If your business only uses AI to search or write the odd email, it’s normal not to notice anything; AI for business starts to show when you give it big problems. Andrej Karpathy, an co-founder and former director of AI at Tesla with more than 4 million followers on X, describes a growing gap between people who use AI that way and people who apply it to real work. This article covers what he said, why many small businesses stay on the side that sees few results, and three steps that help them cross over.
What Karpathy said about the AI gap
On 9 April, Karpathy wrote on X that two groups talk about AI without understanding each other. One judges it by the free version of they tried a while ago. The other pays for the most advanced models and uses them every day in technical work, where the improvement has been huge. In his view, progress concentrates on tasks whose result can be checked with a yes or a no.
On 11 October he came back to the topic and said the gap is widening, to the point that it looks more like a narrow funnel than a simple split. He added some figures of his own: about 6 billion people have barely used a language model, 1 to 2 billion use free versions, about 20 million use the most capable models and around 5,000 people have access to what the labs have not released yet.
These figures are not official data: Karpathy presents them as napkin math, with no source, to sketch the shape of the funnel.
Why your business sits in the part that notices little
According to , more than 1.2 billion people use every week (GPT-6 announcement). In a typical small business, AI usually means someone opening the free chat to draft an email or summarise a document. It helps, but the business runs the same way as before.
On X, Chubby (@kimmonismus) said he had seen the same on a trip to Paris: «When I talk to people, I realize they’re completely unaware of how capable AI already is. They’re unaware of the possibilities it offers». His conclusion: «Currently, we’re living in an ivory tower».
Karpathy names three factors behind the funnel. The first is ambition: a question with a paragraph-long answer barely tests the AI, and you need a supply of big, difficult problems you really care about. The second is that AI progresses unevenly: it shines in work that is digital, has a result you can check and has economic value, and moves much more slowly elsewhere. The third is access: free, paid or internal.
AI for business: three steps to make it count
1. Pick a problem that really matters. Writing text faster is fine, but it rarely changes a profit and loss account. Look for the work that eats the most hours or where a mistake costs money: quotes that take days, customer emails left unanswered, orders copied by hand from one system to another.
2. Start with tasks you can check. AI improves faster, and is easier to trust, when you can tell whether it got it right. Does the order match the delivery note? Does the invoice have the right amount? Is the figure in the right field? These tasks have a clear answer, which makes them the best place to start. Work that depends on taste or judgement can wait.
3. Move from the free version to an agent connected to your business. The free chat does not know your catalogue, your customers or your prices, and it cannot act in your software. AI agents can: they read your data, follow your rules and complete tasks with permission, with a person reviewing what matters.
What this means for a small business
Three illustrative examples: an online shop can have an agent that answers about order status with real data. An accounting firm can check incoming invoices against its books, which matters more with mandatory B2B e-invoicing in Spain. A consultancy can prepare the first draft of a quote with its own rates.
Karpathy closed his 11 October post with these words: «And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.»
Where to start
For a first diagnosis, the next step is the €150 one-off review. We look at how your company uses AI today, which tasks can be checked and which agent would make sense to connect first. If you’d rather talk it through, contact us.