TL;DR
Andrej Karpathy argues that people’s experience of large language models is splitting into a sharper funnel, from little or no use to access to frontier systems. His population figures are rough estimates, not independently verified measurements, and his account of what advanced systems can do reflects his assessment.
Andrej Karpathy, an AI researcher, says the gap in people’s understanding of large language model capabilities is widening, as most users encounter basic chatbots while a small group works with frontier systems on complex tasks. In a post on X, he gave rough estimates of how many people fall into each category, but those figures are his napkin calculations rather than verified population data.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Karpathy estimates that about 6 billion people, or roughly 75% of the global population, have barely encountered LLMs. He places another 1 billion to 2 billion people in a casual-use tier, typically using free chatbot products for search-like questions or writing help. He says some users reasonably find little practical value in them.
At the more intensive end, Karpathy estimates that about 20 million people use frontier-grade models professionally in mathematics and coding. He says agents can now complete some large software projects that previously took weeks or months. He contrasts those users with an estimated 5,000 people who have internal access to frontier systems and, in his account, are seeing more extensive multi-agent work.
Karpathy describes applications including software development, mathematical research, and cybersecurity. These are his characterization of the capabilities and work taking place; the post does not provide independent evidence or detailed case studies for each example. He also warns that human review and understanding may lag behind systems’ output, pointing to work still underway to examine past incidents and large collections of mathematical manuscripts.
Why Access Shapes AI Experience
The divide Karpathy describes matters because public discussion of AI can draw on very different firsthand experiences. A casual user testing a free chatbot may judge the technology by short answers or routine writing tasks, while a specialist using advanced tools on a lengthy coding project may see a much larger change in what can be done.
That uneven exposure can complicate decisions by employers, educators, policymakers and the public. People may overestimate or underestimate what current systems can reliably do when they generalize from tools they have tried—or from reports about systems they have not used. Karpathy’s account also raises a practical concern: if agents can generate work faster than people can inspect it, organizations may need stronger review processes. The post does not establish how widespread that problem is.
Products Worth Considering
Jackery Solar Generator 1000 v2 with Two 100W Solar Panels, 1024Wh LiFePO4 Battery
DJI Mini 5 Pro Fly More Combo with DJI RC 2
Jackery Explorer 1000 v2 Portable Power Station
As an affiliate, we earn on qualifying purchases.
Three Forces Behind the Funnel
Karpathy attributes the disparity to ambition, uneven capability and access. Simple questions do not test systems as heavily as large, difficult projects, he argues, so users’ experience depends partly on the size and duration of the task. A chatbot’s performance on a short answer may not reflect how it performs when given a substantial coding or research goal.
He also describes LLM abilities as “jagged”: stronger in some domains than others. In his view, systems are especially capable in areas such as mathematics and coding, where work is digital and results can be checked. He links that pattern to reinforcement learning with verifiable rewards and to commercial incentives, but the post does not present data that independently tests that explanation.
Finally, access differs between free products, paid services and internal frontier systems. Karpathy says he was still writing code manually less than a year before the post, illustrating how quickly his own work practices changed. The post refers to an April tweet but does not specify its year, so the timeline cannot be pinned down from the source alone.
How Much of the Gap Is Measured?
The population estimates are explicitly rough and are not accompanied by a method, survey or supporting dataset. It is unclear how Karpathy defines having “barely” encountered an LLM, casual use, professional use or access to an internal frontier system. The figures should not be treated as a census of AI adoption.
The post also does not identify specific systems, model versions, evaluation results or examples that would allow readers to check its claims about agents completing complex projects, cybersecurity work or mathematical research. Nor does it quantify how often generated work requires correction or how much human review it needs. The account presents Karpathy’s assessment, not a systematic comparison of users’ understanding.
Evidence Needed to Test the Divide
To establish whether the gap is widening, future reporting or research would need comparable measures over time: who has access to which systems, how people use them, and how well they understand their capabilities and limits. Independent evaluations of agent performance on sustained tasks could help separate demonstrated results from impressions or promotional claims.
For now, Karpathy’s post offers a framework for understanding why people can reach sharply different conclusions about LLMs: they may be using different products, pursuing different tasks and seeing different levels of access. The size and pace of that divide remain unverified by the estimates in the post.
Key Questions
What is Karpathy’s main claim?
He argues that people’s firsthand experience of LLMs is increasingly uneven, ranging from little exposure to work with advanced systems on large projects.
Are the user figures verified?
No. Karpathy describes them as rough “napkin math.” The post gives no methodology or independent data to confirm the estimates.
Who does he say uses frontier LLMs professionally?
He estimates about 20 million people use frontier-grade systems professionally in mathematics and coding, and about 5,000 have internal access to frontier systems. Both numbers are his estimates.
Why might people have different views of AI capability?
Karpathy points to differences in task complexity, uneven performance across domains and levels of access. A person using a free chatbot for brief questions may see different capabilities from a specialist using advanced systems on extended work.
Source: hn
