Ambition
The value of AI grows with the size and time horizon of the problem you bring. In Karpathy’s words, a question with a paragraph answer “barely stresses the system”.
SeparatesConsumers from professionals
Understanding AI · White paper
The Widening AI Gap
Two colleagues can use AI in the same week and come away with opposite views of what AI can do. In October 2026, Andrej Karpathy, a founding member of OpenAI and the former head of AI at Tesla, described this gap as a funnel and said the gap is growing. This paper draws his funnel, sets out the three factors he says decide where a person sits in it, and argues that two of those factors are skills you can build.
Guess first
How rare is the view from the narrow end?
Out of every 500 people in the world, how many have seen an AI agent finish weeks of professional work from a single prompt?
Figure 1
Four groups see four different versions of AI
Karpathy sorts the world into four groups by how they have used large language models, the technology behind tools such as ChatGPT and Claude. The figures are his own estimates, which he calls “napkin math”, not the results of a survey.
-
1. Barely touched it
about 6 billion · about 75%These people have had little or no contact with language models.
-
2. Casual users of free chatbots
1 to 2 billion · about 20%These people use a free chatbot now and then, as a better search engine or a writing aid, and see what Karpathy calls a “derpy chatbot”. Many professionals outside math and code sit here, including executives, who spend much of their day talking to people and so understand AI second-hand.
-
3. Professionals using frontier models for math and code
about 20 million · about 0.2%These people use the most capable models at work and see agents complete, from a prompt, projects that used to take weeks or months: building apps, copying apps and translating apps from one language to another. Karpathy notes that less than a year earlier he was still typing code by hand.
-
4. Inside the AI labs
about 5,000 · about 0.00006%These people have access to systems the public has not yet seen, with thousands of agents working together for weeks on large projects in software, security and mathematics. Karpathy writes that human review is starting to fall behind the work these systems produce.
The bars use a logarithmic scale, so all four groups fit on the page. Switch to a linear scale to see the true proportions.
Figure 2
Three factors decide which version of AI you see
Karpathy names three factors that push a person further down the funnel. Each factor runs from a low end to a high end.
Jaggedness
AI capability is uneven. Karpathy says capability peaks in work that is digital, verifiable and economically valuable, because models are trained with rewards on tasks whose answers a computer can check.
SeparatesOne field from another, with math and code first
Access
The free tier, the paid tier and the systems inside the labs are different products. In our experience, the free tier is the least capable of the three.
SeparatesWhat is available to you at all
The term “jagged frontier” comes from a 2023 study in which consultants at Boston Consulting Group did better with AI on some tasks and worse on others that looked similar in difficulty (Dell’Acqua and others, 2023). The Jagged Frontier Quiz lets you test your own sense of where the edge sits.
The argument
Why this is a skills gap, not a tool gap
Access is the only one of the three factors that money buys directly. An individual paid plan has cost about US$20 a month since OpenAI launched ChatGPT Plus in 2023 (OpenAI, 2023).
The other two factors depend on what you attempt. Choosing a problem large enough to test the tool is ambition. Knowing which parts of your work are digital and can be checked is judgment about jaggedness. Both can be learned, and neither comes with the subscription. This conclusion is our reading of Karpathy’s post, not a claim he makes.
The gap matters at work because decisions about AI budgets, hiring and policy are usually made by senior people. If Karpathy is right that many executives sit in the casual tier, those decisions are being made from the wide end of the funnel about a technology that looks different further down. That is our inference from his post.
Karpathy calls the present moment strange: the general public has mostly not used these systems, most professionals see a modest uplift, and a small group is watching capability climb steeply, all at the same time.
Where this bites
Four places the gap shows up at work
The examples below are our own illustrations of the funnel in an organization, not cases from Karpathy’s post.
Budgets
A leadership team that has seen only a free chatbot can underfund AI work, because the version of AI the team remembers could not do much.
Estimates
A developer who works with agents and a manager who does not can estimate the same project at one week and three months. Both estimates are honest, and the plan needs to say which one it assumes.
Vendor demos
A vendor demo is usually run on a paid model, on a task chosen to suit the model. A trial on your own data, by your own staff, is the better guide to what you will get.
Policy
A policy written with only chatbots in mind can miss agents, which take actions as well as producing text. The Agent Autonomy Ladder sets out the levels to plan for.
Place yourself
Where do you sit in the funnel?
Answer the three questions below for the way you use AI today. The page will place you in the funnel and name the factor holding you back.
The placement is our reading of Karpathy’s three factors. He does not score individuals. The fourth tier needs access that only lab staff have, so the exercise stops at the third. Your answers stay in your browser.
How to read the numbers
What this paper does and does not show
- The four group sizes come from one expert’s post on social media. Karpathy calls the sizes napkin math, and no survey stands behind them.
- Karpathy works near the narrow end of the funnel, so his view of the lower tiers is first-hand and his figures for the upper tiers are estimates.
- The funnel describes large language models only, not AI in general. Many organizations already use other kinds of AI, such as fraud detection and forecasting.
- The argument that the gap is mostly a skills gap, and the placement exercise, are ours.
Use this week
Move one step down the funnel
Take the problem you wrote down above and give the problem to an AI tool this week, with the context a new colleague would need. Check the result before you rely on it.
Choosing problems and checking output are two parts of thinking well in a digital workplace. The Digital Intelligence course builds those skills one day at a time.
Sources
Sources
- Karpathy, A. (2026). Post on X about the widening gap in understanding of LLM capability, last edited 11 October 2026. x.com/karpathy/status/2109361546505966046. The post extends an earlier post from 9 April 2026. Discussion on Reddit r/singularity.
- Dell’Acqua, F., McFowland, E., Mollick, E. and others (2023). Navigating the jagged technological frontier. Harvard Business School Working Paper 24-013.
- OpenAI (2023). Introducing ChatGPT Plus. 1 February 2023.