Open-source models are 5% of enterprise use — and GPT wrappers won

Redpoint's *Unsupervised Learning* sat down with Swyx and Alessio Fanelli, co-hosts of Latent Space, the technical newsletter and podcast that passed two million downloads in 2024. The format is a crossover rather than a guest interrogation: Jacob Effron and Jordan Segall, both Redpoint partners, spend their weeks asking AI founders the same hard questions, and so do the Latent Space hosts. Four people who already live in this market interview each other about last year's surprises, what is over- and under-hyped, which applications actually have product-market fit, and how an app-layer company builds a moat.

The number that hangs over the hour comes from Ankur Goyal, founder of Braintrust, who told Latent Space that open-source model usage inside enterprises is about 5% and going down. Enterprises, Swyx says, are still in use-case discovery. The working method is to take whatever looks like the most powerful model, see if anything useful falls out, and then start over when a new generation ships. Nobody has the energy to maintain a closed stack and an open one at the same time.

That 5% is not a verdict on the LocalLLaMA hobbyist scene, which Alessio says is still loud and sincere. It is a verdict on how large buyers actually adopt AI. The rest of the conversation is the map around that fact: reasoning models arrived with suspicious timing, GPT wrappers won the cultural argument, and the application layer — not a new foundation-model company — is where the interesting fights are.

Asked for the past year's biggest surprise, Alessio splits the answer into a good one and a bad one. The good surprise is reasoning models. Ilya Sutskever had just given a NeurIPS talk that pre-training was running out of road when OpenAI released o1, the system developed under the Strawberry codename. From the outside it felt like "it's so over" and "we're so back" in the same breath. Swyx finds the timing suspiciously neat: Strawberry had been in the works for roughly two years, and the date Noam Brown joined OpenAI is public. Pre-training "taps out" and inference-time compute becomes the new scaling law in one clean handoff. "If there were an Illuminati," he says, "this would be what they planned."

The bad surprise is open models. Community energy is real, Alessio says, but when they talk to enterprise buyers the reaction is closer to "it's cool." People love to argue about licenses; that argument has not changed the adoption path of AI. Jacob's reverse surprise is how fast open source caught up on reasoning, shrinking a gap he had worried might compound the way early scaling did. Swyx immediately corrects the wording: do not say "open source caught up"; say "DeepSeek caught up." There is no unified open-source camp, only companies that choose to release weights or not. After DeepSeek, most other open models are distilling from R1, and distillation catching up is a much lower bar than training something competitive from scratch. The card still worth waiting on is Llama 4.

He grants R1 one unique gift: full chain-of-thought traces. Previous open models were "the same model, just a little worse and cheaper." R1 is the first that published the reasoning process. On DeepSeek as a whole, though, he still calls it overhyped. Replicating a proven recipe is an order of magnitude cheaper than inventing one, and DeepSeek executed a known path extremely well without a fundamental new idea. Latent Space had already covered DeepSeek V3's fine-grained mixture-of-experts architecture in its year-end recap, when the focus was cheaper inference. Swyx admits they were too passive about raising their voice.

The public-market reaction was more theatrical. The day DeepSeek landed, Nvidia fell about 15% in a session, and hedge-fund friends from Swyx's finance days called asking why he had not tipped them off. His explanation is that once a public-market narrative locks in, it *becomes* the story, while the technical work is usually one to two years old. Their job as industry analysts, he says, is to focus attention on things that matter, and sometimes they are too quiet until the tape has already moved.

Latent Space's standing pitch, Swyx says, is AI engineering: what people can do *above* the model to make it more useful. It turns out to be a lot. The cultural tell is the rehabilitation of a slur. Two years ago, "GPT wrapper" was an insult. Aravind Srinivas of Perplexity came on their show and said he was proudly a wrapper. Swyx's summary of the shift: the industry went from making fun of GPT wrappers to an overwhelming consensus that wrappers are the only interesting thing. Anyone talking about differentiation before product-market fit is, in his view, making a ridiculous argument. Build something people want; worry about the moat later.

Jacob's sharper question is about the companies that should have owned this wave and did not. AI builders such as Bolt and Lovable went from zero to on the order of $20 million in revenue in about three months. Low-code companies that already had distribution, users, and the "non-technical people can build software" gene — Zapier, Airtable, Retool, Notion — largely missed it. They had the reach and the ability to fast-follow. Somehow there was nothing.

Swyx pins the miss on founder mindset. When Zapier decided to do Zapier AI, the product was natural language for creating Zapier actions. When Notion did Notion AI, it was writing documents and filling tables. A company that already has a baseline improves the baseline. The people who created a new species asked, with no prior preconceptions, what kind of software the model lets you build from scratch. There is a timing luck factor too: those products started their journey right as model capability hit an inflection. Both, he notes, reported numbers ridiculous enough that the four of them are still talking about them.

On overhyped, Alessio answers immediately: agent frameworks, no names. Workloads are still in violent flux, yet everyone is racing to be *the* framework. He gives LangChain credit for product velocity — shipping every day, even when the first abstractions were not the last ones — but says the market has moved. People now want something they can actually build on and stay on for a couple of years. His analogy: this is the jQuery era of agents, single-file, big helper libraries, and the industry needs React, which almost nobody is building. Jacob's amendment is that it may be too early for frameworks at all. Swyx says the work is the protocol, not the framework. Frameworks embed protocols; protocols can go first. MCP is the current answer, and the comparison is XMLHttpRequest enabling Ajax. On the TypeScript side, Mastra has impressive traction, but they still see a large opening.

Jordan's underhyped pick is Apple's Private Cloud Compute. Apple Intelligence he scores as a negative — the BBC publicly protested a notification that a man had shot himself when he had not — but PCC, which tries to bring on-device security architecture up into the cloud, he scores as a positive. Alessio fills in the industry constraint. Large enterprises are used to running everything in a private VPC. There are simply not enough GPUs for every company to stay single-tenant. The future is a multi-tenant architecture with single-tenant privacy guarantees, and that is a real problem, not a talking point.

Swyx's underhyped bet is memory, or stateful AI. For a conference keynote he surveyed how people define an agent, the way a researcher would, and categorized the camps. The week after, OpenAI launched its Agents SDK and Cloudflare did a version of the same. None of them have memory — conversation memory, yes; a knowledge graph of facts about you that exceeds the context window, no. MCP shipped a memory server in its first official set of servers, and that is where he tells people to start. Letta (the MemGPT team) spoke at their conference; Zep and LangChain's LangMem are in the mix. His judgment is that memory should be as normal in the stack as a database. The investing logic is blunt: anything stateful is interesting to VCs because it is a database, and everyone already knows how databases make money.

Jacob's overhyped add is that new companies training foundation models have resurfaced after he thought that wave had died. Robotics, biology, and materials science still interest him, because they need wet labs and data the big labs do not have. The core LLM-agent space, he says, is a brutal place to compete. A coda on BloombergGPT makes the point from the other side: the team concluded closed general models were better and stopped training. The person who ran it insisted the work still mattered. The data pipeline, the fine-tuning team, and the internal execution muscle all survived; he now runs a dozen GenAI teams across Bloomberg. The model died. Everything around the model lived.

Alessio, who invests mostly at seed through Decibel, says it is hard to buy things that already work, because that usually means you are late. He skews toward bets with little market risk and real technical risk: if it works, people will obviously use it, and it is unclear whether it will work. Swyx, answering more as an analyst than as an angel, says he has to know what has product-market fit because people ask him what is working *now*. Two years ago he wrote "Anatomy of Autonomy" with a $100 million revenue bar. Copilot, writing tools such as Jasper (since fallen off), and Cursor were on the list. The list of working form factors is getting longer.

The newest entry is Deep Research. Grok, Gemini, and Perplexity each have a version; Alessio has a finance-flavored bet in Brightwave. Swyx tried to handicap how much money OpenAI made by launching the product and landed on billions: the sheer upgrade volume from the $20 plan to $200 is that large, even if not every subscriber sticks. OpenAI then expanded access down into the $20 tier. He reads that as standing policy — whatever lives at the top tier, put a smaller version in the lower ones — because the brand of being first and the default choice is paramount. Google, he notes, had the product first. He told the Google team to their faces that OpenAI had "mogged" them, and they agreed.

Google's broader comeback is a running theme. Swyx's Gemini usage is up, from summarizing YouTube videos to native image generation to Flash. He runs a daily AI-news recap that is almost entirely model-generated and bake-offs the frontier models by hand; Flash wins most days. The developer bridges are still missing. You still cannot use Gemini's thinking models inside Cursor, though Logan Kilpatrick has said they are working on it. A model you cannot easily use might as well not exist. Asked for a public-market long/short, Jacob's combo is long Google, short Apple — and Swyx is happy to riff on the same book.

The second solid product-market-fit case is customer support. Brett Taylor of Sierra can raise any amount of money from anyone to do anything, and he is chairman of OpenAI; he still picked support. Swyx's read is that you can walk the income statement and org chart of any large company and pick cost centers off from the outside in, and support is huge, sticky, and durable. Decagon is running the same race. Jacob adds a narrative shift: the last two years of AI apps were about cutting cost; the next wave is growth. Jensen Huang's GTC line went from "the more you buy, the more you save" to "the more you buy, the more you make." First-wave apps hit work companies already outsourced to BPOs, so price competition will be fierce. Second-wave apps that move top line have a cleaner ROI story and may prove more defensible.

Voice sits in between. Jacob has looked at home-services scheduling: electricians miss about half their calls, so an agent that is reliable 75% of the time is still a large revenue lift. These jobs do not need 100% accuracy. Swyx's cold water is operational. If an electrician's inbound doubles, does the shop have bandwidth to do double the jobs? Most businesses cannot 2x or 3x overnight; the bottleneck moves to recruiting and training, a game of whack-a-mole. At a New York AI Engineer summit he met a bootstrapped team that only does appointment scheduling for veterinary clinics and usually attends vet conferences, not engineering ones. They are doing fine. His watch list for what is next: screen sharing that watches you work and offers help, outbound sales, hiring, education, finance, and personal AI, which is harder to monetize. A side quest last year was pushing Singapore's education system to be AI-native, with a stated goal that Python become a national first language; the Ministry of Education's first pushback was that teachers would be unprepared. Companies like Speak, which pick one job and walk the idea maze to the end, are rarer than they should be.

App-layer defensibility is the question Jacob's partners ask him every Monday. Swyx's one-word answer is network effects. Founders obsess over the single-player experience and neglect the multiplayer one. His load-bearing example is Chai Research, a Character.AI contemporary that bootstrapped, has no proprietary models, and runs a marketplace of community-submitted models. Short term, no proprietary IP hurts. Long term, sitting at the choke point between users and model providers, with money flowing through, makes them robust to any model change — and they have outlasted Character.AI, whose core team was absorbed by Google. Jacob adds brand. In some categories ACVs have doubled rather than compressed, because customers pay the premium of "this category equals this company." Velocity matters just as much: every three to six months a new model is an existential event, and whoever figures it out first wins. Internally, he admits, that sounds a lot like how they used to talk about application SaaS. There is no genius story. It is a thousand small things that compound. The early-AI pitch of a unique dataset and a self-trained model was, he says, a head fake.

The next six to twelve months' best show is model labs moving into the product layer, and the first battlefield is Cursor versus Anthropic. Cursor is now worth on the order of $10 billion, which has made it hard to acquire; Jacob jokes they should have sold to OpenAI at $5 or $6 billion and taken the coding data with them. Swyx splits coding into an inner loop, inside the IDE and within a git commit, and an outer loop, between commits in the cloud. To be an outer-loop coding agent you have to integrate with the customer's codebase and sign the security paperwork — schlep the labs do not want — and that is Cognition's moat. Sourcegraph, with a decade of enterprise contracts and existing trust on codebases, should be a natural winner, even if Cody has not yet done much on the outer loop. Long term they expect both sides to converge: you will want the IDE and the cloud agent.

On infrastructure they talk about an "LLM OS," a phrase they share with Andrej Karpathy, and they are more excited about the layer around the model than about bare metal. Code execution — they are investors in E2B — plus memory, search, and security. Security's logic is simple: wherever AI is used on offense, it has to be used on defense, across email, identity, and red teaming that used to be a checkbox. Models can now infer semantics, not just syntax; binary inspection can ask what a piece of code is trying to do overall, not only whether a rule fired. Swyx is less bullish on GPU inference serving. It is a real business, but capital-intensive and a large-fund game, and infrastructure pricing converges to cost-plus while applications can charge for utility. Fine-tuning companies he still cannot see as a standalone spike; they get wrapped into an enterprise offering such as Writer. AI SRE he doubts, calling it log anomaly detection, an old job. Jacob disagrees: the tech is not there yet, but even a 10% cut in mean time to recovery is a large business, and autonomous SRE is a few years out.

Three questions hang over the rest of the map. First, Swyx has been banging the drum on reinforcement learning. It clearly works on verifiable domains such as math and code. Whether it works on contracts, marketing copy, and outbound sales — domains with no single right answer — decides whether the future is autonomous agents or a pile of copilots with a human as the taste-maker. A world of fully autonomous coders and no autonomous sales emails is, Jacob notes, a sci-fi future nobody wrote. Second, former OpenAI exec Bob McGrew's "rule of nines": each jump from 90% reliability to 99% to 99.9% is an order of magnitude more compute, on a two-to-three-year cycle. Hardware supply is an open question, and so is whether Nvidia's CUDA wall gets cracked. Swyx is the only named individual angel in chip startup MatX; his bet is that the Transformer architecture is stable enough to bake into ASICs, and that anyone who started before 2019 or 2020 is still too general because Transformers had not yet won. Third, agent authentication. When Operator hits a website on your behalf, how does it prove it is your agent, not you? Most of the "rebuild the stack for agents" talk, he says, is unnecessary. This one is a new SSO. It is frustrating, he adds, when Sam Altman is right: maybe you do have to scan your eyeballs, and the rest of us are behind him again.

The last new idea of the hour is the one Swyx says came out of his New York conference most clearly: agent authentication. *The Information* had just written that companies are already worrying about it. When an operator-style agent visits a site on your behalf, it has to signal that it is not you. His broader stance on "agent experience" — reinventing every layer of the stack for agents — is that almost none of it is required. This one is. It is, in his phrase, new SSO for agents. Crypto people are amped to own it; he is not sure they will. The uncomfortable version is that Altman has been pushing biometric identity for years, and the rest of the industry is catching up to a problem he already saw.

For builders, the cheapest differentiation the four of them actually pointed to is already sitting in public. MCP shipped a memory server with its first official servers; Swyx's advice is to start there and give an agent memory that survives a session. For enterprise buyers still in discovery, the hour's evidence argues against waiting on open weights to get cheap enough. Take the strongest closed model, get a use case working, and treat R1's traces as study material for distillation rather than as a reason to run a second production stack today.

For anyone writing a plan, Jacob's three-to-six-month model-release cadence is a survival event, not a roadmap footnote. Whoever absorbs the new model first eats the increment, and the more honest question is whether the product cuts a cost center or creates revenue — the latter is the next wave. For anyone evaluating an AI startup, the unique-dataset, train-your-own-model story is the one they spent the hour dismantling. Look instead at network effects, category brand, and iteration speed. Those are the old SaaS answers. After an hour of frontier talk, they are still the answers.