The AI safety conversation is obsessed with the ceiling: the smartest model, the frontier, the hypothetical superintelligence. I think we are watching the wrong number. The change that will actually reshape the world this decade is not that the best model got smarter. It is that a good-enough model got almost free.
The number that moved
Capable models now cost cents per million tokens. Open, permissively licensed models with near-frontier reasoning can be downloaded and run for the price of the electricity. Two years ago, doing something at scale with AI meant a serious budget. Today it means a rounding error. Smart got a little smarter. Cheap fell off a cliff. The second one matters more.
Why cheap is the real multiplier
A genius model locked behind a premium price is a tool for well-funded players doing careful things. A merely-good model that costs nothing is a tool for absolutely everyone, doing absolutely anything, at absolutely any scale. Spam, scams, disinformation, surveillance, and manipulation were never bottlenecked by intelligence. They were bottlenecked by cost. Remove the cost and you do not need genius. You need good-enough, running a million times.
The scary future is not one superintelligence. It is ten thousand mediocre models doing cheap harm at a scale no human process could match.
The part I am not comfortable with
I am a believer in open weights. I have argued in favor of them repeatedly, because they break vendor lock-in, protect privacy, and keep power from pooling in three companies. And openness plus near-zero cost is exactly what makes the misuse cheap too. I do not get to celebrate one and pretend the other is someone else problem. The same freedom that lets a small team build something wonderful lets a bad actor build something ugly for the same price.
So what
I am not arguing to lock it all up. That cure is worse, and it just hands the future to whoever holds the keys. I am arguing that we are pointing our worry in the wrong direction. The interesting risk is not the model that is too smart to control. It is the model that is cheap enough that control stops being economically possible. We should spend less energy imagining the genius and more on the world where competent AI is as cheap and ubiquitous as electricity, because that world is not a forecast. It is this year invoice.