Most folks assume AI just keeps getting smarter across the board. The reality is way more lopsided, and one of the sharpest minds in the field just explained why. I caught Andrej Karpathy’s talk at Sequoia’s annual AI event, and the way he breaks down this jaggedness shifted how I think about building with these tools.
Quick context: Karpathy co-founded OpenAI, led Tesla’s self-driving work, and coined the term “vibe coding.” So when he says something fundamentally changed in December, it’s worth pausing on.
The Key Idea
The original poster argues we’ve moved from Software 1.0 (writing code) to Software 2.0 (training weights) to Software 3.0 (prompting LLMs). Your prompt is the program. The context window is the lever. The model is the interpreter doing the work.
Old Way vs New Way
Here’s the contrast that hit me hardest:
- Old install flow: ship a giant bash script that handles every OS edge case.
- New install flow: ship a paragraph of plain text that users paste into their agent.
The agent reads the environment, debugs in the loop, and adapts. You stop spelling out steps and start describing outcomes. That’s the whole shift.
Why AI Is Brilliant And Also Dumb
This was my favorite section. The expert explains the jaggedness with one word: verifiability.
- Code compiles or it doesn’t. Math checks out or it doesn’t. Easy to verify, easy to train with reinforcement learning.
- Common sense questions (like “should I walk 50 meters or drive?”) have no clean reward signal, so frontier models still flub them.
That’s why a state of the art model can refactor a million line codebase and still tell you to walk to a car wash that’s basically next door.
Vibe Coding vs Agentic Engineering
Karpathy draws a clean line between the two:
- Vibe coding raises the floor. Anyone can build something now.
- Agentic engineering raises the ceiling. Pros orchestrate swarms of agents without dropping the quality bar.
Different skills. Different goals. Don’t confuse them.
What Founders Should Actually Build
The creator’s advice for builders was sharp:
- Skip the obvious verifiable domains (math, code). The labs will own those.
- Hunt for valuable RL environments the labs aren’t focused on yet.
- Build agent first infrastructure. The whole internet still assumes a human is reading the docs.
- Make installation a paste-into-agent flow, not a step by step tutorial.
The Quote That Stuck
You can outsource your thinking, but you can’t outsource your understanding.
If you don’t actually grasp what your agents are doing, you become the bottleneck. Direction still requires comprehension.
Watch the full conversation for the deeper takes on animals vs ghosts, taste as a human moat, and where this savvy professional thinks Software 3.0 goes next.