Doctors Are Rushing to AI. Robotics Will Take a Beating First.

Sarah Guo opens the No Priors year-end episode by trying to wrap a year in which AI broke into the mainstream and took center stage with policymakers. ChatGPT shipped huge user numbers and asked for huge amounts of capital. Google roared back with Gemini. On the application side, coding shifted from autocomplete to agents and started eating inference capacity, while doctors adopted clinical decision support at scale and enterprise use in law and customer support accelerated.

The research race now has multiple live players, and open source is closing the gap. A handful of new research labs got funded this year, and Ilya Sutskever is calling it the age of research. People are trying diffusion, self-improvement, data efficiency, emotional intelligence, large-scale agent collaboration, and continual learning. After a year of attempts to push AI into the physical world, robotics has a fresh wave of optimism, and Guo says those companies will start making contact with reality next year.

Her personal prediction is specific: somebody will make hundreds of millions of dollars trading markets with LLMs in 2026. “It’s inevitable,” she says. We are in the second or third inning. Markets are running a little hot and a little volatile. “It’s hot in the hot tub.”

Elad Gil’s first 2026 prediction has nothing to do with a new model. People will proclaim yet again that AI is not doing much and is overhyped. He points to the MIT report that people kept quoting, which he thought “really didn’t matter.” Technology waves take about ten years to propagate. People are already getting enormous value out of AI, he says, and they will get far more of it.

Those overstated bubble claims will sit next to “AI isn’t working that well” takes, as they do in every technology cycle. Gil calls the resulting punditry a waste of time. Guo’s read of the investing landscape is that a lot of people are stressed about how much capital they have at work, plus uncertainty about adoption cycles and technical bets they do not have first-principles confidence in. She still expects a “negative market current” whenever NVIDIA has a quarter that does not massively overperform—panic that, she argues, has very little to do with the fundamental secular change.

Gil’s second prediction is that the next set of verticals will hit massive scale. This year coding consolidated into a handful of players, as did medical scribing and legal tools such as Harvey. Next year he expects another wave of consolidated verticals to follow the same pattern.

Guo cites a report from the physician community Offcall. Doctors are using AI for documentation, clinical decision support with companies such as Abridge and Open Evidence, and general models, with “massive enthusiasm” across most of the profession. Of all the domains considered conservative, the desire for tools that make work better is obvious, and she expects it to continue in other professions.

Gil says this is “super under-discussed”: the people who tended to be the slowest adopters of technology love AI. Physicians, lawyers, certain accounting types, compliance—the people who never adopted technology are now adopting this fast. Guo’s explanation is simple. Lots of professions live on unstructured data, and models that can reason over it are unusually useful. That is also why a noisy public-markets quarter should not be confused with a stall in the real world.

Gil’s third bet is not the next LLM, but models that make progress in physics, materials, science, and math. There will be one or two cases that work unusually well—a new material, a proved conjecture—and then an overstated hype cycle that science is solved. The one-off will be overstated. In the long run the trend will be understated, and it will be incredibly important.

Guo’s robotics prediction is less polite. There will be a collapse of sentiment around a set of robotics companies next year, not because the field stops progressing, but because people are projecting timelines that not everyone will hit. Humanoid and semi-humanoid robots will be deployed at small scale in consumer or industrial settings. Not everything will work, and because the humanoid hype cycle is already maxed out, any miss will make people freak out and investment will bifurcate.

Gil notes that self-driving took fifteen to seventeen years and is only now really working. Robotics may have a steeper curve but a similar shape: it takes time, then it is very valuable. Self-driving had two or three dozen legitimate companies; the biggest winners so far are Waymo and Tesla, both incumbents. Chinese car companies are also winners, though banned from the US market. The most likely global robotics winners, he thinks, are some subset of China plus Tesla plus maybe one startup, because capital, hardware, and manufacturing needs favor incumbents.

Guo is not convinced an incumbent’s model advantage transfers. Put Tesla’s self-driving model in Optimus and it can do locomotion, but not much else; manipulation and hardware still have to be built. Startups still have a shot; the scariest competition is Chinese companies. Gil agrees startups can succeed, but thinks one or two incumbent slots will default unless something strange happens—the way Google was always going to hold a foundation-model seat given transformers, data, capital, and TPUs. He also predicts that one of the biggest trends of 2026 will be self-driving actually beginning to matter, in your own car and in Waymo and Tesla cabs.

Gil’s fifth theme is IPOs and M&A. Guo says it depends on whether the bottom falls out of the AI market. What people are scared of, she lists, is that demand is not real enough to support the capex cycle; systemic risk from “passing the ball around” on who is actually responsible for the buildout, credit agreements, and pay-on-delivery contracts for data centers and chips; and too much concentration in NVIDIA and a handful of other players.

She relays a conversation with a friend who runs a large tech hedge fund, already an investor in multiple significant labs that may go public. The question is whether to buy the IPO. His game theory: no matter what he thinks, he has to, because retail will want to be part of the AI revolution, hedge funds are benchmarked on annual performance, and nobody wants to miss it the way they missed NVIDIA. Guo’s reaction: “this is not the investing job I know how to do.”

Gil expects many more IPOs next year. If one of the main AI companies goes out, it will probably do extremely well depending on pricing—too aggressive and it will not. There is so much retail appetite to participate in AI besides NVIDIA that one listing will pull a line of followers. For the labs, a public offering is also a way to raise huge amounts of money.

On consumer products, Guo admits she did not believe there would be many unique experiences besides ChatGPT; she now expects a slate of hardware that mostly fails, but she is seeing “magical” consumer agent software she actually wants to use. Gil’s structural worry is that unless a startup hits escape velocity and builds a network effect, incumbents can launch two or three years later and catch up. Two or three years ago, David Song on his team ran a Stanford program giving engineering students free GPUs to build consumer AI apps—the prototypes were neat, and then nothing happened for a couple of years. Guo lists the reasons: incumbents are scary, last-generation founders watched their ideas get ingested into platforms, and many founders rebuild last-generation experiences with this generation’s technology. A well-known consumer founder running a giant public company told Gil there are perhaps a few hundred great consumer product people actually working on it, and Guo would add that they also need to know what is technically possible now, not grind on the fiftieth iteration of a Gmail button.

A surprising 2025 development was the funding of new research labs roughly ranked three through eight. Gil thinks there is still a lot of juice in these models, but scale matters, which means capital aggregates to what is working and the field eventually collapses into a handful of players. Guo paraphrases Ilya Sutskever’s recent interview: he believes in scaling, but there is a floor of compute—not infinite—where you can test ideas at scale. If you have secret ideas for faster or more compute-efficient improvement, it is not just a straight resource battle.

She splits the new-lab argument three ways. Multiple architectures—diffusion, state-space models—may be relevant at large domains of usefulness; they just have not been scaled, and there is enough capital to test them next year. Meanwhile, the top labs have enormous compute but spend a lot of it on inference, which is what generates the revenue to buy more compute. That leaves less for self-improvement, post-training, emotional intelligence, or large-scale agent work, even as inference is the bootstrap into more scale.

Gil’s most speculative thread is evolutionary systems. A brain is a set of specialized modules: vision pre-wired, memory, empathy, something like a mixture of experts. Protein design is the analogy. Analytical design underperformed; phage display and mutagenic scans did better; then AlphaFold mostly solved structure prediction. Maybe AI development ends up there too: spawn instances, apply a utility function, select and recombine.

That is one reason people are so focused on code, he argues. Code is a bootstrap into moving faster on AGI. Code plus self-evolution may be the interesting path to a fast liftoff. Or maybe not. They will see.

Gil’s adjacent prediction is that defense tech will accelerate: drones and autonomous systems are reworking how people think about war, the Trump administration is pushing it, and there is now enough density of startups doing interesting things. He thinks the shift is under-thought because it will be so big, and he also flags SpaceX and Starlink in communications, plus energy and mining. Guo agrees, with a caveat: budget has to actually shift from traditional primes to new companies at scale. She defends boom cycles anyway, because they bring capital, founders, and talent to the table, and you make a lot of progress even if many companies die.

His other non-AI bet is GLP-1 drugs. Despite the enthusiasm, he thinks the impact is still underrated. Adoption is inexorable, and it creates a path for other peptide and hormone therapies. Delivery matters, and that realization will fuel more investment. Guo treats the biohacking community as an early signal: people flying to Dubai for peptide injections, off-label GLP-1, ultrasound neuromodulation, stem cells. Ten years ago creatine and magnesium came out of bodybuilding and became ordinary. Peptides, she thinks, will follow the same path from fringe to mainstream.

They close with a reel of guest predictions, starting with Jensen Huang, who says reasoning systems will make AI much more versatile and will revolutionize industries from biology to self-driving and robotics. Arvind Jain of Glean expects AI to stop waiting for a prompt and become a proactive work companion. Winston Weinberg of Harvey says context and memory will be the most important part of every product, shifting the onus off the user. Zach Ziegler of Open Evidence thinks a mass-scale consumer agentic product still does not exist, the models are ready, and the right interface is “anyone’s ball game.” Aaron Levie of Box calls 2026 year two of enterprise agents in deep verticals: workflow, data access, context engineering, change management, agent harnesses, and economically useful evals such as GDPval.

Misha Laskin of Reflection AI predicts the United States will regain the open-weight frontier after DeepSeek shifted it to China in late 2024. Joshua Meier of Chai Discovery says 2025 already produced full-length antibodies with drug-like properties designed zero-shot on a computer, so 2026 should be the year those models get deployed against targets traditional methods cannot touch. Bryan Johnson wants a cultural shift from “you only live once” to “don’t die.” Sholto Douglas of Anthropic expects other knowledge work to live through what software engineers felt this year, while Dylan Patel thinks the third of Americans who hate AI—and the majority who think it is useless—will have those priors flipped by obvious utility, and Noam Brown expects AI to become a fight in the 2026 midterms. Ben and Asher Spector call it the year of energy-efficient AI: data centers are stuck on power and interconnects, xAI’s Colossus started on on-site gas turbines, so intelligence per watt matters now, even if chips still cost an order of magnitude more than electricity over a five-year depreciation at about 10 cents per kilowatt-hour.