Karpathy: The LLM capability perception gap is widening into a sharp funnel

On October 12, Andrej Karpathy reposted and updated a piece from this April, offering an observation that drew widespread sharing: the public's gap in understanding LLM capabilities is not closing but widening, and its shape has shifted from "two groups talking past each other" to a sharply stratified "steep funnel." His current conclusion: genuine understanding of what LLMs can do is concentrated among a tiny group of frontline users, while the vast majority of the population remains outside the funnel.

Confirmed

The view comes from Karpathy's own repost and update on October 12 (m1, m3), citing his April tweets.
Stratified estimates: roughly 6 billion people (about 75% of the population) have essentially never used an LLM (figures vary slightly across m1, m3, m4, and m5; m2 records it as about 5 billion people / 75%).
About 1–2 billion people (roughly 20%) are light users of the free version of ChatGPT, treating it as a "better search" type tool.
A tiny fraction (m2 paraphrases as about 0.2%) use frontier models hands-on and genuinely understand their capability ceiling.

Not Yet Confirmed

All the figures for each tier are Karpathy's rough estimates with no cited data source, and the absolute headcounts relayed in different posts (5 billion vs. 6 billion) are inconsistent.

Why It Matters

Karpathy's judgment implies that societal understanding of AI capabilities is sharply diverging along "depth of use," which could affect product positioning, policy discussions, and public communication.
The post was heavily shared and re-relayed on October 12 (e.g., m4, m5), showing that this "cognition funnel" framing resonated with the community, making it an influential lens for understanding the current state of LLM adoption.

2026-10-12 ~ 2026-10-12 · 5 related posts

Primary sources

Karpathy: The LLM capability perception gap is widening—75% of people have barely touched one — karpathy · 2026-10-12 03:12 (UTC+08:00)
Karpathy: 75% of people have barely touched LLMs as the capability perception gap widens — karpathy · 2026-10-12 03:10 (UTC+08:00)

[source] Karpathy: 75% of people have barely touched LLMs as the capability perception gap widens — karpathy · 2026-10-12
[source] Karpathy: The LLM capability perception gap is widening—75% of people have barely touched one — karpathy · 2026-10-12
Karpathy: the gap in shared understanding of LLM capability is widening into a sharp funnel — deanwball · 2026-10-12

2 near-duplicate retellings: Anxious-Yoghurt-9207 · dejavucoder