Andrej Karpathy's (@karpathy) keynote yesterday at AI Startup School in San Francisco.
A full breakdown of Andrej Karpathy’s vision for Software 3.0, structured + cited from raw commentary. LLM-ready. 🔗 Full version: pdf.shrinked.ai/andrej-karpath… What’s inside: → Software 1.0 → 2.0 → 3.0 → “Prompts as programs” and LLMs as new operating systems → Vibe coding, agent UX, and human-in-the-loop architectures → Why we need to build for partial autonomy, not AGI demos → “We need to rewrite a ton of code”
@ycombinator @karpathy hands down the best talk of the conference thank you @karpathy
@karpathy breaks it down so clearly it feels obvious in hindsight: software isn’t just code anymore. it’s prompts, weights, GUI, and behavior. his framework is deeply simple: -software 1.0: you write code -software 2.0: you train weights -software 3.0: you write in english LLMs aren’t APIs. they’re operating systems. prompts are programs. context windows are memory. outputs are behavior. most founders are building agents. karpathy says: build sliders. let the human choose how much control to delegate. @cursor and @AskPerplexity nailed this: partial autonomy that feels like power. if you’re a builder: -think like it’s 1965 -make your product readable by people and models -design sliders, not agents -build tools users can verify, not just use
@ycombinator @karpathy Bluemintai.com here you go software 3.0
One of the new key metric goals we will see for consumer AI tools will be: "how can we minimize the amount of iterations it takes for the user to reach their desired output." A big part of this challenge is prompt writing, it’s still one of the steepest learning curves for users. Soon, we’ll see it fully abstracted away. A user can write something simple, and an intermediary LLM loop will automatically enhance the prompt before it’s sent to the core model for execution. This kind of backend refinement will quietly reduce friction and help users get better results, faster.
@ycombinator @karpathy The day an AI explains stuff better than Karpathy we will achieve AGI
@ycombinator @karpathy Software 4.0–> you tell llm what software you want and instead if it generating the software the llm will act / behave exactly like the software you wanted
here are my personal notes focus is on actionable content btw feel free to watch the vid if you're interested in the history of LLM's --- Main Takeaway: full llm autonomy will take way longer than you think. don't try to fully automate your apps. Specific Notes: - there will be a massive long tail of solving edge cases - he took a fully autonomous ride in a waymo in 2013 and tesla still hasn't solved FSD - build partial autonomy apps with an autonomy slider - think how cursor has tab mode, edit mode or agent mode - build tools to convert current tools into ai compatible tools - make your docs and apis llm friendly with markdown and programmatic commands like "cURL" instead of "click here" - ai native tools are the future (MCP to auto grab all creds rather than having to go to clerk and grab them yourself) - the web will be converted from software 1.0 to software 2.0 and 3.0 There will be money falling from the sky for all of us We're all gonna make it 😁
@ycombinator @karpathy 1980's PC era v2 for hobbyists. gpt2 is the new 6502.
@ycombinator @karpathy After seeing that he spoke here, this is one I've been waiting for.
@ycombinator @karpathy Amazing presentation thank you for sharing. Probably the best AI presentation. I’ve seen much better than listening to the AI CEOs talk about their products.
@ycombinator @karpathy Must watch for everyone …
@ycombinator @karpathy We made an incredible GUI for AI. @karpathy , You gonna love it.
@ycombinator @karpathy this was such a great talk, thanks for sharing it so soon
@ycombinator @karpathy turns out web3 meant something else
@ycombinator @karpathy didn't have to wait 2 weeks, thank you
@ycombinator @karpathy Thanks for sharing this yc!
@ycombinator @karpathy Me: Mum I want to go to AI Startup School Mom: We have AI startup school at home These posts:
@ycombinator @karpathy This talk did not feel like 40 mins long watching it live, wish they let Andrej accept a QnA!
@ycombinator @karpathy Insights like these fuel responsible AI innovation. ✨