a builder's codex
codex · operators · Andrej Karpathy · ins_judgment-vs-understanding

You can outsource thinking, but not understanding — verification is the new human job

By Andrej Karpathy · AI researcher; ex-OpenAI, ex-Tesla; founder of Eureka Labs · 2026-04-30 · talk · Sequoia AI Ascent 2026 fireside (Software 3.0)

Tier A · TL;DR
You can outsource thinking, but not understanding — verification is the new human job

Claim

The bottleneck of AI-native work is not intelligence, it is verification. Models will produce more reasoning, more code, more drafts than any team can consume; what they cannot do is make the call about whether a given output is right in this specific context. Thinking can be outsourced to the model. Understanding — the calibration of "is this true, does it fit, what does it imply for the rest of the system" — has to stay with a human. The implication for how teams operate: design every system so the human's role is verification, not production.

Mechanism

LLM capability spikes where there are dense data and clean reward signals; it lags where evaluation is judgement-loaded. The human comparative advantage is the ability to integrate context the model has never seen — the unsaid requirements, the political constraints, the analogous prior failure, the way a number looks slightly off because of last quarter's incident. Calling that "understanding" instead of "thinking" reframes the workflow: production is no longer the scarce step, so production roles must redesign around verification. The teams that fight this and try to keep humans on production work get out-shipped by the teams that move humans to verification and let the model handle production. The teams that go too far and remove human verification entirely get away with it until they don't.

Conditions

Holds when:

Fails when:

Evidence

"You can outsource your thinking, but you can't outsource your understanding."

— Andrej Karpathy, Sequoia AI Ascent 2026 fireside, 2026-04-30.

The same talk argues that verification capability is what makes some labs win and others stall. Independent convergence: Harrison Chase + Eugene Yan reached the same conclusion the following week from the agent-observability and evals angles respectively (see A trace alone teaches nothing; learning requires feedback attached to the trace).

Signals

Counter-evidence

Cross-references

Open the interactive view → View original source → Markdown source →