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 argues that people’s understanding of large language models is diverging as most have little exposure, while a small number of technical professionals use frontier systems for complex work. His population estimates and examples are his own rough assessment, not independently verified measurements.
Computer scientist Andrej Karpathy says the gap in how people understand large language models is widening, as most people have little direct exposure while a small group of technical professionals use frontier systems on complex projects. In a post on X, he offered rough estimates of those groups and argued that differences in access, task type and model performance help explain why accounts of AI capability can seem far apart.
Karpathy estimates that about 6 billion people, or 75% of the global population, have barely encountered LLMs. He places another 1 billion to 2 billion people—about 20%—in a casual-use tier, typically using free chatbot products for search-like questions or writing help. These are explicitly his “napkin math” estimates; the post does not cite a survey or a method for calculating them.
At the other end, Karpathy estimates about 20 million people, or 0.2%, use frontier-grade LLMs professionally in mathematics and coding. He says these users can see agents complete large projects that previously took weeks or months. He describes a still smaller group, about 5,000 people with internal access to frontier systems, as seeing systems work across extended software projects and other demanding tasks.
Karpathy also cautions that exposure is not the only difference. LLM performance is “jagged,” he argues: capability is stronger in some areas, particularly digital tasks with verifiable outcomes and commercial value, than in others. His post characterizes agent-led cyber activity, scientific discovery and advances in mathematics as examples of what a small group may observe. Those examples and the scale of the systems’ performance are his claims, not independently substantiated in the post.
Why AI Experience Varies So Sharply
The argument offers a reason that public impressions of AI can differ without either side necessarily describing the same experience. A person who has tried a free chatbot for occasional questions is testing a different product, task and level of access from an engineer using a frontier model on a large codebase. Use case and access shape what people can observe.
That divide matters to decisions about adopting AI at work and evaluating its risks. If advanced systems are useful in a limited set of high-value, testable tasks, broad claims that they are either generally capable or generally unimpressive can obscure where their effects are concentrated. Karpathy’s post also raises a separate issue: if systems can produce more work than people can readily review, human oversight and comprehension may become constraints. The post frames this as a concern, but does not provide measurements of review capacity or document resulting harms.
As an affiliate, we earn on qualifying purchases.
Four Tiers of LLM Exposure
Karpathy organizes his account as a funnel: people with little or no exposure; casual users of free products; professionals applying frontier models in mathematics and coding; and a small group with internal access. The tiers describe differing levels of contact, not a formal study or a mutually exclusive classification. His percentages are approximate and the post does not specify a data source or date for the population estimates.
He says he had been writing code manually less than a year before the post, contrasting that experience with recent agent-assisted work. The post also refers to the OpenAI-HF incident and 722 manuscripts on frontier mathematics as examples of work that may take people time to understand. It does not explain those references in detail, establish their connection to the wider population estimates or independently verify the claims about them.
professional large language model access
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Estimates Without a Measurement Base
The post provides no survey, dataset or calculation behind its estimates of 6 billion people with minimal exposure, 1 billion to 2 billion casual users, 20 million professional users or roughly 5,000 people with internal access. The categories may overlap, and the post does not define “frontier-grade,” “professional use” or how access was counted. These figures should be treated as Karpathy’s approximate estimates, not established global totals.
It is also unclear how representative his examples are of users and organizations beyond his own technical perspective. The post does not provide independent evaluations of the claimed agent performance, details of the cited incidents or manuscripts, or a measure of whether human review is falling behind. The scale and consequences of that concern remain unquantified.
As an affiliate, we earn on qualifying purchases.
Evidence Needed to Test the Funnel
The post does not announce a study, product release or scheduled follow-up. Testing its central claim would require clearer definitions of LLM exposure and access, repeated surveys across regions and professions, and independent evaluations comparing model performance across task types. Tracking how often AI-generated work is checked, corrected or accepted would also help assess the concern about human review.
Until such evidence is available, the funnel is best read as Karpathy’s account of diverging experiences, rather than a population-level measurement. Further reporting would need to establish whether the groups and proportions are accurate and whether the gap is changing over time.
As an affiliate, we earn on qualifying purchases.
Key Questions
Who says the gap in LLM understanding is widening?
Andrej Karpathy made the argument in a post on X. He based it on differences between people with little exposure, casual chatbot users and professionals using advanced systems.
Are the population figures independently verified?
No. Karpathy calls them “napkin math” estimates, and the post gives no cited survey or methodology. They should not be treated as measured global totals.
Why might casual users see different results from technical professionals?
They may have access to different systems and use them for different tasks. Karpathy argues that LLM performance varies by domain, with stronger results in some digital, verifiable work such as coding and mathematics.
What does Karpathy mean by uneven or “jagged” capability?
He means that performance is not consistent across tasks. A system may handle some technical work well while producing less useful results elsewhere; his post does not offer a formal measurement of that variation.
What remains unknown about the claimed capability gap?
The post does not establish how many people fall into each group, whether its examples are representative, or how often advanced AI work is reviewed by people. Those claims need independent data to assess their scope.
Source: hn
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
