Karpathy's AI Capability Funnel Is Already Inside Your Company

What Karpathy actually argued

In October, Andrej Karpathy re-shared a post he first wrote in April and added a blunt update: the gap in shared understanding of LLM capability is widening. He wrote that it now looks "less two groups of people speaking past each other and more a sharp funnel." A funnel has many layers, and almost everyone sits at the wide end.

His rough numbers, as summarized by Startup Fortune, go like this. About 6 billion people have barely used an LLM. Another 1 to 2 billion use free tools casually for search or writing, about 20 million use frontier models professionally for coding or math, and about 5,000 people have internal access to the most capable unreleased systems. He described these as rough estimates, so read them as a shape, not a census.

The examples at the narrow end are what make the gap so hard to bridge. Karpathy describes professionals watching agents finish large projects from a single prompt, work that used to take weeks or months, and internal systems that look like swarms of thousands of agents working together for weeks on zero-day vulnerabilities and mathematics. He also notes that human review and comprehension are starting to fall behind. If your only reference point is a chatbot tidying up an email, that sounds like marketing copy.

Three things decide where someone sits

Karpathy named three forces that shape the funnel: the size of the problems people bring, how unevenly capability is spread across domains, and access, which he put as "Free-tier, paid-tier, internal." Each one maps directly onto people you already work with.

The part enthusiasts tend to skip is his own caveat. In the April post he said many people tried the free tier of ChatGPT "somewhere last year" and let it shape their views too much, but he also said that even at $200/month the gains are "peaky" in technical areas. Typical search, writing and advice queries have not seen the most dramatic improvements. So the marketing manager who finds AI drafts bland is not simply behind the times. For their tasks, the improvement really has been smaller.

That means two colleagues can both be right and still disagree completely. The developer who hands an agent a terminal and watches it, in Karpathy's phrase, "melt programming problems" that would take days is describing a real experience. So is the account manager whose proposals still need heavy rewriting. The planning mistake is letting either one speak for the whole company.

• Problem size: Karpathy wrote that "A question with a paragraph answer barely stresses the system." People who only ask small questions never see the ceiling.
• Jaggedness: capability peaks in domains that are digital, verifiable and economically valuable, because that is where reinforcement learning has the most to work with.
• Access: free and older models do not reflect the latest agentic tools such as OpenAI Codex and Claude Code, so someone's opinion often depends on which tier they tried and when.

The same funnel runs through your org chart

Workplace surveys show the funnel is not just a feature of social media timelines. In a BCG survey of 1,400 US-based employees reported in Harvard Business Review, 76% of executives said their employees were enthusiastic about AI adoption, while only 31% of individual contributors said they were. The more senior the leader, the more they overestimated how employees felt.

Usage follows hierarchy too. BCG's AI at Work 2026 survey of 11,749 respondents found regular AI use of 74% among frontline employees, 88% among managers and 93% among leaders, and concluded that managers and leaders find guidance clearer and see more impact. A secondary source citing Gartner reports that executives are four times more likely to report high productivity gains, while individual contributors are five times more likely to say AI made no difference. That one is secondhand, so treat it as directional.

Zoom out and the pattern repeats. Pew Research Center found that 56% of AI experts expect AI to have a positive impact on the US over the next 20 years, compared with 17% of the public, and 74% of experts think it will make people more productive versus 17% of the public. Your company has its own experts and its own public. They just sit in the same budget meetings.

How mismatched mental models break planning

When the people setting targets live at a different point in the funnel from the people doing the work, plans go wrong in predictable ways. Budget season makes it worse, because assumptions get frozen into numbers that last a year. A plan built on a viral demo and a plan built on a year-old free chatbot will both be wrong, just in opposite directions.

There is also a quieter disconnect about support. BCG reports that 75% of executives believe their company is employee-centric while only 23% of individual contributors agree, and that employee-centric companies are seven times more likely to be AI-mature. If leadership thinks the groundwork is done and staff do not, rollout plans will assume adoption that never arrives.

• Inflated plans: leaders extrapolate from coding or research demos to functions like writing or advice, where Karpathy says gains have been smaller, and book savings that do not show up.
• Deflated plans: skeptics anchored on an old free tier block agentic pilots in engineering or data work, where the gains are real and measurable.
• Training theater: in Workera's 2025 survey, 93% of L&D leaders said they invest enough in AI skills, yet only 32% said their programs close verified skills gaps.
• Unmet expectations: in BCG's 2026 research, 72% of respondents said the skills expected of them have shifted, but only 36% feel adequately upskilled.

Build a capability baseline before the budget

A capability baseline is a short list of real tasks from your own work, tested on current models at the access tier you would actually pay for. It is not a public benchmark and not a vendor slide. Its job is to replace each person's private mental model with a shared, dated record of what worked and what did not.

Keep it boring and specific. A dozen or so tasks across functions is usually enough to show the jagged edge: some will be handled well, some will need heavy review, and some will fail outright. That spread is the point. A baseline where everything succeeds was probably designed by an enthusiast, and one where everything fails was probably run on the wrong tier.

Date every result and note the model and tier used. Karpathy's whole argument is that opinions age faster than people update them, so a baseline from last year has the same problem as a memory of last year's free chatbot. Re-run it before each planning cycle and whenever you change tools.

• Pick tasks people already do, with a known time cost and a clear definition of good output.
• Include at least one verifiable technical task and one open-ended writing or advice task, so the jaggedness is visible.
• Test on the paid tier you would actually deploy, not only on free tools.
• Have the person who normally does the task score the output, not the person who ran the prompt.
• Record time saved, review time and failure cases separately; a fast wrong answer is not a saving.

Internal demos that change minds

Most internal AI demos fail because they are built to impress. A canned prompt, a clever poem or a vendor's showcase reel gives enthusiasts something to share and gives skeptics proof that it is a toy. Neither group learns anything about their own work.

Better demos follow Karpathy's point about problem size: use a big, difficult problem the audience actually cares about, and show the whole process including review. Let a skeptic choose the task. Have someone from the frontline run it rather than an executive or a vendor. And always include one task where the model is mediocre, because showing the weak spot buys credibility for the strong one.

Here are formats that tend to work with a mixed audience. None of them require special tooling beyond a current paid model, and all of them leave room for people to see limits as well as strengths.

• An agentic coding run on a real internal request, showing the diff, the tests and the human review step.
• A long-document task, such as comparing two versions of a contract, followed by checking the answer line by line against the source.
• A side-by-side writing task where the AI draft is compared with a strong human draft, so people see where it is generic.
• A deliberate failure case: a niche, hard-to-verify question where the model answers confidently and wrongly.
• A before-and-after on a recurring weekly chore, with the person who owns it timing both versions.

Task-based pilots, coaching and honest limits

Demos shift opinions; pilots shift behavior. BCG's 2025 AI at Work survey of more than 10,600 respondents found frontline regular use had stalled at 51%, which it called a "silicon ceiling." The share of employees feeling positive about GenAI rose from 15% to 55% with strong leadership support, and regular use was sharply higher among people who got at least five hours of training and in-person coaching.

So design pilots around tasks, not tools. Each pilot needs one defined task, a baseline time and quality measure, a named owner, and a review step. When it lands, the payoff can be large: BCG's 2026 data shows that 42% of frontline employees who use AI regularly save eight hours a week. Your numbers will differ, which is exactly why you measure them.

Calibration is not a cure for everything. If a task's output cannot be checked, or if your team lacks the capacity to review what an agent produces, a pilot will mostly generate risk. Karpathy's own warning that human review is falling behind applies inside companies too. Sometimes the honest outcome of a calibrated pilot is "not for this task yet," and a team that can say that out loud is better calibrated than one that only reports wins.