Karpathy reshared his April breakdown of the LLM capability gap and it's been stuck in my head since.
His funnel: ~6 billion people have barely touched an LLM. 1 to 2 billion use it casually. Around 20 million actually push it on math and code. Fewer than 5000 have internal access at the labs.
That 20 million is the whole game, not the 6 billion. Every open-weight model, every API play, every agent framework is fighting for the attention of that thin professional slice.
[观点] Man, the part that gets me is how wrong the "everyone uses AI now" narrative is. If you're wiring open-weight models into production you're in a smaller club than the headlines suggest. Kinda good news actually, the bar to matter is lower than it looks.
[观点] And it reframes the open-weight debate. From the outside the Qwen vs closed fight looks like a mass-market war. For the people actually building it's a scramble for maybe 20 million minds, the rest is noise.
So where do you sit: do you think the real user base is way smaller than the hype says, or are we still early and that 20 million is about to 10x?
Top comments (1)
we still early and that 20 million is about to 10x