Karpathy on vibe coding, agentic engineering, Software 3.0

Key takeaways

1. Estimated date. April 2026, AI Startup School / "AIN" conference (reference to Sam Altman having come "last year"). YouTube video: https://www.youtube.com/watch?v=96jN2OCOfLs
2. Opening shock line. "He's never felt more behind as a programmer." — a statement that went viral on X/Twitter and that the interviewer picks up to open with.
3. December 2025 tipping point. Karpathy was on a break, had more time, and noticed that "the chunks just came out fine and then I kept asking for more and it just came out fine". No more need to correct. He insists: "a lot of people experienced AI last year as ChatGPT-adjacent thing. But you really had to look again and you had to look as of December because things have changed fundamentally."
4. Software 1.0 / 2.0 / 3.0.
5. 1.0. explicit code written by a human
6. 2.0. programming via dataset creation + neural network training (learned weights)
7. 3.0. programming = prompting; context is the lever on the LLM interpreter; the LLM becomes a programmable computer
8. openclaw example (Software 3.0 illustration). instead of a shell script that balloons to handle every platform, installation is "a copy-paste of text that you give to your agent". The agent looks at the environment, debugs in a loop — more powerful than a precise script.
9. MenuGen → Nanobanana (extreme case). Karpathy vibe-coded MenuGen (restaurant menu photo → OCR + image generation to visualize dishes) on Vercel. He then discovers you can just hand the photo to Gemini and say "use Nanobanana to overlay the things onto the menu" — Nanobanana returns the original image with the dishes overlaid in pixels. "All of my menu gen is spurious. It's working in the old paradigm. That app shouldn't exist." The neural net does everything, the prompt is the image, the output is the image. No app between the two. Karpathy's takeaway: don't think of AI as an acceleration of the existing paradigm, but as new things made possible.
10. New possibilities. LLM Knowledge Bases — "you get LLMs to create wikis for your organization or for you in person". This isn't code, it's a recompilation/reordering of documents to create a new projection. "This is not something that could exist before."
11. 2026 extrapolation (equivalent to the 90s web, 2010s mobile, cloud SaaS). Karpathy imagines neural computers where the neural net becomes the host process and CPUs become the co-processors — diffusion rendering a UI unique to the moment, from raw video/audio input. "In the 50s and 60s it was not really obvious whether computers would look like calculators or computers would look like neural nets. Of course we went down the calculator path." This branch could reverse itself piece by piece.
12. Verifiability framework. why are LLMs jagged?
13. "Traditional computers can easily automate what you can specify in code; LLMs can easily automate what you can verify."
14. Frontier labs train via RL with verification rewards → peak in math/code and adjacent domains, rough around the edges elsewhere.
15. Combination: verifiable + labs care (what makes it into the data mix based on economic value).
16. Chess anecdote GPT-3.5 → GPT-4: a huge improvement due to a lot of chess data intentionally added to pre-training, not to general progress.
17. Consequence: you're "slightly at the mercy of whatever the labs are doing". If you're inside an RL circuit, you fly. Otherwise, in-house fine-tuning is necessary.
18. Modern jaggedness example. "I want to go to a car wash to wash my car and it's 50 meters away. Should I drive or should I walk? State-of-the-art models today will tell you to walk because it's so close. How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a 100,000 line codebase or find zero day vulnerabilities and yet tells me to walk to this car wash? This is insane."
19. Advice to founders. target verifiable domains where you can create your own RL environments / examples → fine-tuning works as a lever. "Verifiability remains true even if the labs are not focusing on it directly." Karpathy declines to disclose a specific domain on stage: "I don't want to vibe post on stage."
20. On what is NOT automatable. "Ultimately almost everything can be made verifiable to some extent. Even for writing, you can imagine having a council of LLM judges." For Karpathy: everything is ultimately automatable, it's just more or less easy.
21. Vibe coding vs Agentic engineering (key distinction).
22. Vibe coding. = raise the floor. Anyone can vibe-code anything. Democratization.
23. Agentic engineering. = preserve the quality bar of professional software. "You're not allowed to introduce vulnerabilities due to vibe coding. You're still responsible for your software just as before, but can you go faster?"
24. It is an engineering discipline: coordinating spiky/stochastic agents to go fast without sacrificing quality.
25. The 10x engineer is magnified. "10x is not the speed up you gain. People who are very good at this peak a lot more than 10x."
26. On hiring (highly actionable point). "Most people have still not refactored their hiring process for agentic engineer capability. If you're giving out puzzles to solve, this is still the old paradigm. Hiring has to look like: give me a really big project and see someone implement that big project." Example: "Let's write a Twitter clone for agents, make it really good, make it really secure, then have some agents simulate activity, and I'm going to use 10 codecs 5.4x for X high to try to break your website. They should not be able to break it." (Karpathy's rephrasing of the Sierra AI-native interview — direct corroboration of Bret Taylor / Iyengar / Asemanfar / Wang.)
27. Human skills gaining value. taste, aesthetics, judgment, oversight, spec design. Agents are interns — they have excellent recall (PyTorch / NumPy / pandas API details you no longer need to memorize: keep_dims vs keep_dim, dim vs axis, reshape vs permute vs transpose), but they miss the fundamentals. MenuGen anecdote: the agent tried to cross-correlate Stripe and Google accounts by email address instead of using a persistent user ID — "this is such a weird thing to do."
28. The human's position. "You're in charge of the taste, the engineering, the design, that it makes sense, that you're asking for the right things. The engineers are doing the fill in the blanks." Karpathy isn't overly fond of plan mode itself but believes in detailed spec work co-designed with the agent.
29. Heart attack reading the generated code. "It's not super amazing code necessarily all the time and it's very bloaty and there's a lot of copy paste and there's awkward abstractions that are brittle and like it works but it's just really gross." On micro GPT: he tried having an LLM simplify it, "the models hate this. They can't do it. You feel like you're outside of the RL circuits. It's like pulling teeth." — proof that the aesthetics of simplicity isn't in the labs' RL.
30. Animals vs Ghosts. "We're not building animals, we are summoning ghosts." LLMs are not animal intelligences (yelling at them changes nothing). They are statistical simulation circuits: substrate = pre-training (statistical), RL bolted on top to grow the appendages. Karpathy admits: "I don't know that I have like here are the five obvious outcomes of how to make your system better. It's more just being suspicious of it and figuring out over time."
31. Agent-native infrastructure. "Everything is still fundamentally written for humans and has to be moved around. I still use most of the time when I use different frameworks or libraries... they still have docs that are fundamentally written for humans. This is my favorite pet peeve. Why are people still telling me what to do? I don't want to do anything. What is the thing I should copy paste to my agent?" Vision: breaking workloads down into sensors over the world / actuators over the world. Ultimate test: "I would hope that I could give a prompt to an LLM 'build menu gen' and then I didn't have to touch anything and it's deployed." The Vercel/DNS/Stripe configuration deployment was the real hassle, not the code.
32. Long term. "I'll have my agent talk to your agent to figure out details of meetings." Agent-based representation for people and organizations.
33. Closing formula (on education and knowledge). "You can outsource your thinking but you can't outsource your understanding." Karpathy: "I still have to somehow information still has to make it into my brain and I feel like I'm becoming a bottleneck of just even knowing what are we trying to build why is it worth doing how do I direct my agents." He champions LLM Knowledge Bases as a tool for enhanced understanding, through synthetic data generation over a fixed corpus (his articles → personal wiki).
34. Connection to the watch file.
35. Confirms and popularizes, through his voice, the Software 3.0 paradigm (cf. Greyling 2026-03-09 "the development environment is collapsing", Rauch 2026-01-02 "CLI as fundamental coding agent abstraction").
36. Vibe coding / agentic engineering distinction. crystallizes the debate carried forward since Kent Beck (2024-10), Mogère (2025-07), Yegge & Kim (2025-11), Beck Starving Genies (2026-04-03) — Karpathy supplies the stable vocabulary for the opposition.
37. Hiring refactoring. via large adversarial projects = explicit corroboration of Sierra (Taylor 2026-04-20, Iyengar/Asemanfar/Wang 2026-04-22) and Soto (Developer Taste 2026-04).
38. Verifiability. as an explanatory grid for the jagged frontier extends Mollick (2025-11-12 Giving your AI a Job Interview) and provides an operational framework (creating your own RL environments).
39. Animals vs ghosts. falls within the philosophical lineage of the notes on the ontological core (Seale 2025-05-30) and the malleability of the world (Andreessen 2026-02 / refutes Ralmuto 2026-03-17).
40. Agent-native infrastructure. extends Cloudflare Markdown for Agents (2026-02-12), Levie Building for Trillions of Agents (2026-03-07), Sierra (2026-04).