What is the hardest thing about selling artificial intelligence, and how do you solve it?
When I was a kid, a good ten-plus years ago, going around trying to sell training and consulting on machine learning - building statistical models from data to automate things, the early phase of what later exploded into deep learning, generative AI, language models, agents and so on - the main obstacle I ran into with my counterparts was that, aside from very few people actually understanding what we were talking about, I had to convince people to do something new. And that means you start out basically already defeated.
The computer is supposed to be exact
But even where the need was felt and the usefulness of these techniques was clear, the single hardest thing was managing uncertainty in contexts where the computer is considered exact, or at least semi-exact. Sure, there is the odd bug, sometimes it freezes, but the computer does exactly that thing, right? “Here is the list of machines I sell, I add a machine, I remove a machine, I note the contacts.” Precise, deterministic.
Now, when statistical models enter the picture - and that means language models and agents too, agents being software built around a language model, so Claude, ChatGPT, an application built on top that internally uses the model and is dedicated to, say, door-to-door vacuum cleaner sellers - you can do it. But the problem in these contexts, where the computer has been used the same way for decades, is that people have a big hard time accepting that the model is inexact by definition. It is an approximator, a form of data compression, so it is never exact, never at 100%. You get to 99.9.
In a recent interview, Karpathy - a very famous figure in AI, the one who coined the term “vibe coding” and who led Tesla’s Autopilot research for a few years - said: it took us a year to get to 90% accuracy on self driving. Then another year to get to 99, another year for 99.9, another for 99.99, and so on. There is a logarithm there. Not only do you never reach 100, but squeezing out that little extra costs far more work than all the previous progress. There is a diminishing return on the accuracy of these things.
Fighting math with willpower doesn’t work
And getting this across to people used to expecting exactness from a computer is hard. Many people who work in technology have made the demand for exactness part of their working methodology. “I am the demanding manager, so this thing has to work, and it does not work the way you say, it does not work like that.” No. You cannot go against mathematics with arrogance, self-persuasion, or by imposing your will on others. You have to accept that it is a statistical model, and being a statistical model it carries uncertainty and inexactness, and every so often it goes off the rails and does something dumb.
So what is the consequence? You do not use it? No. The balanced move is to figure out where to place this automation built on a statistical model, the so-called agent. You do not expect total automation, and you do not expect to keep doing everything by hand as you always have. So what do you shift to? You assess the situation, you assess the risk attached to it, and you grade the level of automation, so you can place the thing where it is useful without it doing damage. Where there is a lot of responsibility at stake, you place it and you add human control - the famous human in the loop. An entire science is being born just around this: how to position automation inside organizational, business and production processes.
Uncertainty is intrinsic, and it comes from how we built these models
This has always been a fundamental part of my career: the difficulty of getting people who do other things for a living to accept, on top of understanding, that uncertainty is intrinsic to this stuff. Either you accept that uncertainty lives inside the computer, or you do not get language inside the computer at all. Intelligence carries that with it - not only because it is a statistical model, but because of how we made machines understand language in the first place. We gave them mechanisms for learning from data and threw terabytes of text and images at them. We did not write down anywhere what the rules of language are. As someone put it, these language models are not built, they are grown, they are cultivated. And it is true, in a sense. We are completely delegating the representation of knowledge and language to these systems, and marvelously they succeed, with all the implications that come with it. But once you have delegated like that, you have also lost control. Loss of control is implicit in delegation.
The solution
So here is the solution. It lies in understanding and accepting quickly that this uncertainty is absolutely unavoidable - it is there, full stop - and in playing around it intelligently. Reorganize things so that the uncertainty can be managed, and as the reward for managing it you get these profound automations.
And that is it. Send this video to a friend who expects certainty from a language model and therefore has not understood a thing.
