“Dusting off this tweet from April because this gap in shared understanding of LLM capability is widening. It’s now less ‘two groups of people speaking past each other’ and more a sharp funnel.”
“Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.”
“The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about.”
“The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical.”
The company where I am integrating AI sits in the top part of Karpathy’s 20%. It is not frontier. Everyone there has the same licence, but not everyone uses it the same way.
The core idea is that we have four kinds of people in non-tech companies. We cannot expect everyone to go and take advantage of defining agents and skills and automating tasks or workflows, because there are certain types of people that are not able to do it. But the opportunity lies on closing the technical debt.
Context, rules and judgment
A general bot like Grok Bot or Claude Projects is quite superior to anything that I can build by myself. The product that a small team can develop will never be close to what SpaceXSI or Claude can develop. But you need to provide the context, rules and instructions to have this applied in the real world of a company.
Not all people are prepared to transmit the process and judgment. And perhaps judgment is the most complicated part. Not all people have the will to adopt it, either.
Four kinds of people
What I have seen is that we have four kinds of people in a non-tech corporation.
We have the super adopters: people that think in processes and can define what a process is, the boundaries, the rules that govern it. Super adopters are the exception. Then you have adopters, for simple processes and simpler automation. And then you have a certain kind of people that are never going to use it, the non-adopters.
For non-adopters, you have two kinds: the AI deniers, and people that simply follow the rules: they sit and operate, but they cannot define rules or processes.
Generally speaking, a manager should be able to do it, but in real life not all managers can. Or even if they are not managers, they think like an engineer, in processes, and can define systems. It is a process and order mind, in which you order your work, not being a manager. Many people at companies don’t have this mind and they just follow the rules. So having them create agents or skills and establish processes is too difficult.
Two types of integration
When you create simple automations, which is the main part where AI integration can generate value, at least in the first step, general productivity, I’m not talking of an autonomous agent. I’m talking about just automating a workflow.
I think these non-adopters are perhaps big numbers in non-tech companies, and the opportunity lies there: to create the software, to close the tech debt.
We have two types of integration. One is to integrate and provide the super adopters and the adopters the tools to create automations and workflows, to get certain repetitive tasks done by agents, and that’s great. On the other side, there is the reduction in the price of software creation. As I mentioned in one of my articles, the opportunity also lies in closing the software debt that any company has.
For the non-adopters, you create software like a glove, it fits your needs, not like a wood shoe that you must adapt to it. It’s like creating interfaces. Generally speaking, people need to adapt to the ERP: you have a process, but you need to adapt to the ERP interfaces, and the interfaces are not quite good. It is always work to do all the transactions in the ERP, and for some people it takes a lot of time.
Closing the technical debt is also closing this interface with the ERP, and creating new software that can eliminate transactions. That doesn’t mean these rule followers are able to develop the process itself, because for most people it is quite complicated.
The non-adopters will benefit only when the software debt is closing, so they have leverage in software, but not in AI directly.
Funnel. Andrej Karpathy, post on X, 11 October 2026, quoting his post of 9 April 2026: x.com/karpathy/status/2109361546505966046. Tiers and figures are his.
Prepared by Mariano Funes with Claude (Anthropic).