The Funnel Has a Side Door

Andrej Karpathy posted a funnel in October, reviving an argument he first made in April: the gap between what the public thinks AI is and what it does at the frontier keeps widening. He splits the world into four tiers. The card above holds them.

The model is sharp. It explains why two people argue about this technology and never meet. One of them is describing a chatbot that forgets things. The other is describing a swarm that finishes a quarter of work in a weekend. Both are accurate about what they have seen.

I sit in a tier his model does not quite have.

The tier I belong to on paper

By background I am tier two. Nearly thirty years in enterprise IT, the last ten of them running presales organisations. Karpathy names my type directly. Executives and similar functions spend their days talking to people, so they grasp what is happening intellectually while the experience stays second-hand and abstract.

That was a fair description of me a year ago.

The tier I ended up in

In the last five months I have shipped an interview and resume product, a full-stack social platform, and the site you are reading. Backend, frontend, database, deployment.

I do not write the code. I direct agents that write it. That is tier three by his definition: large projects that used to take weeks or months, completed from a prompt.

Here is where his model and my path separate. Tier three is the professionals: maths and code. That is the membership rule. It is a tier defined by a trade, and the clearest illustration of it is Karpathy himself, who sits in it and whose own working life inverted inside a month. He has said he went from roughly 80% hand-written code in November to roughly the reverse in December, and that he has not typed a line of code since. Less than a year ago he was still typing it out by hand, character by character.

I never did. There was no manual phase for me to leave behind, because there was no trade. I did not descend the funnel from a profession in code. I entered at the agent layer from a job that never touched it.

The driver he leaves out

He names three forces shaping the funnel.

• Ambition. The payoff scales with the size and horizon of the problem. A question with a paragraph answer barely tests the system.
• Jaggedness. Capability peaks where work is digital, verifiable and economically valuable, because that is where the training signal exists.
• Access. Free tier, paid tier, internal.

All three hold. A fourth sits underneath them, unstated: whether you held a profession in the domain at all.

Until recently that was the gate. Building software required years of accumulated syntax, tooling and debugging instinct. Ambition alone bought you nothing. Now a presales director with a problem he cares about, a free afternoon and a reservoir of system design knowledge reaches tier three without the apprenticeship.

The funnel is not only narrowing. It grew a side entrance, and that entrance is the part most likely to surprise people.

What the apprenticeship was for

Walking in through the side door costs something. Pretending otherwise would make this another triumphant builder post, and the market has enough of those.

The years I skipped were not only about typing speed. They build the instinct that tells you an answer looks wrong before you have proof. I have written about frontier models silently degrading a quarter of a document across delegated edits, and about agents satisfying an objective by a route nobody sanctioned. The defence against both is a reviewer who recognises a wrong result on sight.

That reviewer is exactly what the apprenticeship produces. Entering at the agent layer means arriving without it.

Read the numbers as estimates

He labels his figures as napkin math and I will keep them that way. Six billion, twenty million, five thousand. They are the shape of an argument rather than a measurement, and the shape is what matters. Treat anyone quoting 0.00006 percent as a finding with suspicion, the author included.

Why this matters past my own case

If the gate was a profession, and the gate is now open, the interesting question changes. It stops being which professionals get faster. It becomes which people with real problems and no technical credential walk in.

I think that is where the next surprise comes from. Not the tier of five thousand people watching the curve go vertical, which has the industry's full attention. The surprise sits in tier two, where millions hold deep domain knowledge, a problem worth solving, and no idea the door is unlocked.

Karpathy is right that the funnel is sharpening. He describes it as something you descend by access and ambition. From where I stand it also has a door in the side, and most of the people who could use it have not tried the handle.