The Gap In Shared Understanding Of LLM Capability Is Widening

TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Andrej Karpathy says public understanding of large language models is splitting as access and use vary sharply across groups. His approximate figures—from billions of people with little exposure to about 20 million professional users of frontier systems in math and coding—are estimates, not independently verified counts.

Andrej Karpathy, an AI researcher and former OpenAI and Tesla executive, says people’s understanding of large language model capabilities is diverging as most users encounter basic chatbot functions while a much smaller group uses advanced systems for complex math and coding work. In a post on X, he offered rough estimates of how many people fall into each group, framing them as “napkin math” rather than verified population data.

Karpathy estimates that about 6 billion people, or roughly 75% of the world’s population, have barely encountered LLMs. He places another 1 billion to 2 billion people—about 20%—in a casual-use group, often using free chatbot products as a search aid or writing tool. These figures are his estimates; the post does not describe a survey or provide a method for calculating them.

At the specialized end, Karpathy estimates that about 20 million people, around 0.2% of the population, use frontier-grade models professionally in math and coding. He says those users can see agents complete large projects that previously took weeks or months. He further estimates that roughly 5,000 people have internal access to frontier systems and may see more advanced applications. His descriptions of agent swarms, cyber operations, scientific discovery and mathematics are characterizations of capabilities and activity, not independently substantiated findings in the post.

Karpathy attributes the widening gap to three factors: the scale and difficulty of the task, models’ uneven performance across domains, and differences in access to free, paid or internal systems. He argues that current systems are stronger in digital fields such as coding and math, where results can be checked, than in many other areas.

Why LLM Experience Varies So Sharply

The divide matters because people may form very different judgments about AI from experiences that are both genuine but not comparable. A casual user trying a free chatbot for short answers may see a fallible assistant, while a specialist using a more capable model on a large, testable coding task may see substantial productivity gains. Neither experience, by itself, represents the full range of model performance.

That difference can shape workplace decisions, public debate and expectations about how quickly AI will affect jobs and industries. Karpathy’s argument also points to a verification challenge: if systems produce large volumes of complex work, people may struggle to review it at the same pace. The post raises that concern, but does not quantify how often human review is falling behind or establish the scale of any resulting harm.

Three Tiers of Model Exposure

Karpathy describes a funnel, rather than simply two groups disagreeing. At its broad end are people with little or no direct exposure; next are casual users of free products; further down are professionals using frontier models for math and coding; and at the narrowest end are people with internal access to advanced systems. The categories are his framework, and the population estimates are approximate.

He says access is only part of the explanation. He also emphasizes what he calls the “jaggedness” of model capability: performance can be strong on some tasks and weak on others. In his account, digital work with verifiable outcomes and commercial value is especially suited to improvement through training and reinforcement. This helps explain why coding-focused professionals could report gains that users in other jobs do not experience.

How Reliable Are the Estimates?

The post does not identify a data source, survey, sampling method or date for its estimates of how many people use LLMs at each level. The percentages and head counts should be treated as Karpathy’s rough assessment, not established global measurements. It is also unclear how he defines “frontier-grade,” professional use, or internal access, and how much overlap exists between the groups.

Several examples of advanced work are presented without supporting documentation in the post. It does not specify which systems completed the projects, what human direction or review was involved, or how performance was evaluated. The claim that human review and comprehension are falling behind is likewise not quantified there. These limits make the post useful as an account of differing user experiences, but insufficient on its own to establish the prevalence or consequences of those capabilities.

Evidence Needed to Track the Gap

Assessing whether this divide is growing would require dated, transparent data on who uses LLMs, which products and model tiers they access, and what tasks they perform. Comparisons would also need to distinguish reported productivity gains from independently measured outcomes, particularly for complex coding and scientific work.

For now, Karpathy’s post is a practitioner’s analysis rather than a formal study or announcement of a new model. The question remains open: as access changes and capabilities shift across fields, will everyday users encounter the kinds of gains specialist users report, and can organizations review AI-generated work reliably at scale?

Key Questions

What is Karpathy’s main point?

He argues that people’s understanding of LLM capability is diverging because most people have limited or casual exposure, while a small group uses advanced models for complex professional tasks.

Are the figures in the post verified statistics?

No. Karpathy calls them “napkin math” and does not cite a survey, dataset or calculation method. They should be read as rough estimates.

Why do users report such different experiences?

Karpathy points to differences in access, task difficulty and model performance by domain. A casual chatbot query is not comparable to using an advanced system on a large, verifiable coding task.

Does the post prove that AI agents can replace months of work?

No. Karpathy describes projects he says agents can complete, but the post does not provide project records, independent evaluations or details about human involvement. Those examples are claims in his account, not proof of general performance.

Source: hn