He Couldn't Convince Anyone to Compete with OpenAI. Two Years Later the Valley Is Full of Them.

Nathan Labenz and Erik Torenberg host Cognitive Revolution with two people who have been scouting AI longer than ChatGPT has been a household name. Sarah Guo, formerly a Greylock partner, launched Conviction, a $100 million AI-focused fund, the previous fall. Elad Gil is the angel whose recent checks include Character.AI, Harvey, and Perplexity; he also co-hosts No Priors with Guo. The conversation is four or five months after ChatGPT, and the whole world has turned its attention to AI. They have been thinking about it, independently and together, well before that.

Guo’s fund thesis is Software 3.0, which she calls shorthand for believing a very unexpected new set of software businesses is emerging. Machine learning has been creating capabilities exponentially. She and Gil even debated, six-plus months earlier, that plenty of ML-first companies had not worked. What changed her mind is that the exponential is real this time. The software is different in what it does: copywriting, illustration, and law were not big software markets. It is different in what the companies look like. How many people does it take to create a one- or ten-billion-dollar company? “Now we have empirical proof like 20,” she says — not true in previous generations of software.

The third difference is still unexplored: how the product should work as UX and human interaction. She expects “a lot more than just the single chat box.” Asked how her thinking has evolved up and down the stack, she points to demand that is obvious in hindsight and still surprising. Anyone who has met Babelfish understands translation and dubbing. Zoom out, and synthetic voice — cheaply turning one form of media into another — has been more in demand, across more use cases, than she expected. The bar to invest is high because the cost to build and the advantage of the problems are high. They have not done a chip company; they have done companies otherwise up and down the stack.

Labenz had spent sixty days red-teaming GPT-4 and came away thinking enterprises would apply the technology fast enough that a language-model startup would not disrupt them. He asks Gil to grade that intuition. Gil’s interest in generative AI started four or five years earlier with GANs, because the art was interesting. Before that he had worked on AI products for ten or fifteen years: mobile and ads targeting at Google, search at Twitter after he sold a company there, then a decade of investing. For about a decade, he says, nothing worked — or rather, a lot worked for incumbents and almost nothing for startups. Facebook’s News Feed and Amazon’s Alexa were huge. Companies started specifically to do ML did not become massive.

Things got interesting around GPT-2, then GPT-3 was a step function. He talked about it on a podcast when a lot of people were still ignoring the area. He started digging around for someone to build an OpenAI competitor and could not convince anybody. “Everybody said well it’s not that interesting of a business” and wondered if the APIs were that good. He pitched person after person. Nobody was willing to try. People who had worked in the area wanted to build applications instead, so he invested there: Character.AI, with Noam Shazeer, a main author of the Transformer paper; he helped an early Adept team; Perplexity; Harvey. A year or two later, some of those companies look like the interesting ones of this wave.

Every technology wave, Gil argues, has a different split of value between incumbents and startups. The first internet wave was about 80 percent startups — Google, Amazon — with Microsoft benefiting too. Mobile was 80 percent incumbent: Apple and Google already owned the platforms; Salesforce on the iPhone was Salesforce; search on the phone was Google; the new names were Uber, Instacart, Instagram. Crypto was 100 percent startup. The first decade of AI, with CNNs, RNNs, and GAN-related approaches, sent all the value to incumbents. This wave he scores maybe 80/20, incumbents over startups. Twenty percent is a lot when the shift may be the biggest in a decade, two decades, or longer. As Guo said, it changes the compute model and how you write code, the user interface, and the baseline of what this wave of computing can actually do. Base models will largely be incumbents — OpenAI, Microsoft, Google — with Anthropic and one or two others as the counterexamples. There is still room to build brand-new things.

If talented listeners should not try to clone OpenAI, what should they build? Gil’s four quick answers: voice applications, social, certain big B2B applications, and certain types of infrastructure. Guo adds that a language model can reason against knowledge it holds, a database, or a repository — and it can take action. The last generation of RPA never got flexible enough. Enterprises still have people moving data between systems and filling forms from policy, especially in the back office and in verticals like healthcare. Natural language plus tool use makes that interesting. Architecturally she wants retrieval and memory solved well enough to use in real enterprise applications; research is noisy, products are not there. And she is interested in emergent media creation. Midjourney democratized illustration, a category she promises not a lot of VCs were focused on before.

Gil has ideas about generative social products and does not want to dox the teams. What he will say is that most social products he sees are trying to rebuild Twitter: decentralized Twitter, Twitter-but-whatever, Facebook throwbacks. After he left Google, people built for the past — SEO traffic plays — instead of developer tools or mobile. He thinks a lot of social is a reflection of the past, not the future. Reid Hoffman’s seven deadly sins, in his telling, sit under every successful social product: gluttony, lust, and the rest. Run those through a generative lens and you get new interactions instead of throwbacks.

Consumer AI, Guo says, requires open-mindedness. Character.AI is one of the most interesting companies in the category. Replika is another that a lot of people dislike. She has been looking at consumer metrics for more than a decade, and if people spend hours a day on a service, that is so rare she is going to pay attention. Parasocial relationships and NSFW demand make it easy not to like. “That’s how a lot of things on the internet start.” Labenz had Eugenia Kuyda of Replika on the show; the product was not super compelling to him given how limited the models were, and he can see it becoming much more so. Replika 1.0, he thinks, will look quaint at 3.0.

Seven or eight years ago Gil invested in mobile coaching marketplaces — nutrition, fitness training. An accountability partner and an emotional relationship can change behavior, which is hard for humans. Bots can already coax people to plan their days and change habits. Education is the larger version: a bot that helps raise a kid, teaches, maybe becomes a best friend, with very positive and very negative implications. Every parent, he says, will want a hyper-customized tutor. Which company controls that, and what it means for how kids are taught, is barely being thought through. He reaches for Neal Stephenson’s *The Diamond Age* and the Young Lady’s Illustrated Primer, and for a science-fiction premise in which a large enough bloc of believers instantiates an AI agent to represent them in Congress. On a day-to-day level he wants an agent watching all his interactions, a personal trainer for everything — and, joking, a suck-up AI that tells him every joke was great.

Labenz’s GPT-4 experiments included an exercise coach and tech support for his 90-year-old grandmother, which hit a real family pain point. If a base model already plays the role, does the value accrue only to OpenAI? Gil thinks there is tons of room for standalone apps that wrap workflow, storage, history, or memory. In the 1990s everyone thought they would run their own email servers because email was a protocol; then it centralized on Gmail and Yahoo. Some things will live in a ChatGPT-like general tool. Many cases will need very specific workflow. The destination is a multi-agent world, with dominant specialists for dominant use cases — the way market structures evolved before.

Guo looks at tools that have succeeded at scale and sees a range of interaction modes depending on the task. Prompting is not easy for 99.99 percent of humans. Asking a good question is hard. Companies that give a particular category of users twenty prompt templates and an easy button, so they do not have to engineer a good output, are companies right now. She does not think generic interfaces will cover every use case. Search, she asserts, is going to break. Getting information from the internet has been stuffed into one Google box for too long. Travel planning and buying something are stereotypical tasks an agent should do better. Every slice of a fragmenting search market is valuable.

Gil’s other answer is that the interface almost does not matter. A sharp hedge-fund investor kept asking whether people would just talk into their phones. “Who cares it doesn’t matter.” The main point is what the technology fundamentally enables; interfaces will iterate, matched to senses and use cases. Kids love yelling at Alexa; nobody has a lengthy information-extraction conversation with it unless they are three. Models already outperform many doctors on standardized exams. Imagine uploading an image, describing symptoms, and getting care on par with a top medical school from anywhere with a capable phone. That impact dwarfs the interface. Guo’s enterprise picture is a compliance bot as an early-warning system in a fintech — hitting merchant-network rules is a definite no-go — and a bot you can debate, an extension of the security champion. Code generation is the near-term consumer version: a junior web or Python developer available to everyone, not autocomplete but write a file, run code, deploy to the cloud, use APIs. Elad’s joke for parents: there are base models, and based models with a D. Who does the RLHF, who adjudicates, and should a Western viewpoint be applied in a country with different values?

Does the coming weirdness push investors toward team over current product? Gil cites Chris Dixon: the next great company starts off looking like a toy. That was true in consumer, crypto, and certain enterprise. The really weird stuff is often the most interesting, mixed with the standard stuff you know will work. Snap started as disappearing selfies that looked insane. He is a market-driven investor, not a team-driven one — teams matter; he has started two companies — but product-market fit comes first. He looks for early signs of it, a big market, a great team, a path to defensibility, and a why-now. Almost every great startup has to be non-obvious, because if it were obvious everybody would already be doing it.

Guo’s warning to peers is not to go blind from your current view of the world, especially if you live with sophisticated customers. Actual demand may start with people who did not have access, or whose use case looks unsophisticated. Illustration is the exhibit: the view that it will never be good enough for picture books or Coke ads, or that it will not work for AAA games and Super Bowl video, will miss what people are actually using it for and where it is going. Investors also over-index on defensibility on day one. What was defensible about Retool or Notion in the early days? Two people, three months. The real first question is whether anybody cares and anybody is using it. Defensibility, scale, and whether this is an n-of-one company come later.

Labenz notes that everything has a waitlist and signup numbers are easy; retention is not. Gil’s favorite waitlist story is a decade-old “AI” company that was actually ops people answering queries. When the co-founder of a well-known large tech company tried it, they pinged the whole ops team to answer instantly. It stayed in private alpha, flashed a giant waitlist, never really launched, and got bought for a bunch of money. Infinitely closed waitlists, he says, are often a negative sign. If you have raw organic adoption, you usually just open the thing. The last decade taught founders the wrong lesson from Figma and Notion: talk to those CEOs and they wish they had gone faster. One company he is involved with was Superhuman-style hand-onboarding every customer until a customer asked why they were getting in the way. They stopped. Usage spiked.

Sustaining usage versus a one-month surge is a tiny percentage of what they see, Guo says, which is the point of venture: you invest in the tail. It is easy to feel smart by being cynical in a hype cycle, but if that is your orientation, do not invest in weird startup ideas. The outliers right now are insane. She knows a founder doing many millions of revenue who hit a cash-flow squeeze — GPUs, SVB, something — removed part of the free tier, and revenue went up a multiple. That is not fake. Those companies have not even played the “good company” games of hand onboarding and sexy waitlists. The hype cycle is useful for fundraising; investors who know what the data should look like will not buy it. Greed, social signaling, fun, and fear are all pointed at AI, which makes waitlists easy to inflate. The distance between demo and product is very large. Her first question to a company with a reasonable waitlist: what would happen if you gave everyone access? What is blocking real usage growth? “Garbage users anyway” is not an answer she wants.

Gil wants founders to ask when to raise versus bootstrap. People get on a venture train when they could just run the thing. If it is growing organically and spinning off cash, go for it. Social-gaming companies made real mistakes raising tons of money, then playing the venture game until the company blew up, instead of dividing the cash and moving on. He sees the pattern in social mobile apps sitting on Stable Diffusion: some ramped to $200 million of revenue in nine months and died off fast. Those companies might have been better served running on cash and distributing it. If you can avoid outside capital, why would you take it? Most people still grab the eight-figure checks. He is not saying most should bootstrap — only that it is another path, and people forget it exists.

Use of funds is shifting in two directions. Headcount can shrink — Perplexity’s Aravind Srinivas has said as much on this show — while foundation-model bills rise. Some companies take investor checks and turn around and spend them on OpenAI. Gil has seen companies raise $50 million to build a model when they should have gone on GPT-4 with some prompt engineering and had the same rough outcome. The subset where the model really matters is narrow: specific verticals such as healthcare with unique proprietary data, or training for particular interaction modalities. Price matters too. Training a diffusion model may cost hundreds of thousands or millions of dollars; a transformer-based model can be tens of millions. Image and video are a different check size from text.

Asked for products people should actually try, Guo starts simple: the magical features incumbents have shipped in Canva, Adobe, and Figma, and everybody should try ChatGPT. Rich contextual email completion is in waitlist beta and coming very quickly — complete with the meme that one agent will make your email longer and another will make it shorter. Character and other parasocial bots are, in her view, great. Agent products are probably too immature for the average consumer, but she thinks we are less than six months from bots that actually do work for you. Gil’s magical demo is dubbing: drop in a video, capture the speaker’s tonality, translate into Spanish. AutoGPT got a ton of attention; anyone in the AI community thought some form of it was obvious. The future is here, just not equally distributed, and a lot of those demos are about to hit broader audiences.

Labenz turns to Neuralink, which has reached great apes, and asks a hypothetical: a million people have implants, they look safe, would you get one? Gil jokes that he already has two, then says of course — an easy yes — if the thing works. Guo also says yes in principle, though they are probably not going to be alpha testers. Communication bandwidth with devices, and knowledge capture as an input mechanism, would become popular if it is an advantage. Gil is the skeptic on time. Understanding of the brain is “so de minimis,” except for a handful of systems like vision, and most of the deep-brain stimulation Neuralink is based on has been for depression and a few other diseases. He is quite skeptical anytime soon, and by soon he means five or ten years. Labenz calls it the most bearish Neuralink take they have had on the show. Guo’s advice: talk to neuroscientists.

Hopes and fears for society: Guo’s hope is the wealth of new capabilities, and a finding — she forgets the paper — that today’s enabling technologies help people with a lower skill base more than a higher one. If you train on the entire set of code, writing, or music, you lift the minimum more than the peak. That is interesting as democratization. Her policy fear is that people, including policymakers, do not understand that alignment and safety research is deeply tied to capability research. There is a version of the world that halts progress, or locks in regulatory capture, before anyone figures out what the systems can do. Nefarious actors will use the tools; the answer, as historically, is to build defensive technologies. Her personal fear is kids. What does she teach them on unstable ground — to be really good prompt engineers, or to go to Waldorf and not interact with tech? Very smart people disagree. She does not know.

Gil splits the horizon. Short term, meaning the next five to ten years, he is incredibly optimistic globally. Education and healthcare are the lines that only go up in price, partly through regulatory capture; this is one of the few technologies that may actually pull them down. Long run, meaning a few decades, he is a huge doomer on eventual species competition with AI or AGI. The biggest short-term risk to the field, in some sense, is regulation itself. The call to regulate is one-sided, and incumbents have an incentive to capture the models and block new entrants. He puts maybe a one-in-five chance that the next presidential race becomes the first AI-driven election — ad copy, targeting, robodialing with real-sounding voices — and then a regulatory storm, the way social companies were blamed when Trump won. Export controls on advanced chips make sense; a NIST-style approach maybe. Most other regulation, he thinks, distorts markets, kills innovation, locks in incumbents, and is way too early. The one-sided calls to regulate now are the wrong thing to do.