The AI Table Is Being Cleared: Two or Three Coding Winners, and Forty Years Without a Biotech Giant

Sarah Guo opens by asking whether the AI market is finally crystallizing into known businesses or still as fluid as it has felt. Elad Gil's answer starts with a confession he has used for years: AI is the one market in his career where the more he learns, the less he knows. Every other market rewards study with more certainty. That, he says, has shifted in the last couple of months. New models and papers are still arriving at a frantic pace, but a subset of markets has consolidated. You can now write down who is likely to matter for the next two or three years.

He walks the board. In foundation models, at least for LLMs, it is clearer what is important and what is not. In healthcare applications, the first wave of services such as medical scribing already has visible winners. Coding has collapsed into two or three players: Cursor, Codeium, Cognition, and then Microsoft's Copilot. Customer success is concentrating around Sierra and Decagon. Those names may not be the winners five years from now. A new startup could still displace everyone, or an incumbent could make a bold move. For the next two or three years, the table is more legible than it has been.

The open markets are just as specific. Something important will exist in sales productivity, in tools for financial analysts, and in accounting. The question is why those have not consolidated yet: nobody has found the right product approach, or the models are not good enough and capabilities still have to improve. Either way, Gil says, it is far clearer than a year ago. When he first invested in generative AI, there was so little competition that you backed strong people in interesting markets. That is when he led Perplexity's seed, and invested in Character and Harvey, before ChatGPT and Midjourney. GPT-3 was out and people called it kind of crappy. The scaling law was already clear. A handful of people, Guo included, saw that it would matter. Then came two or three years of so much innovation you could not see the shape. Now, at least in a subset of things, there is a breather.

Guo says she finally understands some temporary physics of the market. It is a race to find the verticals that matter, then make something users actually want. That usually means proprietary data you can retrieve against, a path to distribution, and users who create or extend knowledge from what they see. The healthcare companies Gil named fit that pattern. She adds two he did not say out loud: Abridge and OpenEvidence. Both, she thinks, have that shape.

Sales is the one market they cannot agree on. Guo does not know how something wins there. You could attack it from data or from adoption, and so far it has stayed fragmented. She agrees with him on finance and accounting, and she would add pharma: document-driven industries where you can already see something coming. Coding is the analog she likes for how confusion resolves. Two years ago there were probably four different approaches running at once, and a dozen imagined entry points. Now the entry points for how you actually win look much clearer. Sales may feel uncertain today and obvious in two years, the way coding does now.

The remaining variables are still real. Open models and small models can already do real work with code; she mentions Codestral. Microsoft open-sourced Copilot, which she reads as a belated decision to fight Cursor from eating its lunch with a fork of Microsoft's own VS Code. Sync IDE workflows and async cloud agents are still different products. OpenAI bet on the async side with Codex, a cloud-based software engineering agent, and then bought the IDE in Windsurf. Guo's comment is the one that sticks: you can believe both. Gil expects two kinds of consolidation from here, product consolidation and actual acquisitions, and he treats the Codeium/Windsurf deal as the first step.

Gil's advice to founders is the sharpest stretch of the episode. If you are number one or number two in a market and you are a startup, consider merging with the other startup. The real fight will be against three or four incumbents. Stop the startup-to-startup war and go win that one. The cautionary tale is Uber and Lyft, who burned capital on each other for years. The positive tale is PayPal: Elon Musk's X.com and Peter Thiel's PayPal were killing each other until they decided there were too many banks and card networks outside to waste the fight inside. Both paths will happen. He thinks more people should at least put merger on the table.

Three things, he says, stop them. Ego: who runs it, I am number two but I will still beat them, what role would I play. Put it aside and go win. Integration anxiety: culture, who stays, what the combined company looks like. Merge anyway. If it does not mesh, shut down parts of it. Everyone will do well financially off the deal; you can write thank-you packages and move on. Valuation between two private companies is the third trap, and his method is almost rude in its simplicity. Pick a metric and divide by it. Years ago at Twitter he drove an attempt to buy a major up-and-coming social network. They added the other company's users to Twitter's, took the ratio, and offered that slice of Twitter. Revenue plus revenue works. Users work. Use whatever metric is core to the business. Plus or minus a few percent will not matter if you all win.

Guo has seen the simple relative-metric trick work. She also thinks the deeper block is social. Founders, board members, and investors will not put a merger inside the Overton window because it looks like capitulating. It is capitulating, she says, in service of winning. Gil adds that the pie gets bigger and the pricing changes: you stop competing with the other startup on every deal. He leaves himself an out. Some markets are big enough for a number one, a number two, a number three, and incumbents. Payments still has Adyen, Stripe, PayPal, and a dozen other processors. That is fine. Sometimes you still want to say: put the differences aside and go win together.

They leave the consolidated AI markets for a different kind of missed opportunity. Gil is careful up front: he is a software and AI investor, the wrong person to pitch on biotech, and the companies he has started have been software-driven. He is not talking about AI models in biology, though those exist. He is talking about basic science that is far enough along that somebody should have a product, and almost nobody does. He offers two or three examples, starting with fertility research out of Japan. You can reprogram a cell into sperm or egg. Researchers have already made mice with two fathers, differentiating one father's cells into sperm and the other's into eggs, with viable offspring. That opens the possibility that any adult can have children with any other adult. A woman over a certain age could produce either sperm or egg.

A simpler version is already sitting in human biology. Girls are born with one to two million oocytes. By puberty they have about 300,000, and the technologies to mature those eggs are weak. A woman should be able to mature oocytes at different points in her life and harvest large numbers if she wants many children. Almost nobody is building that. Guo jokes about the donor market collapsing into everyone having kids with Elon. Gil answers straight: if you meet Elon Musk, LeBron James, or Taylor Swift and manage to swap a few cells, a handshake, you could potentially reproduce them. Anybody could become an egg or sperm donor. Even if you limited the technology to women over a certain age, or to people who are not otherwise reproductively viable, the social implications are enormous. He knows one company doing it, with a very good founder. One company. That is it.

Aging is the second gap. People inject a bacterial toxin into their skin to look younger. That product built a $40 billion company, with about $1.5 billion a year just in cosmetic applications. Why is nobody making actual drugs for aging: skin aging, balding, gray hair? The biology is there. Neurosensory decline is the third. The muscle that holds the lens of the eye weakens with time, which is why everyone ends up with reading glasses in their forties; nobody is rejuvenating it. Hearing loss has pathways. Tooth regrowth has pathways. If you have a cavity, why not grow a new tooth? For dental work, genes such as USAG-1 already allow tooth growth in animal models. In many of these cases more basic science may still be needed. In many, he says, it is already worked out. Why not do it in people?

Guo asks the obvious question. If demand is obvious and the science exists, why isn't it being funded? These are massive markets. Gil gives three reasons, then a fourth. First, the market structure is nothing like tech. Exclude Moderna, which he treats as an accident of Covid, and the last $50 billion-plus biotech started from scratch was Regeneron in the 1980s. It has been almost forty years since a de novo company of that scale. The incumbents are fifty or a hundred years old. Imagine, he says, if tech were still IBM versus HP: no young founder-driven companies, no iPhone, no internet, just logging into IBM mainframes from HP laptops. Progress would be a trickle.

Second, the money is shaped wrong. Biotech capital is very early or very late. A lot of companies start as incubations by biotech VCs, who load a vehicle with $40 million, buy 40 percent up front, and then have to get it far enough for crossover funds, almost public-market money, to kick in. Those funds own so much that they are built to flip the company into the arms of pharma. You therefore build against pharma pipelines. If the big bio-pharma buyers care about six or seven areas, cancer, cardiovascular disease, neuroscience, you only start companies in those domains. The goal was never a standalone giant. The goal was a sale.

Third is regulatory capture. The FDA, he says, will push hard on endpoints for certain things that may not exist, and that blocks a lot of innovation. Fourth is status among scientists. Work that is too commercial is low status. How dare you, as a scientist, work on fixing wrinkles. You are supposed to stay pure. He calls it a prudishness around commerciality. The science is sitting there anyway. Then Guo pulls them back to regular programming: world models, reinforcement learning, and how capabilities actually scale from here.

Guo thinks this is a time of more open research questions than ever. Scaling model size and training data for large language models gave a powerful foundation of knowledge and pattern recognition. What people mean by AGI, she says, is not predicting the next token. They want actions in pursuit of a goal: planning, reading documents and drawing conclusions, using tools, taking feedback, going down different reasoning paths, evaluating their own work. Labs have spent a lot of money collecting traces of humans doing sophisticated tasks, then trying to behavior-clone them. Monkey see, monkey agent do, whether the task is software engineering or investment research. The clone is brittle. The moment the model presses a button no human in the traces ever touched, it lands out of distribution and has no idea what is next. It is the difference between memorizing a navigation route and knowing how to find your way when a street is closed.

The new generation of reinforcement learning is trial and error: take actions in an environment, get a reward or a penalty, play until you are better. Games make that easy because the rules are clear, which is why chess and Go agents worked. Real-world tasks break on two sides at once. What is the task, if it is not winning a game? And how do you make the environment? You are trying to copy a piece of the universe rich enough to teach useful problem-solving and cheap enough to run. She does not know how you get to the Matrix. Rewards are hard to design, the gap from reality is real, and without diversity the model memorizes a path through the game of Elad doing research, or through a software project, and overfits instead of adapting. Researchers are now building more universal environments, world models, or better trace data. She does not find the near-term commercial pitches, generated game assets, robotics training data, very interesting. She does find it interesting as a conceptual path to more AGI.

Gil over-extrapolates on purpose. AlphaGo had a utility function and no other constraints, so it found crazy moves humans had not played. Humans then studied and copied those moves and got a superior outcome. What happens if coding stops copying how people write code and just solves the problem? How different is the code, and which traditional approaches turn out to be broken? Biology already has the analog. In molecular evolution experiments, a catalyst or a binding protein often works in a weird way you would never have designed, because it was evolved, not designed. Self-selecting systems yield strange insights. Guo visualizes it as a model looking in a part of the search space the Go rulebook never taught. She recommends the animation *Pantheon*, about AI and mind uploading: uploaded beings who learn to think past their old constraints, spawn instances, merge back, or refuse to merge. Gil adds the sci-fi control panel, emotion dials, and the identity problem of copies that will not come home. Guo's kicker is that modulating emotion and attention may not need an upload. Ultrasound devices, she is told, are coming to a consumer shelf. They can talk about that on next year's episode.