Using AI /
Why people disagree about whether AI works
I read a comparison recently that named something I had been circling for a while. To the people using these systems to get real work done every day, the people who insist online that AI still cannot write code, that it hallucinates everything, that none of it is real, sound like someone trying to convince you the car in your driveway does not exist. You say, but I drive it to work. I paid for it. I put fuel in it. I could not hold down a job twenty miles away without it. And they tell you that you are imagining the car, or that you only say it runs because you work for the car company. You are both standing in the same street, describing different worlds.
Both sides are telling the truth
Here is the part worth sitting with. Almost nobody in that argument is lying.
The person who says AI is useless is usually reporting a real experience. They opened a free chat window, typed one line, got back something confident and wrong, and closed the tab. That happened. It is a fair account of what they saw.
The person who says it does half their work is also reporting a real experience. They pay for the best model, they have learned how to brief it, and they hand it things it is genuinely good at. That happened too.
Two true stories. Two different machines. The mistake on both sides is the same one, assuming the thing in your hands is the thing in everyone else's.
It is not the same tool
"AI" is one word doing far too much work. It covers a free model you poke at through a browser, and a frontier model wired into your files and paid for by the month. Those are not the same product at different prices. They are different machines.
The distance between them is wider than most people guess:
- The free tier gives you a capable but limited model. The top plan gives you the frontier ones, Opus and Fable and whatever each lab ships as its best.
- A one-line prompt gives you a stranger's guess. A page of real context gives you something that has read your actual situation.
- A cold chat window forgets you the moment you leave. A model joined up to your work remembers what it is looking at.
Someone judging AI from the free tab, one line at a time, and someone running the best model over their own context all day are not disagreeing about the same thing. They have not used the same thing.
The believers did the boring work
This is the part I keep coming back to, because it is most of what I write about here. When AI works well for someone, faith has nothing to do with it. The result sits on a stack of unglamorous decisions.
- They pay for the frontier model instead of judging the whole category by its cheapest version.
- They give it context, the same move as building a digital brain and cleaning the data, so the model is not guessing about their world.
- They have learned what to ask and how, which is a skill, not a setting.
- They use it for what it is good at and route around the rest, instead of trying to catch it out.
None of that is magic. All of it is available to the person who currently thinks the whole thing is a con. The car is real. They have just never been handed the keys to a good one with fuel in the tank.
Why the gap keeps widening
The uncomfortable bit is that this is not evening out. It is spreading.
The people already using these tools well compound. Every week they get a little faster, fold in another workflow, learn another thing the model can carry. The people who decided in a single bad session that it does not work are standing still, certain, and falling further behind the very thing they are certain about. The two realities are not drifting back together. They are pulling apart.
You do not have to take a side on faith. You can just run the test properly once. Pay for a good model for a month, give it real context, hand it something you actually need done, and see which street you are standing in.
The rule
Before you decide AI does not work, check which machine you were holding.
A free model, one line, no context, is a fair test of nothing. A good model, a real brief, your actual context, is the test. Run that one, then tell me whether the car exists.