The gap in shared understanding of LLM capability is widening
Andrej Karpathy described a widening gap between casual users’ experience of LLMs and the capabilities seen by technical professionals with frontier systems.
TL;DR
- Karpathy said people’s understanding of LLM capabilities is separating into a “sharp funnel” shaped by unequal access and use.
- He offered rough estimates for the population tiers, including about 20 million professional users and about 5,000 with internal frontier-system access.
- The figures are his informal estimates, not measured population data or an independent capability assessment.
Andrej Karpathy, an AI researcher, wrote on X on October 11 that the gap in shared understanding of large language model capability is widening. He described it as a “sharp funnel,” moving beyond two groups simply talking past each other. [1] [2]
Karpathy’s post gives rough tiers: about 6 billion people with little contact with LLMs, 1–2 billion casual users, around 20 million professionals using frontier systems for mathematics and software work, and approximately 5,000 people with internal access to the most capable systems. He labels the estimates “napkin math.” [1] [2]
He says these groups encounter different kinds of systems and report different levels of capability, from basic chatbot use to agents completing extended projects. His examples of internal systems working on large software projects, cybersecurity and science are his account; the post does not provide independent measurements or disclose a particular model release. [1] [2]
The Hacker News discussion shows the post attracted substantial attention, but reactions do not verify its estimates. How many people fall into each tier, and whether access differences produce the claimed gap in understanding, remain unmeasured in the cited material. [1] [2]
Why it matters
The post frames uneven access as a reason public and professional debates about AI capability may diverge, while leaving the population estimates and frontier claims unverified.
Editor's note
This is a high-engagement commentary, not a study or product announcement. Karpathy’s numbers and frontier-system examples are explicitly treated as his estimates and account.