Karpathy: build apps not base models

AI startup advice got a useful reality check this week. Andrej Karpathy — former OpenAI researcher, former Tesla AI lead, now founder of Eureka Labs — used a Sequoia Ascent 2026 fireside chat and a written recap posted April 30 to argue that the next big wins are not new base models. They are apps. More specifically, they are products that wrap existing models in context, tools, memory, and workflow until they become actually useful at work. That sounds obvious, but it cuts against a lot of founder instinct right now. (karpathy.bearblog.dev) ### What changed in his framing? The new wrinkle is timing. Karpathy says December 2025 felt like an “agentic inflection point.” Before that, coding agents were impressive but babysitting-heavy. Then the chunks got bigger, cleaner, and more dependable. The unit of work stopped being “write this line” and started becoming “implement this feature,” “research this library,” or “run the tests and fix the failures.” That is a different product world. (karpathy.bearblog.dev) ### Why does that push founders toward apps? Because once the model is good enough, the hard part moves. Karpathy’s Software 3.0 idea is that the real “program” is no longer just code or weights. It is the whole context window — prompts, instructions, examples, tools, memory, and retrieved information. In plain English, the value shifts from training the brain to building the exoskeleton around(karpathy.bearblog.dev) Google on foundation models. (karpathy.bearblog.dev) ### So what does “build apps” actually mean? It means taking a strong general model and slotting it into a real job. Internal copilots. Research agents with access to company docs. Customer support systems that can actually take actions. Industry software where the model sits inside the workflow instead of beside it in a chat box. Karpathy’s framing keeps coming back to orchestration — the huma(karpathy.bearblog.dev)n safe and legible. (karpathy.bearblog.dev) ### Why not just build another model anyway? Because the economics are brutal. Frontier models increasingly look like fabs or utilities in Karpathy’s older YC talk — expensive, capital-intensive infrastructure that a handful of players can sustain. Even if an upstart model is technically good, it still has to catch up on scale, distribution, tooling, and trust. The application layer has more ro(karpathy.bearblog.dev)iant training runs. That is basically the opening he is pointing at. (ycombinator.com) ### What makes these apps hard? Reliability. Karpathy has been consistent on this point. LLMs are powerful but jagged — brilliant in one moment, weirdly brittle in the next. So the winning products are not the ones that merely expose a model. They shape the task, constrain the action space, add checks, and keep humans in the loop where it matters. Think less “fully autonomous employee” and more “very capable junior operator with supervision and tools.” (ycombinator.com) ### Why does workflow matter so much? Because software adoption usually follows pain, not demos. A flashy model demo gets attention. A product that saves someone 45 minutes inside Salesforce, Excel, Zendesk, Figma, or an internal knowledge base gets budget. Karpathy’s point is not that models stopped mattering. It is that for most startups, model novelty is no longer the bottleneck. Distribution, trust, and fit are. (karpathy.bearblog.dev) ### What should founders take from this? Treat the model like a component, not the company. Build around memory, permissions, review loops, domain context, and action-taking. Make the AI feel less like a chatbot and more like a dependable part of the stack. The model labs are building the engine. Karpathy is telling everyone else to build the vehicle. (karpathy.bearblog.dev)ot anti-model. It is anti-delusion. In 2026, the easier story is “we’re training a breakthrough model.” The better business may be the boring-sounding one — take a great existing model, wire it into real work, and make people trust it enough to use it every day. (karpathy.bearblog.dev)