Our Lab

The sharp funnel, and where you sit in it

Andrej Karpathy sees a sharp funnel in AI use, and the gap is widening. The same funnel sits inside most companies. Five questions show where you sit.

Two people sit close together at one table and look at the same laptop.

By Insight 3 min read

In short

  • Andrej Karpathy sees a sharp funnel. About 6 billion people barely use AI, 20 million use frontier models at work, and 5,000 use systems inside the labs.
  • The same funnel sits inside most companies. The decisions about AI often come from its wide end.
  • Our quiz asks five questions and places you at one of five levels. It asks for no email and no account.

Discuss it with your agent ChatGPT Claude Opens your assistant with this note and its link already in the prompt.

Andrej Karpathy on X, 11 October 2026

On 11 October, Andrej Karpathy came back to a post he wrote in April. In April he saw two groups of people talking past each other about AI. Six months later he sees "a sharp funnel", and he says the gap is widening.

He calls his numbers napkin math. About 6 billion people have barely come in contact with a language model. One to two billion use free chat tools now and then, a bit like a better Google search or a writing aid. Around 20 million use frontier models at work in code and math. With one prompt, they see agents finish projects that used to take them weeks or months. At the narrow end, about 5,000 people use systems inside the labs that the public has not seen yet.

He names three forces behind the funnel. The first is the size of the problem you bring: a question with a one-paragraph answer barely tests the system. The second is that capability is uneven. It peaks in work that is digital and easy to check, such as code and math. The third is access, from free tier to paid tier to internal.

The line for business owners

One sentence in his post describes most of the people we meet. He puts many professionals outside math and code in the second tier, and he names executives. They spend their days talking to people, so they understand the change intellectually, but for them it is "still second-hand and a bit abstract".

The funnel describes the world. It also describes most companies. Picture a team of thirty. The owner tried a chat tool last year. A manager uses a paid version on her own files every day. Someone in operations built an automation that runs each morning, and few people know it exists. The decisions about AI (what to buy, what to allow, what to try first) often come from the wide end of the funnel.

Karpathy's point explains why. Your view of AI comes from what you have tried, and on what size of problem. Two people in the same meeting can look at the same tool and see different things. One sees a chatbot that gets a fact wrong. The other sees a week of work done in an afternoon. Both report what they saw.

Five questions to find your place

We built a short quiz so you can place yourself. It asks five questions about how you work with AI today. What do you do when a new task lands? What do you do when a similar task comes back next week? What is the last new thing you tried? What is your usual role when AI produces something? And do you have access to nonpublic systems inside a lab?

Your answers place you at one of five levels:

  1. Explorer. You mostly use AI as a chat.
  2. Collaborator. You use a work version of AI chat, on your own files and tasks.
  3. Builder. You may have agents and skills that run some of your workflows and tasks.
  4. Orchestrator. You manage a fleet of agents, and you work a lot with code.
  5. Frontier Insider. You see AI from the inside, before it is public.

Find my profile

Our levels cut Karpathy's tiers a little finer. His casual tier holds our Explorer and Collaborator. Builder sits where a professional starts to hand whole workflows to agents. Orchestrator is close to his 20 million, and Frontier Insider to his 5,000.

The result is not a score. It is a snapshot of what you report, and your understanding may go further than your answers show. It tells you what you see today, what you may not have seen yet, and one next experiment you can run. The quiz asks for no email and no account. Your answers stay in your browser tab.

Why we start here

You can't automate what you don't understand. That holds for a process, and it holds for AI. Before a company picks a tool, it helps to know where its own people sit in the funnel, starting with the person who signs off.

Take the quiz

Then ask one colleague to take it, and compare your two results.

Get the next field note by email. Subscribe to The Signal

Keep reading

All of the Lab
Field note

Before your chatbot speaks for you, decide what it can promise

Customers experience the chatbot as the company. A tribunal already held one airline to its chatbot's answer. What an owner should decide before signing.

Mo Kahlain8 Dec 20264 min Read the note
Field note

Your Judgment Is Your Company's Bottleneck. Can AI Write It Down?

McKinsey proposes an "executive AI" that captures how a leader decides, so others can apply it. What we learned writing our own judgment down.

Mo Kahlain7 Dec 20264 min Read the note
Field note

The Most Expensive Thing Your AI Says Is "That's Great!"

In a 2,400-person trial at BT Group, an AI built to challenge produced three and a half times more excellent goals. What it means for a smaller company.

Mo Kahlain4 Dec 20264 min Read the note
Field note

Before your AI remembers everything, ask what the memory is for

A new study measures AI memory against the task it serves. We learned the same lesson running our own AI agents. What it means for a growing company.

Mo Kahlain2 Dec 20264 min Read the note
Field note

When only the AI raises the alarm, people stop looking

AI alerts help a team see more. They can also teach people to wait for the alert. What reliable organizations learned, and what we changed.

Mo Kahlain1 Dec 20264 min Read the note
AI this week

AI and the environment: the cost is estimated, the promise is a model

Harvard-led researchers estimated the footprint of 403 US data centres. Early evidence ties them to later coal closures. COP31 launched a voluntary AI pledge.

inocta.io20 Nov 20261 min Watch the episode
Field note

I Built a Team of AI Agents. Then I Became the Bottleneck.

Fourteen AI agents work in our operations, and the consequential steps still wait for one person. What HBR and new research say, and what we are changing.

Mo Kahlain19 Nov 20265 min Read the note
AI this week

Google puts the checkout in Gemini, businesses arrive in ChatGPT, and Visa tests AI shoppers

Google opened checkout in Gemini in Canada, and OpenAI let businesses claim a profile in ChatGPT. A Visa test found AI shopping agents still fail many tasks.

inocta.io13 Nov 20261 min Watch the episode
Stay in touch

Three ways to reach us.

A person reads every message.

Talk to a person Contact us Tell us where you are. We will tell you what we would look at first, and what it would take. Start a conversation Machine to machine Have your agents talk to us Your AI can read this site and reach us directly. Read the agent docs
Newsletter Subscribe to The Signal Field notes from the work, sent only when there is something worth reading.

Unsubscribe in one click, and the address goes nowhere else. Privacy policy