No Priors 第116期｜与 Sarah 和 Elad 对谈

Elad, what's going on?

How are you doing, Sarah?

I'm good. I can't tell whether this is a very stable time in the market—whether it's crystallizing into known businesses and models or whether it's still fluid. What's your take?

AI is the one market in my career where I've consistently said, "The more I learn, the less I know." Every other market, you learn more, you know more, and you keep advancing. I actually feel like that's shifted in the last couple of months, where, despite the rapid pace of innovation and all the really exciting new models, research findings, and everything else, I feel like a bunch of markets have consolidated. It's clear now who the likely players or winners are in 2 or 3 big areas.

That may change, right? In 3 years, another new startup may launch and displace everybody, or an incumbent may make a bold move, or whatever it may be. But I feel like in the foundation model market, at least for LLMs, there's a clearer view of what's important and what isn't. At the application level, I think it's clear who the winners are going to be in at least the first set of services for healthcare-related things, like medical scribing or other workflows.

In coding, it seems like it's consolidated into 2 or 3 players. Maybe that's Cursor, Codeium, Cognition, and then Microsoft's Copilot, right? There probably aren't 2 dozen companies that are all still competing there. In customer success, it seems like things are consolidating around Sierra and Decagon.

You go market by market and you're like, "Okay, there's a bunch of markets where it's clear who we think some of the winners may end up being," or at least who the important companies will be for the next 2 or 3 years. Then I think there's a set of markets where it's still wide open. You look at sales productivity tooling: there's going to be something really important there. There's going to be some financial analyst tool that's really important, and there's going to be an accounting company that's really important.

The question is, has that not consolidated yet because nobody is doing the exact right product approach? Is it because the models aren't good enough and the capabilities have to get better? It feels like there's a bunch of stuff that's still unknown, but it's way clearer than I think it was a year ago.

For the first time in 2 years or so, I feel like there's more clarity. When I first started investing in generative AI, you just went and backed the things where the people seemed really good and the market seemed interesting, because there wasn't a lot of competition. That's when I led the seed round for Perplexity or invested in Character.AI, Harvey, or some of these other things. That was pre-ChatGPT or pre-Midjourney.

Oh, the good old days.

Yeah, the good old days, when nobody cared. When GPT-3 was out, everybody was like, "This is kind of crappy." But the scaling law was clear, right? I thought a handful of people—you being included—we collectively saw that this stuff was going to be important.

But then there was a period of uncertainty for 2 years or something like that, maybe 3 years, where there was so much innovation, so much change, and so much rapid growth. I think now, finally, we're hitting a period where at least a subset of things are consolidating back down. Again, these may not be the winners 5 years from now, but they definitely seem to be emerging as the winners for the next 2 years.

I think it's a nice breather in terms of uncertainty and having a bit more clarity into what's going to happen. I don't know. What do you think?

I feel a little bit like I understand some temporary physics of the market a little bit better. It's like a race to find the verticals of relevance and then get something to work in a way that users actually want. Maybe you have to go get proprietary data sources that you can retrieve against and get distribution, and then ideally have users who can create, derive, or extend knowledge from that, like the companies you just named.

I don't think you explicitly said it, but I put Abridge and OpenEvidence in that category. I think they fit into that shape. One thing you and I have talked about is that I'm actually quite unsure about sales. I don't know how to think about how something wins there. You could go at it from a data perspective or an adoption perspective, but it's been a very fragmented market to date.

I agree with you on finance and accounting. I'd add pharma to that. There are some industries that are really document-driven where you can see something coming there. There are companies in networking and pharma, for example, that are kind of interesting.

And so, to some extent, it's been clear what markets will be interesting, or at least a subset of them. It just wasn't clear who would win and how. Coding is an interesting analog, where there were probably 4 different approaches to coding that everybody was taking simultaneously.

I think some of those approaches will consolidate over time, but the entry points now seem much clearer in terms of how you actually win in that market. Before, 2 years ago, there were a dozen different ways you could imagine somebody winning. I wonder if the analog there is the sales stuff you're talking about, where it seems a little bit less certain right now, but maybe in 2 years we'll be like, "Of course, it was whatever that workflow was."

We had a debate internally at my firm about what it would take for another new entry point to work. I think it would take a lot. I'm open-minded to it, but what is still changing is that you increasingly have open models and little models that can do real things with code. Codestral and this—I think you'll see more there.

Microsoft open-sourced Copilot. We'll see what the impact of that is, but it's like they finally decided they need to fight Cursor from eating its lunch with its own open-source VS Code fork. There's some chance that making specific workflows for engineering work that don't work at sufficient quality today can create enough distribution. That's interesting.

Then it's not clear: you have the synchronous IDE workflow and the asynchronous one, right? One question is how quickly the quality of these asynchronous code agents increases. OpenAI, with Codex, made a bet on an asynchronous, cloud-based software engineering agent, and then they bought the IDE with Windsurf, right?

It's true, it's true. You can believe both. I think a lot of these things will just consolidate over time. My view is that the market is going to see 2 types of consolidation: product consolidation, and then there will be actual acquisitions. The Codeium/Windsurf acquisition by OpenAI is the first step in that.

If I were number 1 or number 2 in a market and I was a startup, I'd consider merging with the other party if there were 2 main startup players, because the real threat will be fighting the incumbents. I would get ahead of it and say, "Okay, let's stop the startup-to-startup war and just focus on winning against the 3 or 4 incumbents that we have to go up against."

You could just keep fighting and getting distracted by the other party, which is kind of what Uber and Lyft did for a while. There are other precedents. The ones that did merge include PayPal, right? There was X.com, which Musk was running, and then PayPal, which Peter Thiel was running. They decided to merge because they were like, "Why are we competing with each other when there's so much competition?"

I think both paths will happen, but it may be something people should consider as well.

What do you think prevents companies from thinking through that or doing that?

Well, it's 2 things. One is ego. Who's going to run it? They want to subsume me? Sure, I'm number 2, but blah, blah, blah. I'll still beat them. Or what role would I play? Put that aside and just go win. Who cares?

Second, people worry too much about integration. What's the culture, and what's this, and what's that? Often, it's just: merge it, and if it doesn't work, shut down parts of it and move on with life. Whatever parts—either in the buyer or the seller—it doesn't matter. Just merge it. Again, it's a "who cares?" pragmatically. You can fix it all sorts of ways.

Either the cultures mesh or they don't. If they don't mesh, you don't have to keep everybody, honestly, because everybody's going to do very well off the acquisition. You can do all sorts of thank-you packages and move on with life.

Third, sometimes there are dynamics around how you value the things relative to each other for private-to-private companies. Sometimes the easiest way to do that is to choose some metric and say it's divisible by that metric.

For example, years ago when I was at Twitter, I drove an attempt to buy a major social network that was up-and-coming. The way we constructed that offer was that we took their users and our users, added them up, calculated the ratio, and made that offer as a portion of Twitter for the company.

I think you can do that. Take your revenue plus my revenue, add it up, and then what's the ratio? Or maybe it's users. It's whatever the right metric is for your business. I actually think you can do really simple things like that and just say, "Look, fair enough. Plus or minus X% isn't going to matter if we all just win."

So people tend to overthink those things. They overthink role—what am I giving up?—or ego or whatever. Culture—what does the surviving thing look like together? And then, what’s the value, or what’s the relative value, of the 2 pieces? Pragmatically, it’s like: Do you want to fight it out for the next 5 years, or do you want to go win? Then your battleground shifts to the incumbents versus another startup.

Yeah, I’ve seen the simple relative metric also work. I also think that founders, board members, and investors are just unwilling to put something like this inside the Overton window. I think people feel like it is capitulating, but it’s capitulating in service of winning. And so I think that’s a big reason people don’t want to look like they’re unwilling to go to war.

Yeah. The pie basically gets bigger if you do that because you’re focused on just winning the market versus competing with each other, but your pricing dynamics shift as well. You’re not competing on every deal with another startup. A lot of things shift, and so I think there are all sorts of positive characteristics. Again, people will win in these markets without it.

Some markets are really big, and there is room for a number 1 and a number 2 and maybe a number 3, or maybe incumbents. Payments was that way, right? We have Adyen, Stripe, PayPal, and a dozen other payment processors. It’s a very big, fragmented market. Some markets can sustain multiple players, and that’s fine too. I’m just saying sometimes you want to say, “Hey, let’s put aside our differences and go win together.”

Okay. Some part of the market is consolidated. Some could be better consolidated in terms of startups winning. There are areas that you and I have talked about where they feel like obvious commercial opportunities, but people are not chasing them sufficiently, I think. We’ve talked about engineering as one that AI will absolutely change. You have a bunch of ideas in biotech. What’s missing?

Yeah, the biotech stuff I’m interested in honestly isn’t AI-related, although there’s obviously really cool things happening in terms of models. There’s a whole separate thread of stuff I just think is neat. I’m not an active biotech investor; I’m the wrong person to pitch on things, et cetera. I mainly do software, AI, and so on as investments, as well as the companies I’ve started, which have largely been software-driven companies.

I just think there’s some really cool stuff now that the science in biotech—or in basic science—is far enough along, and nobody, or very few people, are working on it. I’ll give you maybe 2 or 3 examples. One is there’s some really good data now for fertility out of Japan where you can basically take a cell and reprogram it to turn into either a sperm or an egg.

They’ve made mice now with 2 fathers, for example. You could differentiate one father’s cells into sperm and one father’s cells into eggs, and then you can have viable offspring. That really opens up the capability for any adult to have kids with any other adult.

So if a woman is over a certain age, she can suddenly produce either sperm or egg. You can do it for different types of couples. There’s stuff like that where you’re like, why are so few people working on this?

An even simpler version is that girls are born with 1 to 2 million oocytes, which are egg cells. By puberty, they end up with about 300,000, and there aren’t good technologies to basically mature those eggs. If you’re a woman, you should be able to mature your oocytes at different points in your life, and you should be able to harvest tons and tons of eggs if you ever want to have lots of kids, right? There’s a lot of stuff like that that just nobody’s doing.

Is the outcome of that that people choose the inputs to having kids differently? For example, the sperm or egg donor market is very different. We’re all just having kids with Elon—you and me both.

The crazy thing about that, honestly, is say that you meet Elon Musk, LeBron James, Taylor Swift, or whoever it is somewhere and you manage to swap some cells off of them—you shake their hand or whatever—you could potentially reproduce them.

Yeah. No, seriously. Some of the ramifications of this stuff are pretty crazy if you think about it, right? But for society, it’s so impactful in terms of what you could do with that. To your point, suddenly anybody could become an egg or sperm donor in any capacity. It just seems like it has such big implications, even if you just say, “We’re going to limit it to women over a certain age,” or people who just aren’t reproductively viable otherwise, right? It’s a pretty big deal in my opinion.

But again, the science is there. They’ve worked through a lot of the pathways to get there, and now it’s like, okay, I know 1 company doing it. But it’s driven by a very good founder; 1 company, that’s it.

Another area would be: You look at Botox. People are injecting a bacterial toxin into their skin to look younger—literally, a toxin—and that was a $40 billion company, with $1.5 billion a year in revenue just for cosmetic applications. Why isn’t anybody doing actual drugs and treatments for aging? There’s all sorts of science around it, all sorts of biology. Nobody’s working on skin aging, balding, gray hair, all that kind of stuff.

Then there’s the stuff that’s really impactful in terms of neurosensory, right? The muscle that holds the lens of your eye gets weaker with time, and so why don’t you rejuvenate that? That’s why everybody ends up with reading glasses in their 40s. Or hearing loss—there are pathways for that. Or tooth regrowth: You have a cavity; why don’t you just grow a new tooth? There are pathways for it.

Again, a lot of the biology is worked out. Maybe there’s more that needs to be done from a basic science perspective. In many cases, for example, for dental stuff, there are genes like USAG-1, which allow for tooth regrowth in certain animal models. So why don’t we do that in people?

What’s your hypothesis for why there are areas that, to me, seem like clear demand if the science you suggest exists? Why isn’t it being funded?

Yeah, it’s massive markets. I think there are 3 reasons. Number 1, the biotech or biopharmaceutical market for founders is very different from the tech market, and the overall market structure is radically different.

If you look at biotech, the last time a $50 billion-plus biotech company was started from scratch, excluding Moderna, which was kind of an accident of COVID, was in the ’80s. I think it was Regeneron. It’s been almost 40 years since we’ve had a de novo, tens-of-billions-of-dollars company created. All these companies are 50 or 100 years old.

Imagine if tech were basically IBM versus HP right now, and you didn’t have any young, founder-driven, aggressive companies. We wouldn’t have the iPhone. We wouldn’t have the internet. We’d just be logging into IBM mainframes off of HP laptops. Do you know what I mean? There’d be no progress, or very little progress.

That’s one issue. The funding models also are ones where a lot of biotech money is either very early-stage or very late-stage, and a lot of the companies are started as incubations by biotech VCs. They load up a company with $40 million, they buy 40% of it upfront, whatever it is, and then they kind of have to make it far enough that they can get almost public-market money, effectively. A lot of the crossover funds then kick in.

The way these funds are set up, because they have so much ownership, they’re really built to flip these companies into the arms of pharma. That means you build against pharma pipelines. If there are 6 or 7 areas that all the pharma companies care about—it’s cancer, cardiovascular disease, and neuroscience—you only build companies in those domains because your goal isn’t to build a big standalone thing. Your goal is to sell it to a pharma company.

A lot of the dynamics are driven by that. And then there’s big regulatory capture that also prevents a lot of innovation. The FDA will ask for—they’ll push hard on—endpoints for certain things that may not exist. There are those 3 main factors that make it kind of hard to do anything else, but all the science is just sitting there, right?

Oh, I guess the last piece, the fourth, is that for some of these things, the scientists who would work on them don’t want to work on something that’s too commercial. It’s kind of the purity of science. It’s low status. How dare you work on fixing wrinkles? As a scientist, you need to be doing something that’s much more pure, et cetera, et cetera. So there’s also a little bit of that—what do you call it?—prudishness around commerciality that exists.

I guess, back to our regular programming, I had a question for you on the AI side. In particular, I know you’ve been thinking a bit about world models and RL, and how these things are overall relevant to capability and the scaling of capabilities. Do you want to explain a little bit about what you mean by world models?

People who are in the AI world get all this stuff, but it would be great for a more general-purpose audience if you could walk through your thinking and what you think is interesting and going on there.

I think it’s very important as an overall area because, if you zoom all the way out, I actually think this is a time of more open research questions than ever. Scaling up model size and training data for big LLMs has given us this really powerful foundation of knowledge and pattern recognition. But everybody talks about agents—what people want to do from here. The way people think about AGI is not just predicting text, right? They want to move toward broader intelligence and taking actions.

I think it’s really important to describe what we mean when we say “reasoning” or “actions” more concretely, because I don’t know that everybody has a great mental model for these things. It could be planning, reading documents and drawing conclusions, using tools, receiving feedback, going down different reasoning paths, or evaluating your own work. It’s taking a series of actions in pursuit of a goal beyond just sequential text generation.

My understanding is that the labs—some labs more than others—have spent a lot of money collecting traces of humans doing sophisticated tasks. This is how Elad looks at Japanese stem cell differentiation research, right? He does these tasks, calls these people, and then they try to do behavior cloning: monkey see, monkey do, but for software engineering or investment research or whatever.

But it tends to be really brittle when you go off the path with the cloning techniques. The model all of a sudden presses some button that the human never touched, lands in out-of-distribution territory, and then has no idea what’s next. It fails; you get stuck.

Then people are trying a new generation of reinforcement learning, which is broadly trial-and-error training. I think a lot of people who are paying attention to AI have seen agents play games, famously chess and Go, or more complicated games with human interaction. You’re taking actions in an environment and getting feedback in the form of a reward or a penalty, and then you play until you’re better at the game.

For games, that’s easy because you have clear rules, so you know very easily how to either reward or penalize an action. That’s very different from real-world tasks in some cases, and this is exactly the problem with using RL more broadly: What is the task if it’s not just winning in chess or Go? How do you make the environment? You’re trying to make a copy of the universe, or at least some little piece of it, that’s rich enough to teach useful problem-solving but cheap enough to run.

I don’t know how we get to the Matrix. It’s very hard to design rewards, and then you have a gap from reality. You also need diversity, or you’re just memorizing a path through your game—even if that game is the game of an agent doing research work or the game of a software engineering project—and you’re overfitting instead of adapting.

I don’t actually have a ton of conclusions here, but I’ve spent a little bit of time trying to understand it. There’s an interesting set of researchers now who are working on creating more universal environments and world models, or just trying to get better trace data. I actually don’t know that I believe any of the more immediate-term commercial applications of these models are interesting. People say, “Oh, we can generate games, or we’ll have gaming assets, or we’ll use the data for robotics training,” or some other thing. But I do think it’s a really interesting conceptual path toward more AGI.

One thing I think is intriguing in what you said—and it’s one of the points that I’ll overextrapolate—is that, if you look at the way AI has done certain things, for example in Go, because there is a utility function but no other constraints, it came up with all sorts of crazy moves that a human wouldn’t have come up with, or at least hadn’t come up with to date. Then humans started studying and copying these moves. They were completely out of the box, but they ended up with a superior outcome.

I always wonder what that looks like for other areas of human endeavor. If coding shifted from, “Hey, let’s copy how people write code,” to, “Let’s just solve this problem,” how different is the type of code that’s written? What sort of traditional approaches are just broken that we can then learn from, because you’ve created a utility function with an unconstrained approach to actually figuring it out?

That happens sometimes in biology, right? You’ll do these molecular evolution experiments where you’ll evolve a molecule to do something, and sometimes it’ll do things in a really weird way that you just completely don’t expect. Suddenly you have this catalyst that works in a really weird way, or a binding protein that doesn’t do it the way you’d expect at all. It’s because it’s not designed; it’s evolved. I think this whole notion of evolved systems or self-selecting systems can yield really weird insights, and I’m really excited to see that kind of stuff in terms of the outcomes.

Me too. One way I visualize this is that a model is looking in a part of the search space that humans have not traditionally been taught by the Go rulebook, the prior games, or whatever. It could be in the shape of a protein or any other problem.

Have you seen the TV show Pantheon?

No. What is that?

It’s a TV show about AI and mind uploading. It’s a kind of niche animated TV show. You should watch it. Everybody should watch it.

I think it’s really interesting because the uploaded beings at some point become your full self—or at least, for us, it would be humans learning to think differently. It’s breaking through your constraint of how you might traditionally solve the problem or see yourself. I do think that, thematically, it’s one of the more inspiring things about AI.

Oh, that’s interesting. I feel like there are a lot of sci-fi books where eventually you have your brain uploaded into the cloud or whatever, and then there are all sorts of controls you suddenly have access to that you didn’t have before. For example, you should be able to fine-tune your emotions or your emotional state and dial it up and down literally with dials.

I think there are always these really interesting meta-questions. If a human upload were to occur, what does a transhuman species look like? What are the capability sets that aren’t a priori obvious that you suddenly expose?

I mean, obviously, you could also spawn instances of yourself and have those things go do things for you and then merge back in. Maybe some of them don’t want to merge back in, and then who’s the real identity? You know, all that stuff. It’s kind of fun.

I’m told that the modulation of emotions and attention actually doesn’t require upload. Fred and some professors we know would say it’s just ultrasound devices coming soon to a consumer shelf near you. But we can talk about that on next year’s episode.

Yeah, sounds good.