The Week AI Won Its First Nobels, Two Investors Listed Who Is Most at Risk

Elad Gil welcomes Sarah Guo as the guest on her own show, joking that No Priors is the number-one podcast in her family—the one her mother watches regularly. He is less sure about his own mother, but says hi just in case. The week’s object of fascination is Google’s NotebookLM. About a year earlier, Guo and a teammate named David ran a Stanford cohort: they sponsored compute, met students weekly, took no equity, and asked people to build consumer AI apps. One prototype was a journal whose sidebar interpreted each entry, in the context of all prior ones, through a lens you chose—counselor, friend, and so on. It felt, she says, like something special.

NotebookLM gives her a similar jolt. Upload documents and it can turn them into two AI voices discussing the material as if it were a podcast, a multimodal trick for information discovery that has been ringing around the internet. That audio overview is not how Gil uses it. NotebookLM is his favorite new AI product of late, but he treats it as a pre-built RAG GUI attached to his own data: upload, get an index, interrogate it. Guo has seen earlier versions—two companies in recent YC batches, plus Glean and Danswer in pieces—but not as elegant as this. Any piece of content, they agree, can suddenly become something you interact with.

The deeper point is cold start. If you insert the model into an existing workflow, use existing artifacts, and connect existing data, you should not rely on end users to be creative about how to engage. “Great job, Google,” Guo says: enable more of these interfaces and engagement takes a step-function jump. The consumer apps that have actually landed were mostly unintentional or prosumer. ChatGPT was an experiment that became a major consumer product; Perplexity sits between research and search; Midjourney started consumer and now sells a lot of professional pitch-deck graphics. Last wave: commerce, search, social. This wave: information utilities and content generation. Asked about entertainment, games, and social, Guo is excited about a lot of it and has not seen anything fully there yet.

The concept is old. When GPT-2 and GPT-3 landed, people already asked what it would mean if every NPC could talk like a person—if you formed a guild and could not tell who was real. Gil pushes it to the end: for a subset of people, do most social interactions become interactions with machines? Guo has now seen two demos, one public, that she thought were actually good: coaching NPCs, and, more important, the ability to trigger agentic actions in context at latency that makes sense in a game. That, she says, is finally coming to high-scale titles.

The Stanford group also had a team on generative levels: same mechanics, infinitely many stages. She points at Angry Birds. From one to seventeen the core loop barely moved; the levels and the art did. Gil interrupts that Angry Birds Star Wars was his favorite, and she needles him for having just said the games were the same. The logic still holds: describe a layout, generate it, keep the mechanic, change the setup. Developers may end up doing mechanic design and letting the model scale the rest. Modding is already a huge existing behavior—levels, expansions, visual changes, whole platforms for it. Giving users AI-based expansions to worlds they love, even beyond the publisher, she thinks will be a big deal.

Then there are the products that return every generation. GeoCities and personal sites, Tumblr, LiveJournal: a journaling and self-expression line that goes trendy for kids every decade or so. Guo does not know a modern equivalent right now. She thinks AI will modernize that class of thing, and she still counts herself among the kids who will be part of it. Gil says he thinks of her that way. GeoCities V8, she deadpans.

Gil’s test of whether you are a “real AI kid” is a dinner he hosted. A friend, a researcher at one of the labs, had just had a child. Casually, the researcher said his son is probably going to have an AI partner instead of a human romantic one. How did he and his wife feel about no grandkids? If the child gets support, love, and emotional fulfillment from that, the researcher asked, who is he to say it is not right? Guo’s jab: give him ten years and he will want grandchildren anyway. Gil says the man had seemed middle-of-the-road except for where he works, and that this is the rare topic on which Gil feels very conservative. A well-known company CEO, Guo adds, uses a dinner-table question: what proportion of your kids will at some point date an AI? She thinks that will definitely happen. Dating an AI and reproducing, she says, are separate.

Artificial wombs are where both of them tap the brakes. Science can already grow a lamb in a nutrient-fed membrane for some number of weeks of gestation. Ten years ago Guo would have called that a good thing. Now she worries about it going mainstream. Helping people who cannot reproduce, or a case like placenta accreta, she gets; everybody having kids that way, she does not. The last twenty years, as she reads them, have been dissociation of people from society and from one another, and she thinks that connection is valuable even if subsets of the AI community do not. Gil reaches for Star Wars again: the Clone Wars is mass-cloning one person into the Empire’s rank and file. If it is mainstream, he worries about what it means societally. He may be a little Luddite on the topic, he says, and might think the opposite in five years.

The implication she is describing—less attachment to other humans because you no longer need them to get fed socially—would, Gil thinks, make people less concerned about society. Covid is his example: lockdowns and isolation made a lot of people a little crazy. This would be a different form of isolation. You still get social interaction, with bots; what you lose is human contact. Even in the very near term, no artificial womb required, Guo expects some human–AI dating and emotional dependence. The premium on human-like communication will decrease just by volume. She is trying to cure herself of answering email from strangers because a real person is on the other end, trying to be kind. That obligation, she thinks, will fade when the inbox is infinite. “It’s going to be bots emailing bots.”

OpenAI’s o1 is the other new model. It plans and thinks longer, scaling compute at test time, which helps on tasks that need iterative reasoning—math, code, and, a researcher favorite, crosswords, where clues interlock and next-token recall is the wrong shape. That opens a category of use cases less well served today. Industry reaction has been mixed. Every warning from OpenAI not to use it exactly as you use GPT-4o is being ignored, Guo says, and when people test it, it is not significantly better on every dimension. One school of thought: the base model is just not as good, and that is what matters. Gil subscribes to the other: a new scaling law, a new dimension of competition. How good it is at this exact moment is not the important thing.

Guo’s analogy is GPT-2 for that approach: an early proof that something real is happening, then the same laws of more compute and more data make it dramatically better. She expects people to get a lot more excited once labs attach tools and function calling to this kind of planning. She has been using o1 to find the bounds. A colleague dropped a coin-flipping brain teaser that requires simulating scenarios; she got it mostly wrong herself, then tried to see how much better the model was. The inability to force a correct answer even with clues drove her insane. She spent three and a half hours on it. She is bullish on test-time compute and looks forward to better ways to guide these models.

Then the Nobels. Physics went to two people who did pioneering work on AI models; chemistry went to Demis Hassabis and others on protein folding. Turing Awards have gone to Yann LeCun and others; this is the first time the Nobel has gone for AI. Guo starts with the obvious: these people are amazing and the work is world-changing. The Nobel’s criterion, though, is moving that particular field, and it is not clear to her that the physics prize advanced physics rather than applying physics principles elsewhere. Gil notes the traffic often runs the other way: diffusion models are statistical physics applied to image generation. So what is actually going on in physics that is interesting? Maybe it is a loop, AI for science feeding back into fundamental fields. Chemistry is more obvious, especially to Gil, a biologist by training: predicting structure, then function from structure, in ways beyond AlphaFold, changes what the industry can do. They joke about an AI Peace Prize—Adam D’Angelo for bridging OpenAI’s board, Mark Zuckerberg for releasing Llama, GPT-4 for writing copy—and Guo nominates Garry Tan. She is disappointed the AI community did not sweep. Maybe next year.

Gil’s macro question: what is most at risk—IT outsourcing, last-gen software, services vendors? He is involved with horizontal and vertical AI apps that do eat preexisting SaaS, but the more interesting move is augmenting or replacing people at scale. Customer support and success are the examples: Bret Taylor’s Sierra, and Decagon, a company he is involved with, making reps dramatically more efficient. He does not really think of that as replacing Zendesk. He thinks of it as changing how people are used. The tell is large numbers of people doing “email jobs”—cutting and pasting text, synthesizing, moving information around. Bonus points if the role needs multiple languages, which AI does well, or 24/7 availability, which AI also does well. Real-time voice models from Cartesia, ElevenLabs, OpenAI, and Google will make an agent that speaks twenty languages, never gets tired, and stays on track as support.

Guo adds a software-category lens. Beyond high-margin venture SaaS there are thousands of little vertical products: CRUD apps sold with distribution to non-technical SMBs, campground management, golf-course planning, window design. Constellation Software’s whole model is buying something like a hundred of those a year and aggregating a massive market cap. If those products are extractive and end users can generate the software, they go away; if distribution is the advantage, a new player can replace them. Web hosts face the same fork, because web development is first in line for code generation. For some outputs you will not want Adobe’s infinite control; you will trade it for cost and speed—the “Canva-fication” of those industries. AI attacks the people and also strips tools you no longer need.

Gil’s inverse is the “AI-durable” company, where AI does not matter much as a threat or, sadly, as a tool. The extreme non-software case is a railroad: you cannot lay that track again in the U.S.; the footprint is roughly fixed for twenty years. In software he offers Rippling, which recently shipped an AI product they discussed with its COO. A new AI HR company will not make headway against that bundle; the bigger risk is that AI shrinks customer headcount and they lose seats. Guo likes the example and rejects the lazy version: you cannot categorically say systems of record are protected. Gil is harsher: some products are a data store pretending to be something more, with an ecosystem that is most of the value. Those are much more at risk. The company shape she wants is SMB and mid-market first: generate a higher-quality database from source material—communications, documents—and replace the ingest-plus-approval-flow stack. Core enterprise systems of record are still durable because of process customization; the services spend around them is already attackable. They sign off as Luddites: Clone Wars, artificial wombs, and the future here, just not evenly distributed.