从制药到AGI热潮，再到金融AI开发：Martin Shkreli的历程

You're listening to Gradient Descent, a show about making machine learning work in the real world. And I'm your host, Lukas Biewald. Today I'm talking with a super unexpected guest. Some of my colleagues thought this was an April Fool's joke that I was even setting up this interview. It's with Martin Shkreli, who I originally became aware of probably like a lot of you as the Pharma Bro, a young pharmaceutical CEO who was raising prices on consumers and quickly became one of America's most hated people. I didn't really follow that for years, but then he resurfaced on Twitter and was actually talking about Weights & Biases and had really thoughtful feedback on our product. He had gone incredibly deep into PyTorch and neural nets, and I kind of couldn't believe how technically competent he was. Then I looked at the product he was building with Weights & Biases, and we had a nice connection as fellow entrepreneurs. So I really wanted to have him on this podcast.

1. Hot take: AI in drug discovery is mostly hype

Of course, we talk about what his experience was like getting canceled and going to jail, and I give him an opportunity to talk about his version of the events. But we also go super deep on AI and drug discovery, something a lot of our guests have talked about, but he has a really interesting, different take on that. We also go deep into his trading product, how he built it, and his entrepreneurial strategies. We talk about what it's like to go to jail, what his experience of jail was like, and some advice that he gave Sam Bankman-Fried recently on going to jail. So this is a really wide-ranging interview that I found fascinating, and I hope you enjoy it.

Martin, thanks for being on the show. You're probably the most professional livestreamer we've had on here.

Thanks so much for having me.

2. Shkreli’s pharma background and fragmented drug economics

Yeah, absolutely. Before you started talking about Weights & Biases on Twitter, I really knew you only vaguely as someone who was buying pharma companies and then raising prices. You're known as the “Pharma Bro.” I've seen you online correcting people and telling your side of the story, so I thought I'd offer you the chance to explain it. I'm honestly curious to hear your take on those events before we get into the AI stuff.

Obviously, if you do something really noteworthy that sticks out, it sort of takes over your life and affects how people look at you and what you do. I think that probably happened to the Hawk Tuah girl or any of these other people. It's like, “Hey, you're on the front page of the news. This is what people remember you for forever.”

Before that happened, people remembered me as an investor or trader, and also as a biopharma guy. I'd basically done everything you could do in biopharma, whether it's discovering new drugs, developing drugs, licensing drugs, or acquiring drug companies. I went public at 28—not that that part's particularly relevant—but I had some very successful drug companies.

I did a lot of the things that Vivek Ramaswamy did. In fact, he was my biggest investor. We saw the drug industry as a very fragmented, modular industry where each drug is similar to the others. In the software industry, your company may have nothing to do with, say, Arrival—not Arrival, but ServiceNow or maybe a company like CrowdStrike. They're pretty much apples and oranges.

Whereas in pharma, the drugs have very similar, overlapping properties. They all work a certain way; they have certain toxicology and certain pharmacokinetics. They're very much members of a class, so to speak. We saw the industry as this massive collection of class members that was generally not well arbitraged.

The individual companies would trade on the stock market, but within those individual companies—say, Pfizer or Merck—they'd have 1,000 different assets. The company's management team would have to decide, “Well, we like this one, or we don't like this one.” They'd make smart bets sometimes and not-so-smart bets other times. Sometimes they'd make bets that traded off a certain utility for them—something they were willing to forego for us.

A lot of people have since gotten into that business. It's gotten a little trickier, but it's still a really fragmented business. That was my main focus for a long time, so I know the drug industry and the finance industry really well. For me, moving from the hedge fund world into more of a private-equity-style approach to pharma was really helpful.

A lot of traders come to Wall Street because they want to make money. But what I think you realize over time is that trading is tedious and difficult—it's really going to tear your hair out. You learn all these fascinating stories, and you meet fascinating people like yourself. You start to realize that you love the story, the industry, the people, the products—all the lore is really fun.

Private equity is one of these things. Whether you're a VC, which is sort of a form of private equity if you ask me, or whether you're starting a company, buying a company, or doing a little bit of all of the above, you're getting your hands dirty. I realized that I was 100 million times better at that than I was at investing or trading. That's how I got into pharma.

I've certainly always looked at software as well, but we can talk about that as you wish.

But wait, you originally came from more of a trading perspective rather than from a deep pharma background. You don't have a biology background, right?

No, but when there's money on the line, you force yourself to learn as fast as you can. I was fortunate that, working for Jim Cramer at the start and then working for a so-called Tiger Cub, I was able to talk to management teams, analysts, and others on a one-to-one basis.

Sometimes I feel like I learned a lot of this through hard work and my own perseverance, but I also got to see how the drug industry worked up close.

And were you buying existing drugs rather than doing the science to discover new drugs? Is that right?

No, I did a lot of science. I had a lot of publications in the Journal of Medicinal Chemistry. I did a lot of crystallography and preclinical work, so I've done it all—taking a drug from an idea in my head.

I remember sitting down at McDonald's with one of my business partners and sketching out molecules for pantothenate derivatives.

You really like sketching out a molecule on a napkin?

Yeah. I grew up as a nerd—a self-taught nerd, but a very deep nerd.

Pantothenate is vitamin B5, and in a certain disease, children can't make or process vitamin B5 into its active form because they're missing an enzyme called pantothenate kinase. We said, “Well, if you're missing this enzyme that converts pantothenate to phosphopantothenate, we could create phosphopantothenate.” We were able to get the basic idea down.

Of course, I worked with a lot of brilliant chemists. One of my chemists had worked for a famous Nobel Prize winner, among others. But in management, I think you may have a lot to say about this, too: You want to know how the whole company works.

I was never the best chemist at any of the companies I started, and I certainly wasn't the best biologist. But I knew enough to ask the right questions. For example, “Why won't this work? Why is this drug—this drug looks like it's going to have a very low log P—how is it going to be useful?”

You don't have to be the perfect person at every single job, but I think you should know a little bit about each area. When a lot of the other pharma companies criticized me, I would often point to their CEOs and say, “Do you even know much about medicinal chemistry or some of these other areas that are really knee-deep in the pharma industry?”

Whether it's software, pharma, or any other industry, I really do want to know how the sausage is made and get my hands a little bit dirty. It took 15 years of pretty hard work before some of these things started to dawn on me.

Again, I'm not the perfect person for any of these exact disciplines. I don't push production code at our company, for example.

3. The drug price hike backlash and his thoughts

Well, look, I totally respect that. But I think the criticism was right, wasn't it? Wasn't it that you were buying an existing drug and then raising the price? Is that accurate?

Yeah, that was the criticism. I think it's one of those situations where you should never let the facts get in the way of a good story.

The good story was: “Here's this guy. He's a really unabashed champion for capitalism, and he's doing all these things and saying all these things, so you should not like him.”

What’s a wonderful story for a film or something is often very different from reality. Our drug had many other drugs that were replacement goods. You could have simply taken another medicine for very little, but that would have messed up the story.

There were other things, too, like the fact that I didn’t take a salary at the company or anything like that. There was no need to do that; our company wasn’t profitable. It was a very small drug. It was still relatively inexpensive because, when you do the multiplication on the price, it seems like an awful lot, but you really have to look at the absolute price as well as the relative price.

There were drugs that cost millions of dollars. This drug was a fraction of that—a very small fraction—and today we still have drugs that cost $1 million per year per person. There’s very limited understanding of the orphan-drug industry, which is one of the best things. One of my deep regrets is that I think I obscured some of the great work that pharmaceutical companies do in orphan drugs, which is my specialty.

The PKAN medicine I mentioned earlier is for a disease where maybe 500 people have it. It’s pediatric, it’s fatal, and the biology is fairly clear and simple to get over the hump and make a potential cure for the disease. To me, when you have 500 people in your addressable market, you’re going to have to charge $1 million per patient. There’s no other way.

I think a lot of people, when you put that to them, realize, “Would I rather have the patient, who’s a child, die, or have a massive, rich insurance company with no problem paying $1 million—because it’s only 500 people—pay for it?” The answer is totally obvious. I’m still in touch with some of the families who have children with this disease.

I feel like the price increase, in and of itself—we had done this six or seven times before, and we never had a problem with patient access. In fact, patient access tends to go up because we can now afford to provide for these patients. There was just a lot of stuff that fit into this generic theme of whether you’re going to use your rational mind or your emotional mind.

For so many people, they want to use their emotional mind. I’m really lucky that I started podcasting, webcasting, and livestreaming at a time when nobody did it. It was 2015, and I basically said, “I can’t win. I do these interviews, and they still try to make me look bad. I’m just going to show you who I am.” I would turn my camera on and go at it, and a lot of people would watch.

In fact, most of our seed or pre-seed investors—you can see those on our website—were people in tech that we all know and love. They watched my videos and said, “Well, hold on. Everyone’s got this guy all wrong. He’s one of us. He sits and programs. He sits and hacks away at a spreadsheet. He’s showing us how to be entrepreneurs.”

Some of these guys ended up becoming some of the biggest entrepreneurs. For example, Prasanna Sankar, who co-founded Rippling; Amjad Masad, who co-founded Replit; and Pieter Levels. These guys did really well, and many of them credit my videos with teaching them some fundamental analysis, financial statements, and things like that, which you do have to learn when you become a CEO.

The people who did that had a strong conviction that I wasn’t a bad guy, whereas the millions of people who watched the traditional media clips had very little conviction that I was a bad guy. The people who had a huge conviction that I was a good guy mattered more to me than trying to make every single person love you, because that’s sort of hard to do.

The people who love you, or who do think you’re a good person, will vouch for you, and that will tend to spread through network effects. I’m really lucky that people like the folks I mentioned, and so many more, have vouched for me and, through network effects, have said, “No, no, this guy isn’t as bad as they say he is. In fact, he’s really great.”

I’m super lucky that I’ve been able to do that, because oftentimes, when someone’s radioactive, they tend to get cast aside. They’re a pariah, and there’s no way to redeem them. I feel like we also happen to live in an environment—there’s a bit of luck here—where people are, in general, taking second looks at others and saying, “I don’t know if I should believe that.”

That has good effects and bad effects, because there are things in general that we should be skeptical of. Take ivermectin, for example. Ivermectin isn’t really a great treatment for anything. But ultimately, because there’s skepticism of authority, whether that’s in the form of government or media, people are becoming so contrarian that they’re saying, “In fact, I take ivermectin every day. I’m going to show you.”

It’s a funny dynamic, but I benefit from that dynamic. I’m glad that some people have taken a second look at me after the media coverage and said, “I rather like this guy.”

I also think, though, that it creates this world where admitting any nuance or gray area is taken as an immediate sign of guilt. You get this response of, “Hey, I just completely invalidate the other side.”

In your case, I’m sitting outside of it, and I’m a capitalist. I understand that you want to make a profitable drug so that you can serve a group, especially a group with a rare disease that may not be getting treatment anywhere. But I can also imagine that being involved in a scenario where someone can’t afford a lifesaving drug just seems so horrible. Did that actually happen, or never?

4. Designing drugs and defending orphan drug pricing

I’m a humanist. I would never let that happen. I didn’t get into the drug business only to make money. There’s also this glory, as I said earlier with the PKAN disease, in saying, “I’m going to use this God-given ability—whether it’s just a little bit of knowledge, a little bit of privilege, whatever it is—and I’m going to use it to do cool things.”

Writing code as a kid was awesome. I loved it, but the chance to make a medicine was 100 times cooler because, in the same way that you write code and it does something, you make a molecule and it does something. It’s a little more difficult to program, and it’s a little more abstract, but nevertheless, your body is kind of the compiler and you get the outcome.

Designing a drug is an awesome joy, and it sucks that this price increase—which, again, I still defend to the death—would obscure that. It’s a tough place to be when no one wants to hear your points.

One of the things that I tried to do—and, again, a lot of this is self-serving, to be clear—with infectious disease in general was, in 2015, to try to make infectious disease a more sturdy system. At the time, and really since time immemorial, the general idea with infectious-disease drugs was that we would develop them after the infection showed up.

It was really bad, and then it was, “Okay, now we’ve got to do something.” That’s a crazy perspective, because if you think about it, things like coronavirus were well known at the time. Toxoplasmosis was also well known. The drug—the one drug whose price I raised—was about 70 years old, and nobody had ever made a better version of it.

We should constantly be making better versions of this stuff and having the revision number go up. It’s not a pip install, but it’s close enough: You want to have a better version of this stuff.

For example, in the drug in question, the enzyme is called dihydrofolate reductase, or DHFR. It’s a very commonly studied enzyme, but the parasite’s enzyme and the human enzyme differ slightly. The question is whether you can thread the needle so that you make a drug that only blocks the parasite’s enzyme but does not block the human enzyme.

That kind of idea didn’t exist 70 years ago. We made one that just blocks the parasite’s enzyme. You could take gobs of it and nothing would happen to you, whereas with antiparasitics right now, if you take gobs of them, you would get very sick.

These kinds of upgrades were necessary, but generally the medical and pharmaceutical systems were like, “Look, if it’s not on fire, why fix it?” As a result, you probably read a story at some point about how bad the antibiotics business is, and the drug industry stops making antibiotics.

Well, now if there's antibiotic resistance, we really need that. The problem is that the lead time to make a new drug is about 10 years. So you're really talking about a tough space with really odd diseases: West Nile virus, for example. There's one called Chagas disease where there's no FDA-approved drug.

Chagas disease is actually quite common relative to something like toxoplasmosis. It's mostly a Latin American disease, and it tends to find its way to America often. Even without that, it's a very difficult disease to treat. It's a very severe disease, and no American companies or labs were ever interested in Chagas.

It's like, look, people are dying of Chagas disease. I don't think it was a bad thing to have a company that was going to stand there and say, “We're going to make the infectious-disease drugs nobody wants to make.” That company, again, could have had a coronavirus antiviral ready to go on day 1.

Infectious disease, as you know, becomes more important as the world becomes more local. But also, theoretically—and I hate to say this—the AI doomsday nightmare scenario, which I don't personally believe in, is the Eliezer Yudkowsky theory, the doomer theory: These AIs will let us just print viruses that will infect the world really quickly.

It's probably reasonable that some company has the capability of dealing with infectious disease really rapidly and aggressively. We were sort of trying to be that company at the time. It was an idea that was a little bit before its time, and it's still an idea that isn't going to be liked. People just hate the drug industry for some reason. It's a really terrible thing.

I think people in general should love the drug industry. For example, cystic fibrosis has now sort of been defeated by a great drug company in Boston. Many other illnesses, like hepatitis C, are now totally curable very easily. These are huge human triumphs.

I'm all for celebrating our triumphs in computer science, too, but these are really big triumphs in biopharma as well. We still kind of treat this industry with scorn, and it was really frustrating to see that and also kind of contribute to it.

It was probably a function of when you're desperate and need a medication and feel like it's expensive. That must be a terrible feeling.

It's interesting. We do a lot of work with almost all the pharma companies at this point, and we've had a whole bunch of those people on the podcast. That's why I was surprised to find a tweet of yours where you were saying AI and drug discovery have nothing to do with each other.

It was kind of a hot take because these days, I think over the last couple of years, most of the pharma companies have become Weights & Biases customers—and not small ones, right? They're big teams that are working on drug discovery. I'm curious: Do you think that's a wasted effort and impossible, or where do you come from on this?

I have super-high conviction on this, so I hope every VC watches it. I'm happy to discuss this matter further, but ultimately, in drug discovery, your major bottleneck is the idea.

It's a little like software, right? Your major bottleneck is, “What do I point my clinical-trial, medicinal-chemistry, toxicology apparatus at?” In general, the idea is that there's some protein target in the body somewhere whose function you want to change, stop, augment, or whatever.

A great example is Dupixent, which has become one of the largest drugs ever for Sanofi in France and Regeneron here in New York. It's a wonderful medicine. It's very expensive, but it's a game changer for a few different diseases: asthma, eczema, and so on.

The secret of Dupixent wasn't anything other than picking this molecular target. After that, making an antibody to something is very easy. We've had that technology for 30 or 40 years. Making a small molecule is very easy; this is what a chemist does.

Can AI assist with those 2 things, which are about 10% of the process—the actual invention of the drug? Then you have to actually test it in people, which is the other 90%. Those all kind of break down.

I actually did this on a livestream, where I broke down the entire drug-discovery process and pinpointed the parts where AI could do something. There were very small parts of the system that AI could help with, and probably the biggest part is the LLM part.

One of the things I did with my girlfriend—for full disclosure, she works over at OpenAI—was ask GPT to come up with a good drug target. It gave us a list that was okay, but then we said, “Well, let's put this in a for loop and ask, for every human protein, could you turn this into a medicine? Could an inhibitor, or whatever it might be—could interacting with this protein be useful in a disease context?”

If you actually ran this for loop and had a stronger system—GPT is great, but with a next-generation system—you actually could change the drug industry. The actual chemistry part of it, I think, is all there. If you show a chemist a protein or a crystal structure, you can make a basic inhibitor. It's almost like those kids who do their arithmetic tests very quickly. To a chemist, it's really not that crazy now.

You still have atomic roadblocks where it's like, “Okay, I have to actually make this thing now, and then I have to see if it works in real life.” The systems we have to test those in silico, as we call them, are okay, but you still want to know that it works in real life.

Then you've got to give it to some animals, which AI is not going to help with, because you have the lowest form of intelligence that we want to give these drugs to, to see if there are side effects and things like that.

In general, the idea that chemistry efforts would be helped by AI, or other drug-discovery components like antibody creation, is overstated. For 20 or 30 years, there have been different Bayesian approaches to get slightly better output. Instead of screening 100 million molecules, you can screen 10 million molecules or something like that. But we're not gated by that.

Screening 100 billion molecules is very doable with today's computers, doing it serially and by brute force. You don't get better results by screening 100 trillion or 100 quadrillion molecules. It doesn't really help you very much.

So the gating issue is that we're here because we love this technology, and we're really interested in it, and you've built one of the best companies in the space. But where can you actually leverage it in a way that's real and really going to be game-changing? I think that's the secret.

There's a disease—I'll just pick a random disease—say, Sjögren's disease, this weird autoimmune disease. It's Swedish; I always pronounce it wrong. There are various other diseases like this, but there's no great treatment.

I was developing a drug for a rare kidney disease called focal segmental glomerulosclerosis, or FSGS, another disease that's really tough to treat. Between Sjögren's and FSGS, the big win here is figuring out how we actually treat these diseases.

AI could help a little, say, in the lab. Some people are taking RNA readouts and having AI look through those transcripts, which is voluminous data and kind of difficult to parse. Personally, I think the harder thing is just reading the medical literature.

Very few people know this, but this is what got me into AI. Vivek and I had these rival companies, but we were also friendly, and our theory was what I espoused earlier: We need to go find the world's hidden gems of medicines.

They could be anywhere. They could be at big drug companies, small drug companies, or universities, but there's one common link: Everyone publishes in a database called PubMed. There are 38 million papers in PubMed.

What Vivek and I did was scale our teams up as much as we could for humans to read these papers. This was also when we had IBM Watson come in and try to teach us the latest in NLP.

I started a little NLP company because I basically looked at my team and said, “I love these guys. They're a group of PhDs and other really bright people, but there are about 18 people, and you have to look at the entire world's scholarly output on medicine.”

Vivek had a team of about 80 people, so he was scaling harder and faster. Our basic job was to read every paper about every new medicine and try to see if we should license it, buy it, or fund it.

I said to myself that the gating thing here is intelligence. It's human intelligence. I'll give you a really good example: There was an enzyme called Bruton’s tyrosine kinase, or BTK.

In the 1980s, there were studies around Bruton’s tyrosine kinase that showed it was going to be a key drug target for autoimmune disease.

The protein was a key player in autoimmune disease. Only 15 or 20 years later did somebody say, “Let’s make a drug and block this enzyme.” It turned out to be one of the greatest medicines ever. All that needed to happen was for somebody to read that paper and translate it clinically.

I feel like today you could do that with GPT. GPT could read every paper and say, “I think this could become a drug.” Slowly but surely, if you read all 36 million of those PubMed papers and say, “Okay, GPT, of the top 1,000, what are the very, very best ones?” Pfizer and Merck could say, “Maybe we should make this drug.” I feel like that’s the real area where we can win.

The human-intelligence part is really the gating item in all of these things. Remember, clinical trials are about 50% to 70% of the cost of these things, and that’s just paying people to sit there and wait and see what happens with their medicine. There’s a finite amount of help that AI can provide to begin with, and regulatory time probably isn’t going to be helped either.

There’s animal time, where you have to do the same exact stuff that you do with people, but with animals. So there are very few narrow parts of drug discovery where AI could even be applicable, and even in those parts, it feels like there isn’t a lot to do.

5. Why smart teams are still chasing AI in pharma

When I think of AI mixed into drug discovery, I don’t think about the transformer-LLM side, which is the side where I think it could be useful. What most people think about is, “Can you make a foundation model for chemistry or something like that?” I feel like that’s not going to be too fruitful.

It’s interesting, though. Everyone would acknowledge that most of the expense is in the clinical trials, but setting that aside, it does seem like there are lots and lots of smart people at big companies, and also at companies like Isomorphic and Recursion, that have built these big organizations and convinced investors this is important. Can you sell me on the argument a little bit? I’m trying to channel them because that’s been more of the—what’s happening here? Surely, where’s the difference in their thinking and yours?

I think they think drug discovery is harder than I do. I feel like they think target ID is something they can do better. I’ve seen some of the stacks, so to speak, of these companies, and some of them are correctly pointing to, “We’re trying to look at patient data, combine that with biomedical data, and then take all this data, throw it at a wall, and see what happens.”

The best medicines that have ever been made tend to come from laser-focused thinking: This is a big medical problem, say Alzheimer’s. How do we figure out how to treat Alzheimer’s? Nobody’s figured it out yet. There are some companies that say, “Let’s take blood samples or gene samples from 100,000 people with Alzheimer’s.” That’s a really good idea. It doesn’t require AI.

It’s plain old machine learning, like k-means clustering or something, to look at the data and say, “This is where the bad stuff’s happening.” Sometimes you don’t even need advanced math. You can just see that APOE is a protective allele for Alzheimer’s. Maybe we should make an APOE mimetic or something like that.

I think most of the ideas in drug discovery don’t require these hidden features, this representation learning that we’ve been taught AI can see in ways that humans can’t. If anything, you almost need to be able to see it for it to work. So much of medicine is fairly transparent.

Huntington’s disease is a good example. We know exactly what causes Huntington’s disease. It’s a malformation, a so-called polyQ repeat of the huntingtin protein, and it’s a terrible phenotype. It’s a lethal illness. The problem here is, how do you actually make the drug? The target is well known, and it’s one of these very frustrating diseases because you know exactly what you need to do, but it’s very tricky to do.

For example, CRISPR in general is a curative idea for all genetic diseases. The problem with CRISPR is that you can’t deliver it to target cells. CRISPR won’t get into brain cells, for example. It’ll get into liver cells quite well, but not brain cells. The technology required there is that you have to make a better screwdriver.

Could AI help there? Again, I think that to the extent we get superintelligence, it might be able to say, “What you’re all missing is that this is the way to sneak CRISPR inside brain cells.” But so much of this is serendipity, and so much of it is stuff that’s hard to plan. Innovation in general is kind of hard to force.

The people at the companies you mentioned are all good people, and they’re all really smart people, but we’ve seen this movie before in biotech. When the human genome was decoded, there was a company called Human Genome Sciences, and they said, “We’re going to be the winner here because genomics is a really big revolution.” They weren’t the winner.

When push came to shove, they had to make medicines. As you know from the VC game, once things sort of get realer, it’s like, “Okay, you have revenue now. Now we have to grow the revenue and grow the earnings.” That’s when all the imaginary stuff gets thrown away.

In the drug business, it’s the same thing. Once you put a drug into people, that drug has to work and change people’s lives for the better, or your company’s worthless. So many of those companies are going to cross into that chasm where it’s like, “Okay, all the talk’s over. What medicines do you have, and what do they do?” That’s just a biotech company. That’s just a drug company.

I feel like there’s almost never been a strategic advantage in technology in medicine. The strategic advantage is almost always, “What’s your idea?” When antibodies came onto the scene, within a few years everyone had the technology. To this day, antibodies are a modality for making a drug, but they’re not an end in themselves. AI is the same way.

I could use some sort of AI techniques to filter drug candidates for a disease. But if that disease isn’t severe enough to merit a drug at all, there are some diseases that are so borderline in severity that they’re probably not worth making a drug. Or if blocking that enzyme isn’t really going to do much for the disease, the AI is being misplaced and not used in the right way.

You still need that human judgment to think about, “What target am I going after?” That’s where I really think you could have GPT Pharmaceuticals actually make decisions like, “Let’s make a drug that targets this disease—say, Huntington’s or asthma—and here are the next steps, dear humans.”

I think that would work way better than trying to take machine-learning or deep-learning techniques and apply them to actual wet-lab work or drug-discovery work, which is still a relatively easy and mundane process that doesn’t need that extra intelligence. The extra intelligence—I guess the theme here is that the tail’s wagging the dog. The intelligence goes in the front end, and everything that comes downstream from there is actually fairly easy to do and almost deterministic.

The front-end decision—“We’re making an antibody to this enzyme. It’s this terrible enzyme, and if we don’t inhibit it, it’s going to cause cancer”—is the kind of strategy that has to be delineated first. Then, hopefully, the drug works and does great things.

So is that why you’re not really working in this industry anymore?

Not really. No.

Why not? It seems like you’re super passionate about it. Is it that you want to try something different, or does it feel like a dead end?

6. Moving from biotech to software and starting over

I have one regulatory barrier to it, which I do think could be overcome if necessary. But I also think there are a few other things. One is that software is just awesome. There’s nothing like it. It’s so great to write software, watch people use your product, and watch it grow.

The next thing I’d say is that the drug industry is very capital-intensive. When I started my first drug company, I did an outstanding job of holding on to the equity of the company, and I got to hold about 10% of the equity at the end. Most software people think that’s crazy. They think that’s a really low amount for a software company, but most biotech people are really jealous of me: “How did you get to own that much? That’s insane.”

Meanwhile, you have companies self-funding and owning 100%. So, to me, it’s a financially strange business.

You get a lot of scale packaged in a drug company, but you can get a lot of scale in software, too, and you don't have to take the dilution. So, as an investor, SaaS has just been a miracle. This didn't used to be this way, by the way. In the enterprise software days in the 1990s, you couldn't just open up an enterprise software company. It was really difficult, and it was a huge effort. Now everything's changed, and it's changing by the day. So I feel like the reward is pretty big there.

Again, pharma is one of these places where you can make the next Vertex. Vertex literally made the cure for cystic fibrosis, and Bernie Sanders said, “Why is it so expensive?” It's like, we just did a miracle. We're saving 30,000 people's lives that would normally die. I had a friend die of cystic fibrosis, and the first question is, how dare we reward these people for doing this amazing technical achievement?

What's funny is that you can be a great AI researcher—and you know this way better than I do—and you can plop yourself in San Francisco and make $1 million a year relatively easily, maybe stretching that a little bit. Whereas in pharma, it can require just the same amount of brainpower and you make far less, and your entrepreneurial goals are limited. Just to do a mouse experiment in pharma could cost $1 million, and meanwhile there are guys that reach PMF and profitability with that. So I feel like it's just a better business. It's more fun, et cetera.

I wouldn't rule out a return to pharma. But there are a lot of hoops that I have to jump through legally, one; and two, I have to get my head around whether this is worth doing. I've thought about doing it, certainly in Europe, where, for example, that pantothenate kinase-associated neurodegeneration disease—I still have these residual ideas for drugs that I think could work really well—and doing them in Europe, then sort of backing into the U.S., where I'm literally banned from the pharmaceutical industry, like one of the first people ever banned from an industry.

It's lawfare incarnate. But I still could probably persist, whether it's with the Trump administration or through Europe or something like that. So I would never say never, but it's one of these industries that's also been doing really poorly, especially relative to something like software.

That target ID problem we've been talking about for the last hour is such a difficult problem that, when there is a good target, every company has one. So when the PD-1 miracle happened in oncology, Bristol Myers Squibb and Merck both had this PD-1 antibody that changed the game for cancer: Keytruda and Opdivo. Within years, virtually every single drug company had a PD-1 inhibitor, because there just weren't that many great targets to go seek.

Even the ones that didn't work—there were a bunch of them, like TIGIT and some of these other ones, LAG-3, all these other oncology targets that kind of got hot because of a mouse study or something. Before you know it, 8 companies are already in clinical trials of a LAG-3 inhibitor. So it's this vicious competition to find a good target and prosecute and develop it to the point where you have medicine.

Then there's the capacity problem. It's kind of like the steel industry or something: there's too much capacity in biopharma, and there's too much money in it and not enough actual novel targets. And again, maybe I'll be proven wrong and some great AI tool can actually find new targets, but there are only 40 drugs FDA-approved every year. It's a very small amount.

A lot of those are like Tylenol XR, something that's just not very innovative. So there's really 20-ish drugs that are truly interesting advances. Some of them are not that great, and so you're looking at a very small number of drugs that are actually moving the needle for people, and the number of people that can benefit from those financially.

Well, big pharma wants 8 or 9 of those, or sort of 15 or 16 of those 20, and then really big venture, PE, and public biotech want the rest. So for somebody to just say, “I'm going to do this. I'm going to make it work,” it's a 10-year odyssey to get one of these drugs approved, and it costs hundreds of millions of dollars.

It's a business I fell into as a hedge fund guy, studying these stocks. And then they said I could kind of do this myself, too. It's certainly not a bad business if you can have access to huge amounts of capital and lots of really creative ideas. But even then, it's just not an easy industry either.

You're dealing with a lot of stochastic noise. You do a clinical trial—ketamine was a really good example. We were one of the first companies chasing ketamine, us and Johnson & Johnson, and the data looked so good coming out of Yale University for depression, for really serious depression. And we thought that the phase 3s would be a piece of cake.

It turns out they weren't a piece of cake. Ketamine has a great effect; it's better than SSRIs. But it wasn't the miracle that Yale painted it out to be and got us all excited about. It was a little bit south of that, a little bit north of SSRIs. And it turns out that development would not have been the easiest thing in the world. We ended up selling that medicine.

But it was one of these things where you don't always get what you want in pharma. And unlike tech, where you can make a change to the code base, you can't change a molecule. You start from scratch. So there are a lot of differences where it ends up being really tricky and really hard to always get things right.

So, actually, I really want to ask you about how you got up to speed in AI, because when I first talked to you, I was kind of astonished by your technical depth in that field as well. But I also want to make sure I touch on something I'm sure is on people's minds as they're listening to this: what actually caused you to go to jail, and what was that experience like? I know we've talked about that a lot.

7. What sent him to prison and what jail was really like

It was wild. I mean, it was really wild. It's probably what Sam Bankman-Fried is thinking right now. The guy went to math camp; he's probably the first guy ever to go to math camp and then go to federal prison. You probably went to math camp, too, right?

I didn't. My entire summers were just staying at home programming. But you're a fish out of water, that's for sure. Well, first of all, what was the actual crime here?

Fraud.

Yeah. So fraud's a very vague word, and the more and more we look at Silicon Valley, I think you could probably throw that word up a lot more if you wanted to.

And I feel like, for me, being the bad guy—being a world-renowned bad guy, sort of—I think helped the government say, especially—one of the ways I was a bad guy was not only telling the media that I didn't think their opinion mattered and that I could tell my own story; it was also telling the government the same thing. And I think eventually, sort of, push comes to shove, I was getting into a lot of fights with Hillary Clinton and some other folks.

I didn't care about the party or anything like that. What I cared about was that, ever since I studied economics, it is a company's right to charge whatever price it wants for its products and services. You can raise the price of your product or service tomorrow. Your users might go somewhere else, and that's just something you have to deal with.

You have to decide, balancing the customer's needs with your needs and your employees' needs and your investors' needs and all that stuff. You decide, and you may change the price higher or lower depending on the case. It's one of the first things you learn from Warren Buffett, for example, who—reading all of his work over the last 20 or so years I've been studying this stuff—pricing power is very important.

And so, for me, the idea that a company chooses its price, and the company's own owner chooses the price, is really holy to me. I got the sense that some of these politicians wanted to remove that power. And I was really like, look, you can't change that. That's what makes capitalism, that's what makes economics in America great.

The day that you say you're going to look at my portfolio and decide what the prices are going to be, you might as well move to another country. It's over. That idea was creeping in, and I think the politicians used me as a bit of a foil for, like, look how bad this guy is; we need to be able to have that power, too.

I was frustrated by that. I would sort of get into it with Hillary or whomever and say, look, you don't have that power. You'll never have that power. And I sort of liked poking my nose at them, saying, you can say what you want, but you're not going to change the price of my drug. You're not going to goad me into changing it, either.

Eventually, I think the powers that be tend to coalesce around—and I'm going to sound like a conspiracy theorist here, and I really did in 2015—this idea of lawfare, which is that here's a bad person; we're going to find a way to mess with them.

I had a hedge fund that I started that did very well. It returned about 4 to 5× capital. In hedge funds, that’s a lot more like venture capital; in normal hedge funds, that’s not normal. But it was a great result, and nobody lost a dime. I didn’t make any money either.

I fought the case really hard. It’s very hard to fight these cases. I was acquitted of a lot of the conduct, but a few of the charges did stick, including that I misled investors about our fund. None of the investors were upset, but it was technically a crime according to them. I still maintain full innocence.

If you go into a courtroom, it’s one of these surreal things where everything is stacked against you. There’s no way to win. It’s a 99% conviction ratio. We actually have a higher conviction ratio than Russia and China.

So, what was jail like?

It was interesting. I grew up in Brooklyn, and it sort of hearkened back to that time of low-income people from broken homes, a lot of poverty, and the root of some of these problems. It showed me there’s a really interesting world that you don’t exactly get exposed to, where I think people feel like they want to climb in society, but there’s no ostensible path for them, whether it’s through school or whatever.

I think there are a lot of mental health problems, like people with severe dyscalculia. I tried to do math with folks, and there were people who just could not add and subtract small numbers because of a profound learning disability or something like that, which probably went undiagnosed. Every person I met had one common thread: They came from either a single-parent or no-parent family, and I found that really remarkable.

I also read lots of business books, and so often the business leader would end up coming from a fairly well-to-do family, or a family that had two parents and a loving household. There are certainly lots of exceptions to that, too, but in prison there were almost no exceptions. I met 1 person, out of thousands of people, who had the two-and-a-half-kids-and-a-backyard kind of story and went to college. There seemed to be that common thread.

I ordered tons of books. I had a copy of the so-called Wizard Book, Structure and Interpretation of Computer Programs, the MIT programming book, and many other classics for nerds, whether Gödel, Escher, Bach or stuff like that. I’d just be sitting there—no computer. You’re not allowed to have a computer there.

I have a couple of funny stories. We would smuggle in cell phones and try to use those. There was this computer in MDC Brooklyn that was supposed to be used for discovery, so you could look at evidence on it and prepare for your case. It had a web browser. It wasn’t connected to the internet, but I was able to get to a developer console and change HTML elements.

There would be some guy, like some MS-13 gang member or something, and I would write his name and a disparaging comment. It would be really huge on the screen, like font size 72. I’d be like, “Hey, buddy, you look funny,” and everyone would laugh. They’d be like, “How’d you do that, Martin? It’s crazy.” Good old JavaScript.

It’s got to be really hard for Sam Bankman-Fried. The lack of a computer is almost like corporal punishment. It’s almost unconstitutional, in my opinion, given how wedded we are to these devices. Taking people away from a computer is, for Sam, especially difficult. He sent me some messages about this before he went to jail, and his biggest concern was the internet and a computer.

I think for a lot of us, that would be one of the major concerns. It’s almost unconstitutional, in my opinion, that it’s cruel and unusual punishment to say, “You’re going to get disconnected from the internet. You’re disconnected from all the world’s knowledge, and you’re just going to sit there and do it the old-fashioned way—reading books, using the telephone,” and stuff like that.

Did you feel like there could be benefits to that also? I honestly think if someone took away my computer for a while, I might learn better. I might reprogram my brain. I don’t know.

I talked about this with one of the who’s who in the VC world, and he was joking that I made it sound great. I read 300 or 400 books in prison. I made this list. I have to upload it at some point, but since I’ve been home, I’ve maybe read 1 book cover to cover.

Wow.

You just don’t have the time if you’re running a startup or something. But sitting there, I could read books about whatever. I could read a book a day sometimes, and you just have nothing else to do. It was so enriching and awesome. But you would still rather have the computer. I think you can at least make a silver lining out of it.

Didn’t you feel angry about the situation you were in? You don’t seem angry, but I don’t know. Given what you told me, I would think you’d be feeling really frustrated.

I think you probably know this really well, maybe better than me, as an entrepreneur: You just have to roll with the punches. You lose a client, you lose a key employee—stuff like that happens, and you go to jail. These kinds of things are setbacks, but they’re not going to define you.

Every month at our startup, we try to set a new revenue record, and we try to have it a certain percentage above last month. I don’t know if we’re going to make that number this month. It just started, but I’m nervous about it already. It’s one of these things where I can only control what I get angry and nervous about.

Ultimately, getting nervous about something that was this cascading effect—I was one of the last people who was canceled, so to speak, before cancellation just stopped because people stopped believing in it. I was one of the last people to get skewered by the media before the media sort of fell apart.

8. Mindset in prison and advice to Sam Bankman-Fried

Whether or not those are just copium and my own excuse for it, I still have to live my life. Getting frustrated and angry is just not going to be very useful or healthy for me or anyone around me.

Let me calibrate. I spent the last 15 years of my life worrying about revenue numbers and other KPIs and losing sleep over them, so I totally understand that. If that’s a 5 on a scale of 1 to 10 for anxiety, where is thinking about going to jail in your future? I would imagine that would be a lot higher.

Yeah, it’s like an 8 or a 9. My father passed away a year or so ago, and that was a 10 to me. It was actually harder than prison.

There’s an old saying in criminal justice: You only do 2 days, your first and your last. I do think the Sword of Damocles is actually worse than the actual prison time. I also got 7 years, which means you sort of do around 5 of that. For Sam, he’s got 25 years.

We don’t know what they would have given. I think maybe they did sentence Andrew—I don’t remember—but you’re rolling the dice with the judge. If the judge doesn’t like you, they can give you 30 years. There’s nothing you can do about it.

Did you actually give Sam advice? Did he call you?

I did. I kind of feel for him. A lot of people don’t like him and just think he’s human garbage. Ultimately, I think he should get punished, but I don’t think 30 years was the right number.

He sort of operated a financial institution the same way as JPMorgan. Look at Silicon Valley Bank, which went under. Financial institutions are a little bit wild. He also sort of helped create Anthropic in a way, right? I don’t know that he was this terrible mastermind criminal.

So, what did you tell him?

To be explicit, I basically said that prison is a very different world from the world you grew up in. There are a lot of minorities and people from different walks of life that you’ve never even thought of. You have to respect that there are people out there who are going to like different things and do different things, and you just have to roll with the punches.

There are a lot of cultures that you’re not familiar with. I have a passing familiarity with some of them because of where I grew up, but he really grew up with a silver spoon. He basically has to get rid of that and become a more malleable person.

I think he is fairly dedicated to the person he is, and he’s going to nerd out through prison from start to finish. I don’t think he should become a gang member, but I also think he has to be flexible because you want to be comfortable in prison. You have to communicate with people, and you have to be friendly and charismatic.

I don't know that he's perfect at that. He had a lot of anti-charisma and tended to put his foot in his mouth quite a bit. I wasn't the best, either, but I also feel really bad for the people who get super-long sentences. I think anybody on planet Earth could do a year or two in jail.

I think that when you get 25 or 30 years, it's got to be soul-crushing.

All right. Well, I guess switching back to AI, do you want to talk about the company that you've built?

9. AI to TTS to building a Bloomberg competitor

Yeah. I started doing some AI work, and I ended up pivoting back to more traditional software. We tried to make TTS stuff, and we tried to do different things in the LLM world, but we couldn't find a great fit. The TTS world was really awesome, and I loved it, but I was always sitting there trying to convince others, and also myself, that this would be a big market.

I remember Microsoft's acquisition of Nuance. It's a really big market, but ultimately they succeeded not by making TTS technology or speech-to-text ASR; they succeeded by packaging it with a product, especially one targeted toward doctors and things like that. To me, there's this age-old question: Do you want to build the technology, or do you want to build the product? What layer do you want to be in?

I think it would be fascinating to see how companies like ElevenLabs progress. One of the problems that I constantly thought about, again trying to convince other people, was that we're going to have an audio web. Maybe not every part of the web is going to be audio-driven, but even a little bit more audio is going to be a big deal, because with LLM output, people don't want to read all this stuff.

I almost called the company TL;DR because people just don't want to read and read and read pages and pages of text. They'd rather hear it. That's why people watch podcasts and listen to them: it's less cognitive load than sitting there reading a transcript.

I felt that after a while. I was just convincing myself, and I'm not even sure I believed it. I think the TTS space is around $1 billion in total revenue across every company. A lot of that revenue lives at IBM, AWS, GCP, and Azure—companies that you wouldn't necessarily think of as great providers.

I don't know if people are going to make money in this space. I feel like it's a very niche space. We still wanted to do it because it was a lot of fun, but I feel like it's just super-niche.

Even though I thought the Sesame AI demo was really amazing—this AI voice thing—I feel like there will be room for that. I'm of 2 minds about it. I feel like every content company, like Disney, should have a Darth Vader or whatever characters they have custom-made for you and your kids.

There's no reason SpongeBob shouldn't have a dedicated episode just for you every day. This is all doable with TTS, and increasingly with other companies, but you also have the bigger companies that have better scale. OpenAI does TTS and things like that.

We gave up on it because we couldn't get real revenue traction. One of the funniest things, and I think you'll appreciate this, is that we took a saying by Greg Brockman really seriously. He said that it's easier to teach good software engineers AI than it is to teach AI people good software engineering.

We tried, but it wasn't that easy. We were sitting there reading everything we could find on deep learning. We rewrote a bunch of Torch, and I made a pull request—my own pull request—to DeepSpeed, the Microsoft library. Ultimately, this stuff isn't easy. There's a reason it's very difficult, and it's frustrating, too.

10. Why trading data is like code, and what makes his product work

When I write software, I get the satisfaction of shipping every day or every week. With an ML project, you can easily spend months on it and just think, "I got nowhere, and I don't know why."

Yeah.

It was really challenging. We followed this wild TTS story about a guy named James Betker, who more or less solo-developed Tortoise TTS. He's at OpenAI now, but at the time he had open-sourced Tortoise TTS, and lots of companies tried to take it and make it a managed service.

I think the key there was that first-mover advantage was really important. Very high-quality audio was important, too. Every little mistake that a model made could be disastrous to your reputation. It was one of those word-of-mouth things where people would say, "What's the best TTS?" "Oh, it's definitely this."

When you get that kind of word of mouth going, you're going to do really well. But I still wonder if this market is going to be that reliable. One of the fears you have to have in the back of your mind is that, as compute becomes cheaper and data becomes cheaper and these models become open source, what's to stop some model on Hugging Face that's really good and can run locally from just being used? Why do we need anything else?

Speech still has some issues with trying to get the context of the whole page or the whole book. A lot of times, people just chunk speech into sentences and pray that it has the right prosody and things like that.

It's very different if, in one sentence, somebody says that this character is a bad guy. When they say the bad guy's name, it might inflect differently from how it would if the character were his lover or something. That kind of context is missing from TTS.

These last-mile problems will be met with dialogue models and really cutting-edge technology, where people are talking over each other and the speech sounds really natural. It's going to happen. I don't know that there's some big customer that's going to do really well with it.

My dream was that customer agents and service agents would need to be equipped with these TTS voices. We'll see how these companies do. I think it's a really cool space, and we really liked that part.

I also thought about doing AI for physicians, which was its own interesting thing. We see every day that knowledge workers are going to get disintermediated. We're not going to have knowledge workers anymore.

Healthcare costs are something I know a lot about and have been in the news about. A big chunk of healthcare costs, if not the majority, is human labor. A lot of that labor consists of workflow charts and flowcharts and things like that.

For hypertension, if the patient is not on any medicine and they're at 130 systolic, use this drug. If that isn't working, use this drug. If that isn't working, use that drug. It's actually fairly simple stuff. A doctor is important, but you could pretty much encapsulate all of medicine in an LLM if it's basically already happened.

The question is, how do you actually get that to cut actual, real-world costs? I think we're going to learn that doctors have this very serious lobby, and they enjoy these huge reimbursement fees. The idea that a machine could do it better is very scary to people.

We had a nice talk with a Middle Eastern country that didn't care. They said, "We can cut our healthcare costs by a huge amount by just forcing people to use an AI doctor." This was a country that was willing to just say, "The hell with it."

I feel like eventually that's going to happen in the U.S. Two-thirds of this deficit—or 3/4 of this budget deficit that we have; I don't know, maybe it's half—is Medicare and Medicaid. This gigantic deficit that's causing all these problems, and that could result in U.S. insolvency someday, is just healthcare.

All of that could theoretically be done by AI, and maybe someday by humanoid robotics, too.

Why did you do it?

I think we did okay with it. It felt like getting commercial traction would be difficult. Other people have danced around this bush. A lot of people are taking it from the clinician-payer side, which I think is smart.

We were trying to do it from the consumer side—to get people to reduce their costs and, one day, maybe get insurers to pay for it. I feel like we didn't get that much traction. We got a lot of people trying it, but we also felt at the end of the day that this could potentially be the big value proposition for the foundation-model companies.

Elon actually tweeted that people should use Grok as a second opinion. You probably have to be pretty careful about what he says, but that's true. It's probably a better second opinion than most doctors, and what an awesome thing.

But do you want to compete with that and have your own dedicated LLMs for specific areas?

I don't know that those are going to work. Look at Cursor and stuff like that: you're just piping into the foundation model LLM. So you're not making a code-specific LLM, at least with the same technology or the same sort of framework or architecture.

11. Lessons from building a finance product and achieving product-market fit

I don't know that a medical LLM makes sense. It's just a regular LLM, and that's really been done well by a bunch of companies. Trying to do that didn't make too much sense for us. It was one of these things where we got usage but not revenue, and B2C in general is something where it's really hard to get revenue.

We settled on this financial product that I've always had this dream of. As a lifelong trader and lifelong investor, building my own sort of financial suite was always my dream. In fact, that's how I fell in love with JavaScript. I started with BASIC and Pascal and stuff like that, and I never programmed in JavaScript until I literally got arrested.

That was the darkest of moments. I turned to JavaScript. I had some free time—I was going to fight the government and prepare—and I basically said, "Let me go back to an original love." I very quickly realized that this was awesome, and I missed it from my youth.

I started to make some little financial programs, and it evolved into this really big suite that we have now. It's the most fun I've had in a really long time. But in terms of AI, I think the pressure to put AI on everything is being seen in finance a little bit, in finance software. It still hasn't really been done right, so I hope to do something right there.

We're almost completely traditional software now. I don't think we have any Torch in anything we do right now. At some point, I do think that we can do some cool stuff with AI, but I feel like the contrived kind of "chat with this," "chat with your documents," or whatever is a little bit too forced. I do think there's a huge use for AI in finance; it's just not so obvious.

I was watching your livestreams, and you gave me a super cool demo of your product. For people who aren't traders, it appears to me that you're using this monster BI tool where you pull up tons of graphs and tables and stuff, and you seem to like having zillions of metrics spread out all across your screen.

I was looking at that as a nontrader and thinking, "I'm not sure that's how I would want to ingest data about my own company." Is there something different? Are you just feeding your ADHD and anxiety as a trader by putting so much information there, or is it actually helping you trade better?

It's a really good question. If you walk around a trading desk, it looks like that. People are constantly on the phone. They're ADHD whether they like it or not, and sifting through lots of securities—all these things are, to me, very reminiscent of a programmer's job.

If you're going to spend 8 to 10 hours a day coding, you want the exact right command-line setup. People get very serious about their Vim setup. Every little thing has to be perfect. I'm not like that, but there are some things that I like, and if I'm going to do it for 10 hours a day, I might as well get those little things done right.

I think financial software is very similar. If you're going to spend 10 hours in the office every day—and it's very well known how long bankers spend in the office—you are going to look at stocks or bonds or whatever you look at. You're going to do it constantly, over and over again.

I feel like you have to give people a command-line-style tool that helps them get that information quickly. It's going to be a little bit ugly and scattered and all over the place, but that's sort of a feature instead of a bug. If it looked like a SaaS site, it would frustrate the investor, because they would think, "Okay, well, I just looked at Apple, but now I want to look at something else. I want it to take a millisecond. I don't want it to take refreshing this page or trying to hydrate the page. Just show me the data right away."

I feel like that's really important to Wall Street. In terms of AI, there are a lot of places where this could help, and I know that a lot of the quants are starting to do work with LLMs. We're trying to race to provide those tools, too.

I think the stuff that I've seen—chat with a 10-Q document or a 10-K document—can be a little bit simple, but a lot of the time it works. I don't know that, but what is a trader doing? You're looking at a bunch of graphs and then trying to make some prediction. That sounds like—how are you possibly going to use AI for that?

I think AI trading is going to make quant trading even better and even harder for humans to do well. Humans have a small edge when it comes to judging markets, the emotional states of other traders, and integrating lots of disparate pieces of information.

If you see a little clue about inflation from one company and then another little clue from another company, you start to connect the narrative: "Okay, maybe inflation is going up. I'm going to do something with bonds." That's the kind of thing that maybe AI can't see right away.

But in terms of super-high-frequency trading, computers have dominated that. Trading the news is something that I'm really interested in, and I know a lot of other LLM people are interested in it, whether they're coming from the quant side and moving toward the LLM side or from the LLM side and moving toward the quant side.

The question is, can you have a computer that just reads breaking news really quickly and reacts to it really quickly? That's one of the products we're working on that I think is going to be really helpful. You'd be surprised at how much breaking news hangs in limbo, with the market still trying to figure out whether the information is real.

The LLM can make that call in a second and then make the trade. I feel like that's going to be a big game changer. But, of course, like all things in finance, everyone will have it within 6 months or a year.

You can very quickly react to things. Just today, Hims announced that they were going to carry a GLP-1 product, which was going to be good news for HIMS, but it took the market a few minutes to really digest it. The LLM would immediately say, "Yeah, buy that."

I feel like that is a product that certain big firms are making, but nobody's making it for the little guy. I'd like to make those types of products for firms, too, but also for the little guy. There are all kinds of other little nuances in finance that could really benefit from this.

Traders themselves have mostly been wiped out by computers. There are still hedge funds out there with humans, but they usually pale in comparison to the quant hedge funds. We're also considering how to bring quant tools to people, too.

Wait, are you trading? It looks like you're trading on your livestream. Is that so? Have you yourself not been wiped out by computers?

No, I think there are still people who could do it. I think it's a losing proposition—a fool's errand. It's sort of like Kasparov versus Deep Blue. You could see the writing on the wall even years ago, and today there are still good chess players, but it's really hard to play against Deep Blue.

I feel like you're going to have a time and place where private equity or venture capital are areas that AI is not going to compete with for another 10, 20, 30, 40, or 50 years, maybe. Somebody's going to come up with a startup that competes with yours, and you'll say, "Who are these guys? Is it that guy I met at MIT, or is it this guy I know from Anthropic?"

Then it's like, "No, it's a machine. It's an AI." That would be really, really ironic, because you'd sit there thinking, "Let's unplug it or something."

The reality is that the markets have always been competitive. I feel like there are some people who believe—they often trick themselves into thinking—that they have an edge. One of my goals is to help people see that signing up for Robinhood, putting your weekly paycheck into it, and trying to guess whether the stock is going to go up or down is probably not something you should do.

Is there a way to do it a little bit more intelligently, in a way that is less risky and maybe more rewarding? That's one of the things we're trying to think about.

I agree that trying to make trading software, or software for research analysts, in 2025—when the research analyst has kind of become a dinosaur and a thing of the past—is something where, certainly for us, we understand that Wall Street has completely changed.

It used to be a cigar-smoking, blue-blood, sort of frat-bro-type environment in the ’90s, and now it’s mostly quants, IMO winners, stuff like that. I feel like you have to make tools for those guys, and so we are sort of trying to get there. I think we will, but I also think Wall Street hasn’t collapsed yet. There are still sort of people doing that stuff.

You also have a lot of other things you could do with this type of framework. Bloomberg is a really interesting company: they’re the time-series data company, and they do it really well for stocks, but they kind of stopped doing it for other things. I think that you could actually look at tons of unstructured or structured data and render it more easily. Google and LLMs can do that better with unstructured data, but if you know what data you’re looking at and you look at it every day, we can graph that. It doesn’t have to be a stock or a bond; it can be weather, or some other shipment data or something.

I hope that we can sort of show that stuff. It’s working really well. We have millions of dollars of revenue now. We’re really lucky. It was really hard finding product-market fit. We tried 3 different things, and none of them sort of connected. The first one didn’t connect at all with anything. The next one connected a little bit with users, but not revenue. The next one, the TTS product, connected a little bit with revenue and a little bit with users. But this product has just sort of gone off the charts.

The thing that I admire about your product, from, I guess, one entrepreneur to another, is that I think it’s very similar to Weights & Biases. We were kind of building this really specialized thing for an audience we really understood, and more than I think we liked them, we loved them. We came from that place of really wanting to make something good.

I can kind of see from your livestream that you’re doing the same thing. You’re a user of your product, you know the people using it, and you have a very idiosyncratic UI that clearly comes from a lot of caring about it being a good product for your audience. You can really feel that in your product. I just really admire that from afar, although I’m absolutely not a trader.

Yeah. This is what Tobi Lütke said. I saw a video the other day where Tobi Lütke said something like, “The only way you’re ever going to make a great product is if you really care about it, you really love it.”

I feel like we probably should have seen that all along, because I love investing. I love trading. Every little detail is something that I obsess about, because I know what I would feel if somebody were selling me this software. I’d say, “No, I don’t want to use this anymore if it’s going to have this thing there or that thing is going to be wrong.”

The developers don’t always understand that. We actually try to find people who invest, or encourage all of our team to invest, so they really understand what the user’s going through. Maybe not more than 10 minutes a day sometimes, but at least a little bit of market activity so they can sort of understand what the user’s going through.

I feel like it’s a really amazing area, because traders have money. They love spending money on software, and it’s a really remarkable, awesome thing. We sort of sat around the first year or 2 of our company, meandering. I don’t regret it, but now that our revenue is going crazy and we’re hiring like crazy, I’m really lucky and happy that we’re hitting that.

I also now have to start thinking about harder things, like: How do we 10x our revenue? How do we make products that people really, really like and that are really valuable for them?

I think, again, stuff like Robinhood and other companies helps users gamble, and I want to help our users actually make money. It’s something I’m pretty passionate about, and we’ll see if I can get that to work.

We still haven’t actually released our product, so it’s all been in this weird beta stage. One day we’ll tweet about it or put it in The Wall Street Journal or whatever, and it’ll hopefully 10x or 20x then. But for now, this is actually very hard software to make.

Do you want to put a link in our notes or tell us what it is?

That’s fine, but I saw Perplexity was trying to do a finance thing, and I think Chamath sort of made a wisecrack about it: just showing 1 or 2 charts doesn’t replace Bloomberg. I do think that’s true.

It’s really hard to negotiate with the Frankfurt Stock Exchange, and then Osaka and Tokyo, and herd all these cats and get all these vendors and write tests to make sure that, if any of that data is late or something like that, you do some emergency thing. It’s mind-boggling how complex this machine is, and that’s sort of a moat in a way.

The UI, I think, is where things really shine. I mean, it’s Vim for finance in a way, and I think that, to anybody who’s stuck sitting down in an office just hacking away at finance, it’s code in a different language, and this is the sort of thing you need.

I’m really lucky and happy. Again, I had a lot of patience from investors to sort of find our footing, and now we’re in the same grind that everybody that gets product-market fit is in: What’s the revenue? What’s the growth? How fast can you grow? Every year is going to be—every month is going to be—a challenge.

It’s more fun searching for product-market fit, I’ll tell you. It’s better.

Yeah. I think, looking at the Levels of the world, I actually tried to buy a company, Levels. I called him. I sent him a DM, and he said, “Oh, great, Martin. I love your stuff. I’m a fan.”

I said, “That’s great. Give me your number. Let me give you a call.” He said, “No, I’m too—what are the words? I’m too—I’m grinding right now,” or “I’m too locked in.” I was like, “Nice. I’m trying to buy your company.” I didn’t even say anything. I was just like, “All right, man. Let’s stay locked in.”

I think these solopreneurs are really something. I think that you’re going to see more and more of these. We may get our first solopreneur IPO someday, where somebody’s like, “Yeah, I built this all by myself. I built a whole SaaS company. I own 100% of the equity, and it’s a multibillion-dollar company.”

I think we’re going to see that more and more, thanks to companies like yours and others that have built this amazing ecosystem.

12. AI bubble predictions and market rationality

You can never do it for our company, by the way, so don’t even try. But the complexity is just—you need at least 10 people, or however many you have.

Exactly.

Okay. Well, actually, one of the spicy takes I found of yours from, I don’t know, 6 months ago or something, was saying that AI is going to enter a bubble and the bubble’s going to pop, which you could argue was extremely prescient. I’m not sure where you feel like we are in the hype cycle, or if you’re expecting a bigger bubble. It does feel like we actually had maybe a peak 2 months ago, and maybe we’re coming down from there. What’s your take?

Yeah, it’s hard to say. I think that in the ’90s, this happened with dot-coms, too. It wasn’t just dot-coms; it was a mix of dot-coms and other tech, including communications technology. You had this bubble, and it felt like maybe you were getting to a sigmoid, but then it sort of did another big gap. I feel like that’s probably going to happen here. I think this is maybe one of the bubbles to end all bubbles.

One of the funny things about every bubble is that sometimes you actually sit there and think, “Maybe it’s real.” I think the smartest people, after a while, are like, “I don’t know. Maybe I should do internet stocks.” I feel like that’s exactly what’s going to happen here.

I think OpenAI will go public. I think it’ll be a trillion-dollar market cap. I think the bubble will get bigger. I feel like it’s possible it all comes tumbling down tomorrow, but I also feel like, almost more than the internet, AI has an immediate impact for a lot of organizations.

The other key thing is that the biggest beneficiaries are going to be companies like Verizon and Procter & Gamble and stuff like that, where they’re going to be able to save a billion dollars here and there, which is a huge amount for the S&P 500. They can save some money deploying AI tools, and maybe they can spend money using products like yours to say, “What can we do with the data that we do have internally?”

Really smart companies will pass those savings on to efforts like that. But I think the internet did that for so many different S&P 500 companies. It’s what fueled global growth.

And so AI has been fueling global growth and will continue to make GDP growth and productivity go up. I feel like that is when a bubble can actually continue to inflate. Once you have OpenAI with a lot of revenue—in the dot-com era, nobody had revenue—you have actual progress, an actual product, and amazing technological advances. This is sort of the real deal.

The big question is, what do investors do about it? I feel like investors are somewhat sober. You have some valuations that look a little bit large, but by and large, folks aren't funding companies just because they have AI in the name. Back then, if you had dot-com in your name, you'd go public or instantly take it public. Today, it's not that easy.

The big firms are not just funding anything. They're being fairly judicious about where they put money to work, whether it's Sequoia, Founders Fund, or a16z. I think that when you start seeing people get really sloppy, it's kind of like when the bubble has really hit. Once those firms start doing really dumb stuff—which they tend not to do—but even the smartest people get wrapped up in this, I think then we'll know the bubble's really here and going to blow up.

But for now, it sort of worked. I feel like you're closer to the vest on this. You're in the thick of it, but so far, the valuations that I see for brand-new startups—the founders want a lot, but oftentimes the investors really do get to make the call. I don't sense that it's really, really bad yet. I feel like 100 times sales for a SaaS, fast-growth startup is actually fairly rational. You see that number thrown around a bit, so there's real substance here.

I think one of the problems is going to be: How do you actually take AI and make something real, where just having the idea itself is probably not going to do very well? If you try to say, "I'm going to do AI for X," that's sort of a recipe for disaster. But I do think that really focused ideas and really good teams will be able to do it.

I also feel like, at the end of the day, your customer and business rules haven't changed. Your customer matters, your product quality matters, and all those things matter. I see some companies—and you probably see this too—where they feel like they have a really good engineering team, they feel like they're going to do something really special on the technology side, and they don't connect at all with customers, and they don't do anything. I feel like that has happened, but I wouldn't call that a bubble bursting as much as it is a misapplication of technology. This happened to dot-coms too.

So your view is that we're at the start of a bubble? Is that a good summary of your view?

Yeah. It’s so hard to explain, right? I feel like in the dot-com bubble, there was just indiscriminate craziness. It was mania. You don't get that mania today; it's still very rational and sober. When you see something really wild—sometimes some of the robotics stuff that's happening right now is a little bit of a sign of non-sobriety.

If you see a robotics company with no product and no revenue raising at $100 billion or something wild, that's obviously bubble behavior. But at the same time, you sit there and think: It's human labor. Maybe that could be a $100 billion company. Maybe it could be a $100 trillion company.

I feel like there's relatively little of that. I don't know that we're even in a bubble. A lot of it is rational. We'll have to see. Again, you see more than I do, but Wall Street hasn't had this parade of IPOs. There have been some IPOs where companies have had to temper their expectations from the private market.

So, to some extent, there's a lot of VC money and the roles have reversed. In the dot-com bubble, there wasn't much VC money, and companies went public to get the stupid money and the crazy prices. Now you kind of get the crazy prices privately, and when you go public, you have a reality check: "No, no, Wall Street's not going to buy that."

I feel like because the private markets have been disintermediated by these mega VC and growth funds, we actually can't tell, because price discovery isn't necessarily there. You'll have a $5 billion company announce it's shutting down, and you're like, "How's that possible?" It's like, "Oh, the VCs quit on it, and it actually wasn't worth $5 billion. It was just a mark."

I think so many things about the ways you can even detect a bubble have changed. We'll probably know the answer to this 5 years from now way better than we know it today.

13. Wu-Tang, art, and collecting old computers

Awesome. Well, I've kept you over. I want to let you go, but I did have one last question. I saw you bought the Wu-Tang album. Were you a Wu-Tang fan?

Yeah, I'm a rap fan. I'm from New York, the home of hip-hop. I think art is one of these cool things you can do with money, and when you make a lot of money, there's the patronage of art. So many people will buy a Picasso or Monet or Manet or whatever and say, "This thing's really expensive. What a trophy." I feel like if you sell a startup or come into some money, one of the best things you can do with it is make your own trophies. I feel like this is the coolest thing you could do, assuming you want to do it.

The most complete dinosaur ever found was recently sold at auction for $40 million. That's something I could see a founder like you, or somebody who had a big success, saying, "Oh yeah, I want the world's best dinosaur in my backyard."

For me, here was this music album where nobody had ever paid a lot of money for a piece of music. I thought it was a really cool idea that should be explored more because music's so important to us, whereas Impressionist art doesn't play a role in my everyday life. For a lot of us, we have our headphones on, we're listening to music; it's a huge thing for us.

So why doesn't Taylor Swift or Drake or somebody like that make albums for a patron? They get to really channel their voice without thinking about commercial output and just say, "I'm going to make the best art I can, and whether people like it or not, it's irrelevant because Elon gave me $10 million to do it." I feel like that's potentially a new paradigm for art, and it was kind of the old paradigm for art. I feel like that was a really interesting statement to make.

I've done some other things like that where the traditional things you do with money can often be really trite, and it's not you. You buy a painting, you're like, "I did this," or you buy a summer house, or you do all the rich-person stuff, and then you're like, "This isn't me."

For me, one of the things I really wanted to do was collect really old computers, like the first Apple computers—really, really old things like that. I collect letters. There was a museum in Seattle. Did you ever go to it? They had all of them working. It was amazing.

I haven't been yet.

Oh, wow. I think it was one of Microsoft's founders, Paul Allen, who was also a collector of this stuff and made it available. So, if you do get a good collection, let me come over and play with it.

No, it's funny. I actually loved the Wu-Tang. I remember when the Wu came out when I was a kid, and it really spoke to me as definitely not the target demographic. I remember going to a show in 2010 and running into 20 other startup CEOs that I knew in San Francisco, just being like, "Huh, this is an absolutely terrible show." I think they had gotten old and were really bad, but it's just funny to see that connection with a lot of other people at once.

Someone did a numerical analysis, or some other data-science analysis, of Wu-Tang's vocabulary, and it was many sigma to the right of virtually any other rapper. It's a group of guys, so it's a little bit different, but even still, the lexicon of Wu-Tang is massively larger than almost any other group. I thought that was really interesting, and the use of metaphor was also very significant in their work.

They had a weird mythology that I really got into. So, yeah. Anyway, I'm also a fan.

Anyway, thanks for taking the time. I really appreciate it.

Thank you so much.

All right, take care, man.