TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Andrej Karpathy says people’s experience of large language models is splitting across levels of access and intensity of use. His estimates—from billions with little exposure to roughly 5,000 with internal access to frontier systems—are informal, not independently verified population data.
AI researcher Andrej Karpathy says the public’s understanding of large language model capabilities is increasingly divided, with most people having little exposure while a small group of professionals uses frontier systems on complex coding and math tasks. His figures are explicitly described as “napkin math”, not survey results, but the post highlights how differing access and work patterns can produce sharply different impressions of what LLMs can do.
In a post on X, Karpathy divided users into rough tiers. He estimated that about 6 billion people, or 75% of the population, have barely encountered LLMs. Another 1 billion to 2 billion, he said, are casual or infrequent users of free ChatGPT-like products, who may treat them as a search aid or writing tool. These are his estimates; the post does not cite a dataset or explain a method for calculating them.
Karpathy placed around 20 million people, or 0.2%, in a group using frontier-grade models professionally in math and coding. He said some users in this tier are seeing agents complete large projects that previously took weeks or months. At the narrowest end, he estimated roughly 5,000 people have internal access to frontier systems and described experiments involving many agents working on large software, science, and mathematics tasks.
Those examples and claims about what internal systems can accomplish come from Karpathy’s account. The post does not identify specific systems, provide independently checked results, or establish that such capabilities are broadly available. He also argues that human review may be falling behind the volume of machine-generated work, citing continuing examination of an OpenAI-HF incident and mathematicians’ work on 722 frontier-mathematics manuscripts.
Why Uneven AI Access Matters
The difference Karpathy describes matters because judgments about AI often come from personal experience. A person who has only tried a free chatbot on short questions may reasonably see a limited assistant, while a developer using paid or internal tools on a sustained coding task may encounter a much more capable system. Treating either experience as representative can distort decisions about workplace adoption, training, and investment.
His account also raises a practical oversight issue: if agents can produce large amounts of work quickly, people may need time and expertise to check it. Karpathy’s warning that review is lagging is an interpretation, not a measured finding in the post. Still, the gap between reported system output and human capacity to evaluate it is relevant to organizations deploying AI in technical work.
The post does not establish that the most ambitious examples are reliable, repeatable, or economically useful across settings. Its value is as a description of uneven exposure and access, and as a prompt to ask which systems, users, and tasks are being discussed whenever capability claims are made.
As an affiliate, we earn on qualifying purchases.
Access Shapes the Experience
Karpathy’s argument builds on a post from April that he says he is revisiting. He describes the widening gap as more than two groups disagreeing: in his account, experience narrows through a funnel based on how frequently people use models, what kind of work they do, and which tier of system they can access.
He identifies three drivers. First, larger and longer-running tasks put more demands on a model than short questions. Second, he says capabilities are uneven across fields, with stronger results in digital tasks that can be checked and have economic value, particularly coding and math. Third, access differs among free users, paying customers, and people with internal access. These points are Karpathy’s explanation of the pattern, rather than independently established findings in the post.
As an affiliate, we earn on qualifying purchases.
How Much the Estimates Show
The post does not provide a survey, sampling method, or source for its estimates of exposure and access. The figures should be read as Karpathy’s rough approximations, not established measurements of the global population. It is also unclear how he defines “frontier-grade,” “professional use,” or access to internal systems.
Specific claims about agents completing large projects, working in swarms, or contributing to science and mathematics are not substantiated with examples or evaluation results in the post. The claims about cyber activity and newly discovered work are not independently verified there. The source also gives no new dated evidence showing how quickly the estimated tiers have changed since April.
As an affiliate, we earn on qualifying purchases.
Evidence Needed to Test the Claim
Karpathy’s post does not announce a study, product release, or scheduled follow-up. Testing the argument would require clearer measures of who has used which models, on what tasks, and with what results, alongside transparent definitions of access tiers. Independent evaluations could also clarify whether agent performance on long, complex projects is repeatable and how much human review those projects require.
For now, the post is best treated as a commentary on uneven experience, not a population-level finding. Further reporting or research would need to verify the estimates and examine whether the gap is widening over time.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main development?
Andrej Karpathy has renewed his argument that people’s understanding of LLM capability is diverging because exposure, use cases, and access to advanced systems vary sharply.
Are the user figures independently verified?
No. Karpathy calls them “napkin math.” The post does not provide a survey, methodology, or independent verification.
Who does Karpathy say is seeing the largest gains?
He points to a small group using frontier-grade systems professionally in math and coding, and a much smaller group with internal access. These are his estimates and descriptions, not independently confirmed counts.
Why might people have different impressions of AI?
Karpathy attributes the gap to differences in task size, uneven model performance across fields, and access to free, paid, or internal systems.
What remains uncertain?
The accuracy of the population estimates, the scope and reliability of the reported agent capabilities, and whether the gap has measurably widened over time remain unclear from the post.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
