a breakthrough product, but then a company usually has to generate some sort of financial return. Yeah. What has been the kind of evolution of that process? Where do you see things going in terms of commercializing it and all that? Yeah, so our products are commercialized at this point.

We work with universities, and we sell licenses to them. So, we are already working with and running pilots at multiple universities including USC, RPI, UT Dallas. Some of these are pilots that we hope will convert into licenses. Some are licenses. But that's kind of our model. We have the products, and then we work with the universities. We set up pilots.

Usually, we get the pilots up and running in less than 30 days. We don't usually charge for that, where there's a good faith decision that, yes, this is a problem we're very interested in. We want to see a solution. We build it, we pilot it, and then we kind of convert to a license. So that's our model. We're a B2B business. So the sales cycles are long.

Higher ed, in particular, has very long sales cycles Of course and we are facing fiscal crunch right now. So the sales cycle is even longer. But that's the challenge for us, and it's a challenge that we are willing to go after because that's why we began. It all began with that, with the frustrations of spending too much time on administration. So we do believe that there's a huge opportunity because universities are still using platforms and legacy products from the '90s and 2000s in many cases.

So there's a real need to build for academia, and a lot of people are not building for academia. They're building for others and then bringing them to academia. But we think of academia and universities as first-class citizens. We're very focused on that. Yeah. Yeah. Yeah. So what was that decision, though? Why necessarily selling to the university kind of as a whole and having some sort of group license or organization license versus individually selling to researchers?

Are you open to being more of a business-to-consumer business later on, or are you thinking of just staying kind of university to university? Because obviously, as you mentioned, there's a lot of bureaucracy and boards and versus selling to just somebody saying, "Hey, pay $20 a month," or whatever it is. We thought about it. Actually, it's interesting.

We initially started that way, and this is where the customer discovery was very instrumental. What we found out was that professors don't really want to pay out of pocket. They expect the university to pay, and maybe that will change someday, but right now they are not willing to pay for all these products. I think there's a psychological effect here.

It's not just about the $20 but there's a psychological effect. And also from a business perspective, it doesn't make sense for us to sell to individual professors because many of our products work for entire departments. Some of our products are meant for entire departments. Either the whole department uses it or no one uses it. It's a little bit like that.

So we started almost as B2C with sort of the LaTeX editor and the research operating system, but we've come to realize that the enterprise model is the right model for us, so we're focused there. But that being said, we love working with faculty, and we often give them free access to some of the work that we're doing. So for example, the LaTeX editor, and we have an AI reviewer for reviewing your grants and papers, the AI scientist. We give limited access to faculty.

We give them very generous credit, so they're able to use it for quite a period of time, and they don't charge anything. And then after that, if they want more, they can actually pay into the research operating system, and we do have an individual way to pay for that. But we don't really see that as the main business of the company. Yeah. You can have them pay in as an individual to get more credits Yeah but you'd still probably have some sort of contract with UT Austin USC.

We would aim for that, yeah. Aim for that, yeah. At some point, if enough faculty are using it, our model is go to the chair or the dean and tell them, "Hey, all these faculty are using. Some of them are paying for it. Why not sign a deal?" And in many cases, they are willing to. Yeah. So just in terms of your central motivation for it, and I think it is so powerful is that so many of the scientific and technological discoveries have come out of the universities.

There is kind of this rise of industry labs and all that, obviously. Yeah. But the cure to cancer might come out of some grad school's research group. Yes. And so you're trying to free up the administrative burden, all this other stuff that surrounds doing research, and get them to spend as close to 100% of their time focusing on the cure to cancer, the That's correct solution to nuclear fusion.

That's kind of your overall Mission overall mission. That's our North Star. Yeah. It's like how much time can we give to research? Can we maximize it by freeing up the administration? So that's a very important part of what we're doing. And then the more future-looking part is that then using AI agents through our research operating system, which is why that's a much more future-looking product.

And ironically, that's why we give free access to that product because we really want people to use it and start getting used to it. But the research operating system is a way for human scientists to use AI agents to boost their research even more significantly. So I am trying to now see that can I set up an autonomous lab where you have AI scientists that are actually doing work all by themselves and improving all by themselves, and it won't lead to papers because we still have to write the papers as humans, otherwise it won't get accepted. There are ethical issues over there.

But at this point, all I care about is can we actually turbocharge the science in whatever way possible? Yeah, not just in the specific avenue, but biological research anthropology research, doesn't matter. Just help to push the frontier of knowledge, regardless of what type of research it is. Correct. Do you see that sort of research operating system being more not as mature as some of the help with the administrative tasks?

How do you sort of view the research operating system components of Grail in relation to the other parts? It was our first product, right? So in some ways, it is actually the most mature. Wow. I think, yeah. I think the enterprise products tend to be more customized. I think they are mature, but we are still collecting a lot of data on them from our existing pilots and so on.

So they are still production-ready and they are ready for deployment, and they have been deployed in a few places, right? So they're all mature. I think all of the platforms are mature. They are certainly not lab products anymore. We work very hard on that. But I do think that the research operating system we still need better models. I think that the AI models are not quite there yet.

It's going to take one or two more generations of models, which could occur in the next six months, or it might be another year or two. I don't think it will be longer than that. But we need a couple more iterations on the models from the hyperscalers before I think the AI scientists will really be there, and I think cost is still a major issue over there. So they're coming, but the auto scientists, the AI scientists, are still not completely there yet.

But we are doing a lot of research on that. Actually, that is an area where the research intersects with the company. So a lot of our research has to do with the operating system and the AI scientists. Yeah. And we're not too obsessed right now with how do we turn it into a product. It is being used by a lot of people, so it's kind of out there, but the auto scientist itself is actually not available.

We have not made it available. Got you. It's only internal to the group. But the tools that we use to build the auto scientists are available to everyone. That makes sense. Yeah. Yeah. I will ask quickly about just, because that is a topic that a lot of people that are skeptical of AI and AI products, they think that the costs are just going to continue to scale, and then regardless of what value AI agentic products provide, the costs are going to get to such where it's not commercially viable.

How do you view the sort of cost structure of it? A lot of people think it's just going to continue to skyrocket. A lot of people think it's going to trend to zero at some point, almost like a commodity or something. Yeah. How do you view that? The data science, data centers in space and all this stuff. Yeah. Without getting too wild, I am leaning towards the latter, not the former.

So I do believe that the cost will come down, not up, and there are a couple of reasons why. One reason is that just historically, we look at every single technology out there, and technology is always deflationary. Everything else seems to go up, like food and oil and all of that goes up. But technology actually goes down. Technology is deflationary.

Some people think, well, is that really the case? My phone goes up each year. But that's because you're getting a new phone. It's new technology. Right. If you buy the old phone, then it's actually much cheaper, right? If you buy a pre-owned car or pre-owned phone or whatever it might be, it actually gets cheaper year by year. There are some exceptions, like there was a shortage of memory chips, and so anything with memory went up significantly.

But now that's gone away, I think. So there will be spurts when there's a supply shortage. But let's leave that aside, assuming that's not the case. Generally speaking, technology is deflationary. The price will deflate, right? TVs. Yes. And the GPT-4o, for example, 4o mini 4o, they were much more expensive when they first came out. Now it's much cheaper.

The small models keep getting better. So yes, the very best models their price might keep going up and it's possible that the very best model that you may need to use for an auto scientist or some very complicated problem you have to pay big bucks for it, and there may be some proprietary technology behind that so that right now it's seeming like this arms race between all the hyperscalers. But who says that one of them does not leap ahead technologically? It could happen.

They may discover something new or they just have a new way of doing things that just puts them well ahead, and you may have to pay a lot for that, or you may not. It really depends. But the small models are getting better, and that I think is the biggest reason why I think that the cost will go down. Yeah. And then, like you're kind of saying, you found for Sonnet, it might not be the most advanced AI model that they have, but for writing It's good enough it's good enough for writing the research grant proposals.

Right. Some of that scaling of the cost, even for the absolute best model or whatever, it might not matter Is not necessary. Yeah it might not be necessary for whatever product you might need. Exactly. Okay. Awesome. Yeah. Yeah, and I think going back a little bit, you mentioned about how this AI is being used in doing scientific discoveries writing grants, applications, as well as papers.

I think I read one of your Substack articles saying that there should be a new way of evaluating the grants application. Yes. It's not about how beautiful you write your proposal. It's probably more about the track record of the person or something like that. Is that also happening in, for example, paper reviewing? Are they looking for something else when there's AI able to help?

Unfortunately not. Oh. But I'm glad you read the Substack. That is the point I was making at least for grants, which is that everyone is just using AI to write the grants. And so I feel like we should get rid of the system where we have to write all these long grants And some agencies are adopting a different system. For example, in DARPA, the pitch format has become much more common, where you don't write a 15 or 20-page technical volume.

That is just a technical portion of the proposal usually. But now what they do is they put out the program and they say that, "Submit an abstract. It's four or five pages, and then if we like your abstract, we will invite you to pitch." And then the pitch is much more efficient. You go and you have a five, 10-slide deck. You get 10 minutes, 20 minutes, whatever the time is you pitch, and then in a very short amount of time, they will let you know whether you're selected or not.

So, I think we need systems more like that, where when you are pitching, the truth is that you have very little time, so it's going to depend on how creative your idea is, what your track record is, and yes, there is a personal element. Are you able to connect with the program manager? Are you Charisma Yes. Those things do matter, and unfortunately or fortunately, whatever the consensus might be, but it does depend on the person as well, and on their rapport with the program manager.

And I think that system is a lot better than writing a 15 to 20-page dense technical volume Right that no one will understand, and that you won't end up doing anyways. By the time you get funded, the technology will have moved along. Right. And we don't actually end up doing what is in the proposal. This is more common than uncommon. Now, speaking of papers, we're not seeing that.

I think the written word is still very important for papers because it's original research. But I am trying to push for a movement, especially with AI scientists, where I'm saying that we should forget about the paper. We should only write papers on some topics that we want humans to read about. So we have to, instead of publishing more and more papers let's publish fewer papers, but let's have a different way of expressing our results.

So you can still express your results without writing a full paper on it. Anthropic, for example, is doing this. Also OpenAI. Anthropic puts out a blog post sometimes where it's conveying some important research. Now, sometimes that will also be accompanied by a pre-print. But I've seen a lot of research from Anthropic where it's just a nicely written blog post, and it's not following the exact structure of a paper.

It's more conversational, it's more understandable, but the work is still very rigorous. Yeah. I think I read a paper of theirs evaluating people that used AI to complete a task versus didn't, and evaluating their understanding, and they communicated it very clearly. Yes. And it propagated and it really reached through. It was on LinkedIn, social media, all across the way.

Yeah. Whereas if it's too dense Write a dense paper if it's too dense, it's not going to have the same reach. Correct. Yeah. And I think we are also curious about you being a professor, researcher, at the same time, a founder of startups. Does that two part similar or different or do they help each other, or do they not help each other? They're very similar in some ways and very different in other ways.

So the culture is very different in both fields. In research it is much more like I would say you need to be more humble and skeptical and people don't like a lot of bravado and Overconfidence Yes. Even if your result is good, it's like there's a certain way in which you're expected to present it and talk about it, and there's a certain humility. So I think the communication culture is different, and there are some good things about it and some not so good things about it.

The not so good thing is that we can be very overly pedantic. Especially in some fields, people can be very pedantic. They can really dig their heels in and you kind of step back and you're like, "Wait what are we really arguing about? I'm not seeing any impact over here." But for whatever reason, they are dug into this argument. So that can happen.

I think that what is common between them is that they are both very innovative. So when you run a research lab, you have to be very innovative. You have to fundraise. We've talked about that. You need to, and the odds are not great. It's five, 10, 15% odds. When you fundraise as a startup, the odds are lower than that, but not too much lower. It's like one, two, 3%.

Depends again on who you are. Are you an ex-founder? Are you in the Bay Area? Do you have support from Y Combinator, et cetera? Depending on who you are and what kind of company you have, the odds might be higher or lower. But that in itself is a market signal. So you have to follow the market much more. You have to really listen to the zeitgeist, and you have to be willing to pivot much faster.

And it's less about formal outputs and just more about surviving. I think when you're doing a startup, your number one goal is you have to stay alive long enough. That's what Jensen Huang said, I think in a recent NVIDIA talk where his main point was that, "Listen, I'm here today because I managed to stay alive." For the longest time, NVIDIA was just making GPUs for gamers. Who could have imagined?

But I stayed long enough, live enough. I lived long enough, and I stayed alive long enough that I was able to see this era come to fruition. Right now, we are one of the most valuable companies in the world. So the lesson over there is you need to stay alive long enough. That's a very different way of thinking. And when you're a researcher, you're also thinking about longevity, but it's not so hyper-competitive, and you can think about the long term, and you can have off periods, and you can slow down, and you have more of that.

It's still very competitive. You still have to look for funding. You still have to support your students. But it's more relaxed and long-term, and you're not as intent on how do I pay my bills for the next six months? Right. And where will our next customers come from, and are we reaching out to 100 customers each week? Are we keeping our customers happy?

Are we keeping our investors happy? Are we on top of everything? So it's much more frenetic, much more fast-paced. Much more of a winning culture where you really have to play out your wins. And you have to be out there, you have to talk about it. You cannot be humble. Right. You have to talk about your wins and your successes, and you want to change the world, and you have to be very vocal about that.

And I think the way I'm able to bring it together is that I'm able to almost take a metaphysical view of it, where I'm like both of these are cultures. I study complex systems. So I'm like, both of these are complex systems. One is not better than the other. They're both complex in their own ways. They both have their own stakeholders, their way of communicating, what they're trying to accomplish, their metrics.

They have their own cultures, in other words. And if you want to be successful in either one of them, you have to learn the culture. The culture will not learn you it will not accommodate you. You can't say, "I'm an academic, so treat me differently." You have to learn the culture and get with the culture and succeed in the culture. And you can do both.

It is possible to do both. Yes. Did you keep two sets of personas? I do. I have two sets of personas. Believe it or not, I really do have two sets of personas. I have the one with my students and my colleagues, and I have one with my investors and my customers, and they're both me. It's not fake. They're both me, but it's like Jekyll and Hyde. Yes. Yeah.

So when you're talking about your research and your research group, you're more humble and like, "This is what we've done." Yeah. But then when you talk about Grail, you've got to be willing to drive it home right? Yes. You've got to drive it home. Yeah. Yeah. I'd imagine that is difficult, though, because you've been a researcher for so long, and predating any of this founding of a company.

Yes. And then you kind of have to switch modes and have bravado and Sometimes within hours. Yeah. One hour I'm doing research, and then the next hour Although it's become a lot more about the company now. I've basically gone into the company full time. Since we got our first round of investment and so on, it's like I pretty much work on the company now full time.

The research group runs itself in a way. I trust my students to do what they're doing. There are no weekends, there's no evenings, there's none of that. You just have to keep going. And you have to love it. If you don't love what you're doing, and if you're very stressed out, then it's not fun. It's very total. The whole experience is total. It's more than fun, it's total, but you have to have an intense degree of fun.

You have to be comfortable with the pace of it, and with making mistakes and all of those things go with the turf, but it's day by day. I feel like every single day is uncomfortable. Yeah. That's what has changed. Yeah. I think it was Elon Musk that said, maybe he was quoting somebody else, he said being an entrepreneur is like chewing glass and staring into the abyss.

Yes. As if one or the other wasn't enough, you're kind of doing both at the same time and yeah. You're very uncomfortable and you're very uncertain. Yeah. And I think that's what he meant. When you're chewing glass, then it's very uncomfortable. So that's an extreme way of saying it's very uncomfortable. When I'm talking to a customer, it can feel a little bit icky because that's not me.

I'm not a salesperson by nature. Pushy and feel a little bit like you're pushing something on somebody. Yes. Yeah. Yeah, and you have to do it in a nice way, and you have to find the right way to do it. But there's no question in your mind when you go into the meeting that, okay, my whole goal in this meeting is to cultivate a sales contact. I am trying to make a sale.

It might take me three meetings to do that, so I don't want to push it and ruin the meeting because then it won't proceed any further. But you're not going in to learn about research or anything like that. You're going in with the goal of ultimately making a sale. And so that's what you have to do. Whereas if you're talking to someone about their research or their papers, et cetera there's no agenda really over there.

You're just having an intellectual conversation. But you can't afford to do that when you're running a business. You have to use your time very carefully, and every call should have a purpose. If a call does not have a purpose, you're wasting time. Yeah. So you have to have those metrics. If I'm taking this call, if I'm writing this email, if I'm reaching out to someone, what's the goal over there?

Maybe it's not that I want to make a sale now, but why am I doing that? What do I hope it will lead to? And so that can be uncomfortable, and you have to learn how to deal with it. And then the other side is it's uncertain. You just don't know. Will you actually survive six months from now, 12 months from now? Yeah. No one knows. There's only a certain period of time to which you have runway and then every startup is always thinking about whether they will survive two years from now, one year from now, three years from now.

We just don't know what the long-term future is. We can tell ourselves what we want, but we know what the failure rates are and all of that. So it's just at the back of your mind you know that what you're doing is very high risk, high reward. You want it to grow very quickly and grow very big, but that by its nature, it is very risky. So you are doing something very risky, and you can have a lot of confidence in it but it is risky and it is dishonest to assume that it is not risky.

But you have to have that degree of craziness and confidence to leap over the abyss. To do your work while you're still staring into the abyss and really believe that you can do it, you can cross the abyss. Yeah. But there's no proof. You have to run on blind confidence and really believe in what you're doing. Yeah. And especially doing something that you're doing because your sort of tech startup is not just iterating on something that somebody's done before.

It's kind of just navigating your own space that you're sort of creating here. Yeah. So there's no test case to say It's a new market it's a new market. Yeah. Yes. Yeah. So how do you manage your time? How do you keep sort of emotional stability with all this existential risk hanging over your shoulder? It's like the company grim reaper is always over your shoulder.

I've never heard it put that way, but I like it. Yeah. Okay, how do I manage it? I don't know. I think I deal with stress well and I think it does take a certain type of person to deal with something like this. Anyone who comes to me for advice and asks me "Should I do this or should I not do this?" That's my question to them. Do you feel like you can actually deal with 10 things at the same time?

And not get too upset and not take things too personally, and are you able to brush things off very quickly and just move on? And are you comfortable cold calling and just doing what you need to do to fulfill the mission? And if you're not able to deal with that, if you feel too uncomfortable, if you're overthinking too much, if you take things too personally then it's probably the wrong field. Or you have to grow out of it, and that's hard to do because it might be a personality trait.

I think in my case, what also benefited me was that I come from an entrepreneurial family, and so I grew up around small business owners. And I saw them take risks, and I saw the work that they put in and how they talk to customers, and the uncertainty because one year a lot of money is coming in, the next year not quite a lot is coming in, and so you have Yeah to have reserves and all of that. And so I just kind of feel like all of that was internalized in me. I never thought about doing that until much later on.

Obviously, I spent 10 years here and then I decided to do that. But I feel like it was always there in my bones because I kind of grew up around it. So that helped. And then I think the personality type helps, which is that I'm just generally energetic and optimistic. You have to be very optimistic Right to be in this field. You cannot be pessimistic.

You cannot be cynical. You have to love the future, see what's coming, believe in the product, believe in the world. You really need a high degree of optimism. And you need to be able to brush things off quickly and have confidence and have self-awareness. I think that helps. Some might call that an optional trait. There are startup founders who are not self-aware, unfortunately.

There are researchers who are not self-aware. I think it helps to be self-aware because in the end, you can step back and say, "Okay, I went through the day. What worked, what didn't work? Where could I do better?" And "Where was I trying to make a sale?" Or, "Where was I trying to do research?" Or, "What was I trying to accomplish?" And I think it's very important that you don't BS yourself over that.

And it helps you have a better relationship with your clients because you shouldn't lie to them. You should be authentic. And it helps to know what you're trying to accomplish when you're trying to do that. And you have to be a good communicator and hard worker, and all of those things goes without saying. I think those things can be taught to some extent but emotionally, you have to be able to deal with stress in this job.

And if your blood pressure goes up too frequently, it's not for you. Your heart will retire you from the startup at some point. Yes, it will. If you go too up and get too crazy about it. Yeah. Yeah. Go ahead, Philip. I think we're kind of coming to wrap up this episode. Just really appreciate your time. Know you're pretty busy running both of your companies and the professor's duties at the same time.

Yeah, really appreciate you being here and we are looking forward to hear more about your progress on your startups in the future. All right. Thanks for having me. Yeah, we look forward to Grail being at however many universities there are in the US, but it's cool. From your lips to God's ears. Thank you, Professor. I appreciate it. All right. Thank you.

Yeah. Thank you so much.