How Karpathy Fell Behind (and Why You Did Too) 🫠

Recently, Andrej Karpathy spoke at Sequoia Ascent 2026 and admitted that he had never felt so behind as a programmer.

Most people still perceive LLMs as just an advanced autocomplete. An accelerator for writing good old Software 1.0. But the essence of the current shift is not about writing classic code faster. It's about huge swaths of code that should no longer exist at all.

Here are three takeaways from his talk and subsequent post:

1️⃣ Code for code's sake is dead
Efficiency is not about "writing an app faster." It's about not writing it at all. The model acts as a natural language interpreter: it evaluates your environment, executes commands, and debugs errors on its own.
What was fundamentally impossible to solve with classical algorithms (computations over unstructured data from articles/PDFs/logs) can now be solved architecturally in a few clicks.

2️⃣ Neural network schizophrenia
The model can flawlessly refactor 100,000 lines of legacy code, and in the next prompt, seriously suggest you walk to the car wash to clean your car.

Why? Economics.
Labs pour millions into RLHF where there is clear verifiability of results (math, code) and a clear market. There, the neural network runs on rails. Step aside, and the model machetes through the jungle, hallucinating at every step. If you don't understand this boundary and where you are, the LLM will kill your prod.

3️⃣ Agent-native architecture
Products are finally decomposing into sensors, actuators, and logic (LLM). We are moving toward neural networks driving the main computation flow, with the classic CPU becoming just a coprocessor for deterministic tasks.

In short, "Agentic engineering" is already a hard skill. If you don't know how to make information maximally machine-readable for LLMs and orchestrate agents, your skills are becoming obsolete right now.