Learnings from Karpathy’s Software Is Changing (Again) Talk — Notes

Andrej Karpathy: Software Is Changing (Again)

Software is evolving — from 1.0 to 2.0, and now to 3.0
- LLMs functions like a mix of electricity, fabs, and OSs — metered usage and easily switchable like electricity; centralized with a few companies controlling hardware and many renting services like fabs; you can run apps (like Cursor or Perplexity) on different models, like an OS.
Using AI effectively requires its own techniques
- Learn how to leverage the super-human capabilities while avoiding the “jagged edges.
- Optimize the generation–verification loop: AI generates, human verifies. This is much faster the more you are able to visualize the results. It can take you as much time to verify something as it would to do it yourself.
💡 Writing the core logic with AI is easy—deployment, auth, payment integration is the hard part. This work is still done manually via browser GUIs, which is inefficient for agents. It took 1 day to write his app and 2 weeks to make it “real.”
The hottest new coding language is English
- Big implications for how to design “AI Ready” processes — need clear goals, procedures to be followed, and when and how to check in. (Much like good human instructions!)
- Prediction: we’ll soon see “Job Descriptions” for Agents. Hopefully these will be better maintained than JDs for people! (When was the last time your sat with one of your employees and talked through a change to their existing job description?)
💡 The most AI-ready companies will use “Living Docs” that have specific goals and sub-processes, clear instructions on accountabilities and when to escalate, and can grow over time through collaborative check in docs. Like the GitLab or PostHog Handbooks.