The Honest Questions

It starts with a switch that is either off or on, and ends with something that can be given a job and left to do it. There is no gap anywhere along that line, no step where something inexplicable had to be added.

Knowing all of it settles less than you might expect.

Some of what remains open is technical, and the field is honest about it. Nobody fully knows why scale produces the abilities it does, how to read what is stored inside the weights, where the ceiling is, or how to keep systems aligned as they become more capable. Serious researchers hold flatly opposing views on several of these, and the disagreement is not laziness.

The rest is not technical at all, and that half is the harder one.

A system being able to do something tells you nothing about whether it should, who gets to decide, or who pays when it fails. Those come apart cleanly once you separate them: what the system can do, how many steps it takes unsupervised, what it is connected to, and who answers for the result. Only the first is a fact about the model. The other three are chosen by whoever deploys it, and the frightening scenarios almost all need someone to hand over a key.

Behind that sits a harder asymmetry. Training a frontier model costs more than almost any organisation on earth can raise, which means a very small number of groups decide what gets built, what it refuses, who may use it, and which problems are worth pointing it at next. Effort spent on one problem is not spent on another, and the money flows toward what can be sold.

None of this requires bad intent. It requires only ordinary incentives, working as designed.