No Priors Ep. 116 | With Sarah and Elad
Summary
Elad Gil sees a subset of AI markets consolidating, with apparent winners for the next two years—even if today’s leaders may not be five-year winners. Foundation LLMs, medical scribing, coding, and customer success have recognizable frontrunners; sales productivity, financial-analyst tooling, and accounting remain open, and Sarah Guo adds pharma. Gil wonders whether the unresolved markets lack the right product approach or still need better models. After years when “the more I learn, the less I know,” he finally sees “a nice breather in terms of uncertainty.”
The emerging application playbook combines a valuable vertical workflow, proprietary data, distribution, and users who can create, derive, or extend knowledge from that data. Guo places Abridge and OpenEvidence in that category while remaining unsure how sales will be won in a historically fragmented market. Coding suggests the ambiguity is narrowing: plausible entry points now include Cursor, Codeium/Windsurf, Cognition, and Microsoft Copilot.
AI startup consolidation may become strategic defense against incumbents, not an admission of defeat. Gil advises the top two startups in a category to consider merging to stop the startup-to-startup war and shift the fight toward three or four incumbents. Guo says founders, boards, and investors often treat even considering a merger as capitulation, though she calls it “capitulating in service of winning.”
Gil argues that several enormous biotech markets are commercially obvious but structurally neglected. Examples include deriving sperm or eggs from reprogrammed adult cells, maturing a woman’s remaining oocytes, treating skin aging and hair loss, restoring near vision or hearing, and regrowing teeth through genes such as USAG-1. The startling implication of cell-derived reproduction is that “any adult” could potentially have children with another—and cells obtained from someone by swapping some off them in a handshake could potentially enable reproduction involving that person.
Biotech’s financing architecture selects for assets that pharma will buy, rather than enduring new companies serving unconventional demand. Gil says incubators may invest $40 million for roughly 40% ownership, then steer programs toward cancer, cardiovascular disease, or neuroscience pipelines sought by acquirers. Regulatory capture and scientists’ view that fixing wrinkles is “low status” further strand usable science outside the venture funnel.
Gil presents world models and reinforcement learning as a possible route beyond text prediction, not primarily as near-term gaming products. Agents need planning, tool use, feedback, alternative reasoning paths, and self-evaluation, while behavior cloning is brittle outside demonstrated paths. The challenge is building “a copy of the universe”—an environment rich enough to teach adaptable problem-solving, yet cheap and diverse enough to avoid overfitting. Gil is unsure about immediate commercial applications; Guo’s extrapolation from Go and molecular evolution is that unconstrained search could produce unusual but superior code, molecules, or strategies.
Deep dive
1. AI’s temporary winners are becoming visible
Gil’s changed view is the episode’s anchor: AI was the rare market where “the more I learn, the less I know,” but recent months have clarified several categories despite continuing research progress.
In foundation LLMs, healthcare applications such as medical scribing, coding, and customer success, he now sees likely leaders for the next two or three years—not necessarily five. He names Sierra and Decagon in customer success and Cursor, Codeium, Cognition, and Microsoft Copilot around coding.
Guo calls this understanding the market’s “temporary physics”: find a relevant vertical, make a workflow users want, add proprietary retrievable data and distribution, and give users a way to create, derive, or extend knowledge from it. Abridge and OpenEvidence fit that shape; sales remains unclear.
2. Coding previews both product and corporate consolidation
Coding once supported perhaps a dozen plausible ways to win. Its entry points now look much narrower, though Guo points to open and smaller models that can do real things with code, citing Codestral, as possible enablers of specialized engineering workflows that do not yet work at sufficient quality.
Guo also points to Microsoft’s open-sourcing of Copilot and frames it as an effort to stop Cursor from eating its lunch with its own open-source VS Code fork; she says the impact remains to be seen.
The remaining divide is synchronous IDE work versus asynchronous cloud agents. Guo notes that OpenAI bet on async with Codex while buying the IDE with Windsurf: “You can believe both.”
Gil expects product convergence plus acquisitions. His advice to two leading startups is blunt: “Stop the startup-to-startup war” and combine against incumbents, as X.com and PayPal did, rather than repeat years of Uber-versus-Lyft distraction.
Ego, integration anxiety, and private-company valuation block such deals. Gil suggests choosing one metric—users or revenue—splitting ownership by the resulting ratio, and accepting that “plus or minus X% isn’t going to matter if we all just win.”
Gil does not think every market requires a merger. Some are large enough for several players; he cites payments, where Adyen, Stripe, PayPal, and a dozen other processors coexist in a fragmented market.
3. Reproductive and longevity biology hide obvious markets
Gil’s first neglected opportunity is fertility: he says there is good data from Japan showing that cells can be reprogrammed into sperm or eggs, and that viable mice have been produced with two fathers. The eventual promise, as he frames it, is for “any adult to have kids with any other adult.”
A simpler version would mature existing oocytes. Girls begin with 1–2 million, retain roughly 300,000 by puberty, and lack good technology to mature and harvest many eggs across life.
The ethical risk is inseparable from the opportunity: cells obtained from someone by swapping some off them—through a handshake or other contact—could potentially enable reproduction involving that person, such as Elon Musk, LeBron James, or Taylor Swift. “Some of the ramifications of this stuff are pretty crazy.”
Aging offers another demand signal: people inject a bacterial toxin through Botox to look younger. Gil says it represents a $40 billion company with $1.5 billion a year in revenue from cosmetic applications. Yet he says nobody is working on actual treatments for skin aging, balding, gray hair, weakened near vision, hearing loss, or tooth regrowth through genes such as USAG-1, which he says work in certain animal models.
4. Biotech’s structure filters out unconventional ambition
Gil’s historical comparison is stark: excluding Moderna’s COVID-driven rise, he thinks the last de novo biotech company to reach $50 billion or more was probably Regeneron in the 1980s. Without young founder-driven challengers, biotech resembles a technology industry limited to IBM and HP.
Venture formation reinforces that stagnation. A biotech fund may seed an incubation with $40 million, take 40%, and take it far enough to attract almost-public-market money; crossover funds then kick in. That ownership model is “really built to flip these companies into the arms of pharma.”
Consequently, startups follow pharma acquisition pipelines—cancer, cardiovascular disease, and neuroscience—instead of building standalone markets. Gil adds regulatory capture and FDA pressure for endpoints that may not exist, while acknowledging that some areas may still need additional basic science.
His fourth constraint is status: scientists can regard highly commercial work as impure. “How dare you work on fixing wrinkles” captures the prudishness that leaves valuable biology sitting unused.
5. World models could teach systems to leave the human path
Gil frames world models against the progress of large LLMs: scaling model size and training data has produced a powerful foundation of knowledge and pattern recognition, but broader intelligence requires more than sequential text generation. Agents must plan, read documents and draw conclusions, use tools, receive feedback, explore alternative reasoning paths, and evaluate their own work in pursuit of a goal.
Human task traces enable behavior cloning—“monkey see, monkey do”—but remain brittle. Once an agent presses a button the demonstrator never touched, it enters out-of-distribution territory and may have no idea what comes next.
Reinforcement learning supplies trial and error, but real work lacks the clean rules and rewards of chess or Go. A useful environment must approximate some piece of the universe, close the reality gap, and generate enough diversity to teach adaptation instead of one memorized route. Gil says he does not know how to get to “the Matrix.”
Gil is skeptical of many immediate commercial pitches—generated games, gaming assets, or robotics training data—and says he does not have many conclusions about the near term. He nevertheless sees world models as “a conceptual path toward more AGI.”
Guo then extrapolates from Go: with a utility function and few other constraints, AI found moves humans had not devised, which humans subsequently studied and copied. She wonders whether coding and molecular evolution could similarly produce unexpected solutions—strange catalysts, binding proteins, or code—that humans can learn from. Gil agrees, visualizing models searching regions of a problem space that humans were never taught to explore.