Andrej Karpathy says only about 5,000 people have access to the real frontier of AI. He posted this on X on October 11, 2026, and the post had over 6.4 million views when I captured it three days later. His point: the gap in how people understand AI is widening, and it is now, in his words, "a sharp funnel." If you judge AI by a free chatbot, you are judging the wrong tier.

The post builds on one he wrote in April, where he said two groups were "speaking past each other." Now he splits the world into four tiers. He calls the numbers "napkin math," so read them as rough estimates, not data.

The bottom: derpy chatbots

  • About 6 billion people (about 75%) have barely come in contact with LLMs at all.
  • About 1 to 2 billion (about 20%) are casual users of free tier ChatGPT style products. In his words, this group sees "derpy chatbots" and treats them a bit like a better Google search or a writing aid. He includes many professionals outside math and code here, such as executives, who understand the shift intellectually but experience it second hand.

The 0.2%: agents doing real work

  • About 20 million people (0.2%) use frontier grade LLMs professionally for math and code. In his April post he names the tools: OpenAI Codex and Claude Code.

This group sees first hand that large, complex projects that used to take weeks or months can now be done by agents with a prompt. He lists building apps, copying apps, translating apps and decompiling apps from binaries. He also notes that less than a year ago he was still writing code by hand, character by character.

The 5,000: frontier systems internally

  • About 5,000 people (about 0.00006%) have access to frontier grade systems inside the labs.

He describes what the outside world has seen of this tier as a preview: swarms of thousands of agents working together over weeks on software mega projects, finding zero days, running cyber attacks and defenses at machine speed, and doing new science and mathematics. He adds a warning: human review and comprehension are starting to fall behind.

The part you control

Karpathy says three factors drive the funnel:

  1. Ambition, problem size and horizon. "A question with a paragraph answer barely stresses the system." You need a supply of big, hard problems you actually care about.
  2. Jaggedness. Capability peaks in domains that are digital, verifiable and economically valuable, because models are trained with reinforcement learning on verifiable rewards. That is why math and code moved first.
  3. Access. Free tier, paid tier, internal.

You cannot buy your way into the internal tier. But the first factor is fully in your hands. If you only ask a model short questions, you will never see what it can do. Give an agent a real project with a long horizon and you see a different system. If you are new to the idea, I explain what that means in What Is an AI Agent.

My take

I agree with the shape of this, even if the numbers are rough. Most opinions about AI still come from a free chatbot used for a minute. The move that matters is the cheap one: take a task that would cost you a week, hand it to an agent like Claude Code or Codex, and see what happens. That is how you move from the 20% tier to the 0.2% tier.

Sources: Andrej Karpathy on X (October 11, 2026), Andrej Karpathy on X (April 9, 2026)