Karpathy LLM capability gap: what his funnel post says

Andrej Karpathy's post on the LLM capability gap says most people are judging AI by the weakest version they have used. His tiers run from about 6 billion people who have barely touched an LLM down to roughly 5,000 insiders who use the strongest systems before anyone else, and he says the distance between those tiers keeps growing.

He posted it on Sunday, October 11, as an update to an April post of his, and it had more than four million views by Monday.

What Karpathy posted

In his post, @karpathy says the gap in shared understanding of LLM capability is "widening." It used to look like two groups of people talking past each other. Now, he writes, it is "more a sharp funnel." He is building on his April post, which opened: "Judging by my tl there is a growing gap in understanding of AI capability."

He calls his numbers napkin math. Here is the funnel:

• About 6 billion people (roughly 75% of the world) have barely come in contact with LLMs.
• About 1 to 2 billion are "casual and infrequent users of free-tier ChatGPT-like products," mostly treating them as a better search engine or a writing aid.
• About 20 million work in math and code and have seen agents finish "large, complex projects that used to take them weeks/months."
• About 5,000 have internal access to frontier systems, where he describes "swarms of thousands of agents collaborating over weeks on software mega projects."

His last line is a plain observation: "And it is all happening at the same time."

His three reasons the funnel exists

Karpathy gives three reasons the experience splits so sharply.

The first is ambition. He writes that "the impact scales with ambition, problem size, and horizon." Ask for a paragraph and you barely stretch the model. Hand it a project that runs for weeks and the difference between models becomes obvious.

The second is the domain. "Capability peaks in domains that are digital, verifiable and economically valuable." Code and math fit that description. When an answer can be checked automatically, labs can train hard against it. Work with fuzzier feedback improves more slowly, so in those areas the models look less impressive.

The third is access: "Free-tier, paid-tier, internal." Someone using a free chatbot and someone on a paid frontier model are using different products, even when both are called ChatGPT.

The reaction on X

The post set off a long thread of replies. @kimmonismus wrote from Paris that the people they had been talking with use AI now and then for search or writing, and that many professionals only follow its progress in theory. Their summary of the heavy users, themselves included: "we're living in an ivory tower."

It's a fair jab at people who live on AI Twitter, and it fits Karpathy's framing. The casual users see a weaker product aimed at easier tasks, and they draw sensible conclusions from what they've seen.

Our read

The funnel is the most useful thing in the post because it explains why two smart people can watch the same technology and disagree about it. A plumber who tried a free chatbot once and got a bland email draft is reporting what they actually saw. An engineer whose agent rebuilt a billing system over a weekend is too. Both sit in the funnel, at different depths.

Karpathy's second reason matters most for businesses. The tasks that improve fastest are digital and checkable. A lot of small-business work fits that shape, even though it doesn't feel technical: matching invoices to payments, confirming appointments, routing leads, reconciling a spreadsheet against the CRM. Each one has a right answer you can verify. Those are good jobs to hand to an agent. "Write our brand story" is a poor one.

The funnel also says something about owners. Most of them are in the second tier: they tried a free tool, saw a slightly better search box, and filed AI under "not yet." What usually moves someone to the third tier is a big, checkable task on a paid model, with a person watching the output. A cleverer prompt rarely does it.

What this means for a small business

If your picture of AI comes from a free chatbot, test it again on one real process with a clear right answer, using a current paid model. Measure the result against how the work gets done today.

If you'd like help choosing that first process, New Face Design's free process audit goes through your week and finds the repetitive, checkable work an AI agent can take over safely. That's often where an owner first sees what the people in the third tier keep describing.