The Bewilderment Test
Karpathy built the foundations of modern AI and says he can't keep up with it. That's the most credible signal the field has produced.
By Geordie Everitt
Andrej Karpathy has not written a line of code since December 2025.
He will tell you this himself — he mentioned it in a recent interview, with the matter-of-fact delivery of someone reporting a minor scheduling change rather than a professional rupture. He hasn't stopped because he burned out. He hasn't stopped because he retired. He stopped because the AI does it better, and he is — by his own account — overwhelmed by what's happening in the field he built.
That last sentence is the important one. Karpathy is not a casual observer who stumbled into a press cycle. He did the foundational work on convolutional neural networks. He ran AI at Tesla. He was among OpenAI's first researchers. He taught the Stanford deep learning courses that trained a significant fraction of the engineers now building the systems in question. If you drew a provenance map of the technology currently restructuring every knowledge-work industry on earth, Karpathy's contributions would appear close to the root.
And he's overwhelmed.
The Confidence Economy
AI commentary runs on a particular fuel: certainty. The assured voice captures the audience; the uncertain one loses it. The clean causal story — why the technology is doing what it's doing, where it will go, what you should do about it — circulates widely. The person who says "I'm finding it genuinely difficult to track" does not trend.
The problem is structural. Anyone paying close attention to what is actually happening — not the narrative that has formed around it, but the raw rate of change — is finding it difficult to keep up. That difficulty is a calibration.
The field improves between the time a benchmark is published and the time it is read. What was true of the leading model eighteen months ago was already partially obsolete at twelve. Capabilities that were supposed to plateau haven't. Research that represented the consensus last year has been revised — not once, but several times. The people working closest to the frontier describe something closer to managed uncertainty than mastery.
Karpathy is the most credentialed person in that group. His description of his own experience — confused, unable to keep pace with a field he contributed to creating — is a more accurate signal than any confident framework on offer. Calibrate accordingly.
What the Stack Looks Like From Above
"Moving up the stack" borrows its vocabulary from systems architecture. At the lower layers: hardware, network protocols, assembly, the kernel. At the higher layers: applications, services, user intent. The pattern that holds across every layer transition is this: when the layer below you becomes reliable enough — predictable enough, capable enough — you stop managing it directly and start assuming it.
A web developer doesn't write assembly. A pilot doesn't build the engine. The delegation is trust earned through track record.
Karpathy has stopped writing code because, by his own assessment, the layer below him has crossed a threshold he finds persuasive. He is now operating above the layer where code is produced — not because he has forgotten what code is, but because something else is producing it more effectively than he would. That is a specific technical claim about a specific threshold. He is the person best positioned to make it.
His threshold is not the same as yours or mine.
The Bewilderment Is the Credential
There is a particular signature in commentary that has lost contact with what's actually happening: everything coheres. The frameworks generalize cleanly. The narrative has a clear arc — disruption, adaptation, new equilibrium. The predictions arrive with percentages attached.
The cleanness is the tell.
The slide rule masters who set down their instruments for the calculator were not diminished by the handoff. The draughtsmen who traded their parallel bars for CAD were not acknowledging defeat. What changed was the layer at which expertise was required. The judgment — what to calculate, what to draw, whether the answer makes sense — remained essential. The execution moved to a different kind of machine.
Karpathy's move is the same one, a generation later, at a higher layer of abstraction. He still carries the full understanding of the stack — no one has taken that from him. What he has stopped doing is expressing that understanding through code. The judgment is his. The keystrokes belong to the model.
What he is describing, when he says he feels overwhelmed, is the experience of a person whose map of the territory was the best available — and who now watches the territory revise itself faster than any map can track. That is the most accurate description of the current moment anyone with real standing has offered.
Who to Trust
The field is full of guides. Some of them are selling frameworks. Some are selling courses. Some are genuinely trying to make sense of something that is, by the account of the person who understands it best, difficult to make sense of.
A reasonable heuristic: if the guide you're following has no uncertainty about any of this — if the framework is clean, the predictions are confident, the tone is composed — they are either not paying close attention or they have decided that your reassurance is worth more than your accuracy.
Karpathy is confused. He says so plainly. He is also the best-positioned person alive to distinguish between genuine confusion and the performed kind.
That's the test. If your AI expert isn't at least as bewildered as Andrej Karpathy, you should ask yourself what they're not telling you.
Published by Geordie