In this interview hosted by Sequoia Capital, Andrej Karpathy (former Director of AI at Tesla and founding member of OpenAI) speaks with Stephanie Zhan about the rapid evolution of AI in programming. Known for coining the term “vibe coding,” Karpathy explains a profound recent shift where he realized he felt “behind as a programmer” because AI models have become so competent at independently writing and executing large chunks of code.

The Shift to “Software 3.0”

Karpathy outlines the evolution of computing paradigms:

  • Software 1.0: Writing explicit rules and logic (traditional coding).
  • Software 2.0: Creating datasets and training neural networks (learned weights).
  • Software 3.0: Prompting an LLM. The context window is now the “lever” over the interpreter, and the neural network performs the computation in the digital information space.

He shares an example of building “Menu Gen,” an app he coded to scan restaurant menus and generate images of the dishes. He realized that under Software 3.0, the app itself shouldn’t even exist—you simply hand a photo of the menu to an AI and ask it to return a rendered image. The neural network does all the work, eliminating the need for traditional app infrastructure.

Verifiability and “Jagged Intelligence”

Karpathy explains that AI accelerates fastest in domains where output is highly verifiable (like math and coding) because AI labs can easily set up reinforcement learning (RL) environments to train them. However, this leads to a “jagged” intelligence: an AI might perfectly refactor a 100,000-line codebase or find zero-day vulnerabilities, but still fail a simple logic test (like telling you to walk rather than drive to a car wash because it’s only 50 meters away).

Vibe Coding vs. Agentic Engineering

Karpathy draws a clear distinction between two modern AI coding styles:

  • Vibe Coding: This raises the floor. Anyone can use AI to build software simply by prompting, making creation accessible to all.
  • Agentic Engineering: This preserves the professional quality bar. It involves coordinating highly capable (but sometimes unpredictable) AI agents to build secure, robust software at unprecedented speeds without sacrificing engineering standards. He suggests technical hiring should no longer be puzzle-based, but rather observing how a candidate coordinates multiple AI agents to build and defend complex systems.

The Future of Human Skills: Understanding over Thinking

Just as Hinton advised learning a physical trade like plumbing, Karpathy offers a profound piece of advice for the knowledge worker in an age where agents do the heavy lifting: focus on high-level aesthetics, judgment, and understanding.

He references a quote that fundamentally shifted his perspective: “You can outsource your thinking, but you can’t outsource your understanding.”

While agents can handle the details of APIs, write the code, and execute the “thinking” tasks, humans are still the bottleneck for understanding why something is worth building, how the fundamental architecture should be designed, and what the ultimate goal is. Karpathy notes that the human’s role is to direct the agents, maintain taste, and oversee the project, because LLMs still do not truly excel at genuine understanding.