Andrej Karpathy: Software Is Changing (Again)

tl; dr:

  • AI is similar to large-scale utilities, like electricity. Huge infrastructure costs, metered access, etc.
  • text formats, like Markdown, are important. If you do your documentation in Markdown, the LLMs can effectively use it.
  • AI needs to be kept on a short leash – small incremental steps, test.
  • Work will shift to AI generating, and humans verifying

Notes:

  • Three types of software:
    • Software 1.0: code we write
    • Software 2.0: neural networks
    • Software 3.0: LLMs
  • Good idea to be proficient in all of these paradigms
  • Part 1: How to think about LLMs?
    • AI is the new electricity – Andrew Ng
    • CAPEX to train the LLM (build the grid)
    • OPEX to serve intelligence of APIs
    • Metered access
    • Intelligence “brownouts” when OpenAI goes down
    • LLMs have properties of fabs – Huge CAPEX, deep secrets
    • LLMs have properties of Operating Systems
      • increasingly complex software ecosystems
      • LLMs are software, trivial to copy & paste
      • LLM apps like Cursor can run on GPT, Claude, Gemeni, DeepSeek, etc.
      • 1950s - 1970s time-sharing era
        • we are in the mainframe, time-sharing mode
      • feels like we’re using the terminal
  • Part 2: LLM Psychology
    • stochastic simulations of humans
    • encyclopedic knowledge
    • can hallucinate
    • jagged intelligence – good at some things, but make really silly mistakes
    • no continual learning
    • gullible – prompt injection risks, leak private data
  • Part 3: Opportunities
    • Apps like Cursor and Perplexity
      • package state into a context window before calling LLM
      • orchestrate and call multiple models
      • application specific GUI
        • audit work of these falliable systems
      • Autonomy slider
    • keep prompts small and concrete and make small steps
    • how to keep AI on the leash
    • Vibe coding will be a gateway drug to programming
    • Vibe coded https://www.menugen.app
      • code was easy (few hours)
      • devops was hard part (week)
  • Part 4: build for agents

Some slides that stood out: