Appearance
Back in February 2025, Andrej Karpathy dropped the Vibe Coding concept and blew up every developer community on the internet. Just over a year later, he's back to correct his own idea, explain what Software 3.0 actually means, and lay out the ground rules for building real, production-ready things with AI right now.
For context: Karpathy is an OpenAI co-founder, led development of Tesla Autopilot, and is uniquely able to explain cutting-edge AI research in plain, practical terms. He recently joined Anthropic to work on next-generation foundation model training -- so these are not just hot takes, this is the thinking shaping the future of the technology.
What actually is Software 3.0?
This is the base framework you need to understand everything else.
We have now completed three fundamental paradigm shifts for how software gets built:
- Software 1.0: You write every line of exact instructions for the computer
- Software 2.0: You define loss functions, and train model weights
- Software 3.0: You only tell the system what you want. Not how.
Prompts are the new code. Context windows are the new IDE.
Karpathy uses a perfect example here: He once spent weeks building a small app that would take photos of restaurant menus and add product images next to every dish. That entire app does not need to exist anymore. You just drop the menu photo into a modern LLM, say "add pictures for every item", and get the finished result back. All that middle layer application code? Completely obsolete.
Vibe Coding didn't die. It split.
If you missed the original hype: Vibe Coding was the idea that you let AI write all the code, and you only check if the end result feels right. No line reviews, no edge case testing, just a vibe check and ship. It was fun, it let anyone build working prototypes in hours, and the entire developer world went crazy for it.
Now Karpathy is correcting the record:
Vibe Coding is not dead. It works perfectly for prototypes. It will never work for production.
Vibe Coding raised the floor for everyone -- now literally anyone can test an idea in an afternoon. But if you are shipping something that real people will rely on, you need the new discipline that replaced it: Agentic Engineering.
This is the critical line for product managers: You can and should use Vibe Coding to validate a hypothesis fast. But the second you decide to build that thing for real? Throw that prototype vibe code away. If you try to ship prototype code to production, you are not moving fast. You are just accumulating catastrophic technical debt that will explode later.
LLMs are ghosts, not animals
This is Karpathy's single most useful mental model for working with AI right now.
Stop treating LLMs like dumb interns, smart dogs, or any kind of living thing. They are ghosts. They have jagged, illogical intelligence. The exact same model that can reverse engineer a 100,000 line codebase and find an undiscovered zero-day vulnerability will also sincerely tell you to walk 50 meters to the car wash. It will not remember you drove there. It will not see the obvious contradiction.
For anyone building products with AI this means:
- Never assume competence in one area means competence anywhere else
- Always keep a human verification step, especially for use cases outside the model's training distribution
- Don't yell at it. Don't praise it. It has no feelings. Stop anthropomorphising. Look at the data.
The one rule for what AI can actually automate
Forget all the hot takes about which jobs will disappear. Karpathy states there is only one variable that matters: verifiability.
Old traditional computers automated things you could define exactly. Modern LLMs automate things you can verify exactly.
That is the entire boundary.
This is also the biggest untapped opportunity for AI products right now. Look for niches where:
- There is a clear, objective way to check if an output is correct
- No one has done targeted reinforcement learning training for this use case yet
Think specific legal compliance checks, industrial quality control steps, specialised code audits. These are not the flashy demo use cases everyone posts about. These are the places AI will actually create real, reliable value first.
What human skills are still valuable?
Everyone is asking the same question: if AI can write all the code, what am I even supposed to do?
Karpathy's answer is very clear:
You can outsource thinking. You can never outsource understanding.
Learning how to write good prompts is table stakes now. The actual rare, valuable skills right now are:
- Actually understanding your real user's needs
- Developing intuition for when the AI is wrong, before it breaks something
- Deciding what is even worth building in the first place
AI will execute almost anything you ask for. It will never tell you if you are building the wrong thing.
3 Actionable takeaways for AI product managers
Build an evaluation harness before you build the feature Stop just being impressed by demo videos. Before you ship any AI feature, first answer: How will we measure if this output is good? Who checks it? What do we do when it fails? If you can't answer these, don't ship.
Filter every product idea through the verifiability test If you can't clearly explain how you will confirm the AI did the job correctly, walk away. You will end up with something that looks great in demos and falls apart the second real users touch it.
Stop practising tool usage. Practise understanding. Anyone can learn to use Claude or Cursor in a week. No AI can replace the work of talking to users, learning your domain, and building good judgement about what matters. That is your moat now.
Closing thought
We are not at the end of this shift. We are at the very beginning.
Vibe Coding was the fun, wild first year of this new world. That party is over for production work. Now comes the actual engineering part.
Software 3.0 is not a threat. It is the most powerful lever any developer or product person has ever had. But it only works if you stop pretending AI is magic, learn its actual boundaries, and learn how to work with it properly.