TL;DR Link to heading

  • Treat LLMs as statistical ghosts, not curious animals. Use them empirically.
  • Build agent-native infrastructure (good docs, APIs, logs, sensors) rather than just human UIs.
  • You can outsource thinking, but you cannot outsource understanding. Human taste and judgment remain irreplaceable.
  • Future interviews and hiring will test agent-native skills: build, deploy, secure, and defend systems against other agents.

Software 3.0 Has Arrived Link to heading

Karpathy reframes the evolution of programming:

  • Software 1.0: Humans write explicit code
  • Software 2.0: Humans curate data; models learn weights
  • Software 3.0: Humans program via prompts, context, tools, and examples. The context window is the new program.

In this new paradigm, you no longer write every line. You delegate macro-tasks to agents: “implement this feature,” “refactor this module,” “set up this service and tests.” The LLM becomes an adaptive interpreter that reads your intent and executes.

Key Examples Link to heading

MenuGen: Take a photo of a restaurant menu. In the old world, you’d build OCR, backend, image generation, frontend, deployment, etc. In Software 3.0, you can often just prompt a multimodal model to directly transform the photo into the desired output. Much of the traditional software stack simply disappears.

Adaptive Systems: Instead of brittle shell scripts, agents dynamically debug environments and adapt instructions on the fly.

Vibe Coding vs Agentic Engineering Link to heading

  • Vibe Coding raises the floor — anyone can now build by describing ideas.
  • Agentic Engineering raises the ceiling. Professionals must master directing fallible agents while preserving correctness, security, taste, and system understanding.

The skill shifts from writing code to orchestration: writing clear specs, building evals, adding guardrails, inspecting outputs, and knowing when the model is hallucinating plausible nonsense.

Jagged Intelligence & Verifiability Link to heading

LLMs excel where outcomes are easy to verify (code that passes tests, math with clear answers, games with scores). They remain weak where verification is hard. This “jagged” capability landscape explains why coding improved so fast while other areas lag.

Founders should look for high-value domains that are verifiable, economically important, and still under-trained.

The Bottom Line Link to heading

Code generation and boilerplate are becoming commodities. The scarce resources are now deep understanding, orchestration skill, security mindset, and taste.