Artificial Intelligence
An Overview

What we are actually talking about

Nearly everything called AI today comes from one idea repeated at an enormous scale. A system is shown a very large amount of text and asked to guess the next word. When it guesses wrong, it is corrected. Then it does that again, roughly a trillion times.

Nobody was certain that would produce much. It produced most of this. To guess the next word well, a system has to pick up spelling, then grammar, then facts, then the shape of an argument, and eventually something that behaves like a working knowledge of the world. None of it was written in by hand. All of it was learned from being wrong.

That single fact explains both halves of what you see. These systems are startlingly good at anything close to what they were trained on, and they fail in ways no careful person would, because nothing tells them when they have wandered off the map.

The rest of this page is a consequence of that.

Three words that get used as if they were one

A model is the trained thing itself. You give it words, it gives you words back. On its own it has no memory of yesterday, no access to your files, and no way to act on anything it says.

A reasoning model is the same thing allowed to work before it answers. Instead of replying with its first thought, it writes out a long private attempt, catches some of its own mistakes, and answers from that. It costs more and it takes longer. On anything with a right answer, it is worth both.

An agent is a model put in a loop with tools. It can read a file, run a command, look at what came back, and decide what to do next. That last part is the whole difference. A model answers a question; an agent takes a run at a job, and can be wrong several times along the way and still finish it.

Most arguments about AI are really three arguments at once, because these three things have different limits, different costs, and different ways of going wrong.

What happens when a first attempt costs nothing

The change that matters is not that machines can write and draw. It is that the first version of almost anything is now close to free.

That sounds like a story about speed, and it is, but it changes creative work more than it changes output. When a draft is expensive, you commit to your first idea and defend it. When a draft costs a minute, you can hold five of them side by side and throw four away. Judgement becomes the scarce part, not production.

It also moves where the effort goes. Saying precisely what you want turns out to be most of the job, and it is the same job whether the thing being described is an essay, a photograph, or a piece of software. The people getting the most out of these tools are rarely the fastest typists. They are the ones who know what good looks like and can say why.

The floor rises for everyone. The ceiling does not move on its own.

Why it still matters who is deciding

Alignment is the plain question of whether a system does what the person actually meant, including the parts they did not think to say out loud. It is unsolved, and not for lack of effort. We are simply not very good at writing down what we want.

The everyday failure is smaller than the ones in the headlines, and far more common. A model states something false in exactly the same confident tone it uses for everything else. It has no sense of the edge of its own knowledge, so nothing in the answer tells you which sentence to go and check.

Which is why trust has to be built into the product rather than claimed in the marketing. Put the source next to the claim. Make it plain what the system looked at and what it did not. Keep a person answerable for anything that matters, and show them enough that they can be.

Care here is not caution slowing things down. It is the thing that makes the output usable by anyone who has something to lose.

What the last few years already tell us

Forecasts in this field age badly, so this chapter only repeats what has already happened more than once.

Capability arrives before anyone knows what to do with it. There is usually a year or two between a system being able to do something and the first product that does it well, and most of the value in that gap goes to people who were already paying attention.

Cost falls faster than almost anyone plans for. Work that is too expensive to hand over this year is routinely cheap the next, which makes not economical yet a weak reason to stop looking.

And the hardest part stays human. Deciding what is worth building, what good enough means, and what should not be built at all are not jobs a better model takes away. They matter more as everything around them gets cheaper.

The useful posture is neither excitement nor dread. It is attention: use the tools on real work, notice exactly where they break, and change your mind when they stop breaking there.