🪴 Brain

Software 3.0

5 min read

Software 3.0

🗒️ Description

A breakdown of Andrej Karpathy’s Software 1.0 → 2.0 → 3.0 thesis (from his Sequoia Capital interview) reframed as a business playbook: which digital business models AI is making obsolete, and what is still worth building. The core reframe — when Karpathy says “software,” read it as all digital businesses, because every business is being projected through the software lens and restructured to be AI-first.

The promise: those who grasp the 3.0 paradigm early “will find more opportunity than has ever been possible before.”

🚀 The three paradigms

ParadigmHow you “program”Installer exampleBusiness analogue
Software 1.0Write explicit code — every rule, step by stepShell script: step 1 accept terms, step 2 find folder… brittle, not self-healingUdemy course taken in fixed lecture order; Premiere Pro manual editing
Software 2.0Program by curating datasets and training neural-network weightsInstaller reports errors back to Apple → ML feedback loop improves install rateYouTube engagement heatmaps + comments reshape the course; Descript AI-infused editor
Software 3.0Prompting — the context window is your lever over the LLM interpreterOpenClaw: copy-paste a skill to your agent; it inspects your machine, debugs in a loop, self-healsA personalized expert agent that does the work with you and delivers the finished outcome

“Software 3.0 is kind of about your programming now turns to prompting. And what’s in the context window is your lever over the interpreter that is the LLM.” — Andrej Karpathy

🧩 The four real-world examples

  1. The installer — 3.0 is an agent you summon with one command (the OpenClaw bash command). It owns the goal (“get OpenClaw installed”) and loops through problems it has never seen before, instead of following a fixed rulebook. This is a little “skill” — a copy-paste blob handed to your agent.
  2. The app (MenuGen) — Karpathy’s own 2.0 app (photograph a menu → render dish images via Nano Banana through an MCP connection) became “instantly useless.” The 3.0 version: just hand the photo to Gemini and say “use Nano Banana to overlay the items.” No app to download — “That app shouldn’t exist.” This is coming for almost every app.
  3. The course / specialized knowledge — 1.0 = sequential Udemy lectures; 2.0 = course reshaped by engagement data; 3.0 = a personalized agent (a “Max Verstappen in your headset”) that coaches you as you take action. Example: Alex Hormozi’s acquisition.com LLM trained on his whole brain — next step is “go build it with me,” not “ask me a question.”
  4. The service (video editing) — 1.0 = manual Premiere Pro rules; 2.0 = Descript auto-trimming silences/mistakes; 3.0 = a text box: “edit this in a viral MrBeast style, under 8 minutes” → finished render in 5–10 minutes. Software 3.0 sells the outcome, not the tool.

☘️ The four moats left

When the big LLMs (Gemini, ChatGPT, Claude) can do the task directly, four defensible moats remain:

  1. Your proprietary data — a model trained on your own knowledge/style (Hormozi’s LLM; an agent that edits in your last-100-videos style). Your value lives in the training data, used well so it isn’t AI slop.
  2. Prompt engineering / context — a good outcome needs both the knowledge and the instructions + context you give the engine. Get good at this.
  3. System design around the AI engine — the LLM is “just the engine”; you design the car, the wheels, the UX. Example: Higgsfield wraps Nano Banana in a UI that helps people get the most out of it.
  4. Trust from your audience — “I trained an LLM on all of Elon’s knowledge” from a random on Twitter vs. from Elon himself. Trust in the person and their access to genuine data is the most valuable asset.

📒 Takeaways

  • Don’t build static courses, single-purpose apps, or manual services that an LLM can now do directly — they’re being made obsolete.
  • Do build outcome-delivering agents wrapped in your own data, prompts, system design, and audience trust.
  • Every business is being restructured to be AI-first: AI as the engine in the middle, you design the system around it.
  • The shift mirrors Agentic Coding — you stop writing explicit logic and start designing the environment/context an intelligent agent operates in.

✍️ Quotes

“All businesses are literally being restructured to be AI first… AI in the middle being the engine and driving them forward. But that is all the engine is. You get to design your own car.”

“The trust that you have in Elon Musk… his access to his own data is the most valuable thing in the future that we’re heading towards.”

🔗 Related

  • Agentic Engineering — the primary Sequoia interview this video summarizes, with the vibe-coding→agentic-engineering distinction and verifiability
  • Agentic Coding — the same paradigm shift from the builder’s side: design agent environments, not code
  • Context Engineering — “what’s in the context window is your lever” made concrete
  • Self-Improving Company — Karpathy-adjacent thesis on AI-first org loops
  • Agent Skills — the copy-paste “little skill” install pattern Karpathy describes
  • LLM Knowledge Bases — Karpathy’s method for using your proprietary knowledge (moat #1)

📖 Further reading/watching

Template: knowledge_note_info

Graph View

Backlinks