The people building the future and the people living in it aren't even using the same technology yet.
The Summary
- Andrej Karpathy estimates 75% of the world's population has "barely come in contact" with large language models, six months after he first flagged a widening AI literacy gap
- The OpenAI cofounder now describes the adoption pattern as a "sharp funnel" rather than two groups talking past each other
- Among those who do use LLMs, most still treat them like glorified search engines instead of reasoning tools or collaborative agents
The Signal
OpenAI cofounder Andrej Karpathy is watching the AI revolution happen in real time and noticing something uncomfortable: it's not actually a revolution for most people. In April, he flagged a growing gap in how people understood LLM capabilities. Six months later, that gap didn't close. It turned into a chasm.
His updated math puts roughly 6 billion people—75% of the global population—at "barely come in contact" with LLMs. Not "don't understand them deeply." Not "use them wrong." Barely touched them at all. The general public's exposure, when it exists, looks like what Karpathy calls a "derpy chatbot." A cute toy. A thing that gives weird answers sometimes.
"The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot."
But here's where it gets interesting. The remaining 25% isn't a monolith either. Karpathy describes a funnel:
- 75% haven't really used LLMs
- Most of the remaining 25% treat them like fancy Google
- A smaller slice uses them for actual productivity
- An even smaller group builds with them
The people at the top of that funnel—the ones doing "vibe coding," automating workflows, treating Claude or ChatGPT as thinking partners—live in a different reality than the people at the bottom. The gap Karpathy identified in April wasn't two groups speaking past each other. It was a preview of stratification. Now it's structural.
This matters because we're not talking about early adoption of smartphones or social media. Those tools changed *how* you did things. LLMs change *what work is*. If you're using them right, you're already operating at a different speed, with different leverage, solving different problems. If you're not using them at all, you're not just behind—you're solving yesterday's problems with yesterday's tools while the game board rebuilds itself.
The Implication
The funnel Karpathy describes isn't going to flatten on its own. Adoption gaps this wide don't close through osmosis. They close through economic pressure or deliberate infrastructure. If you're building for the agent economy, your market isn't "everyone." It's the people already at the top of the funnel, and the next cohort trying to climb it.
For individuals, the message is simpler: the gap is widening, not shrinking. Treating LLMs like search engines means you're still at the bottom of the funnel with 75% of the world. The people winning are the ones who figured out these tools think, not just retrieve. Start there.