the cost of trying

a model doesn’t have to become more intelligent to become more useful. sometimes the same answer, at a lower cost, is enough to make a new use practical.

running a model is called inference. every answer uses computation, whether the model is writing a sentence or spending several minutes working through a problem.

when that computation gets cheaper, two things become possible:

1. more attempts. a coding tool can try several approaches, run the tests, and keep the one that works. this helps when the result can be checked. asking five models whether an answer sounds right is not the same as testing it.

2. smaller jobs. a task that saves someone a minute might not justify an expensive model call. at a lower price, checking a broken link, sorting a receipt, or explaining a compiler error can become worth doing automatically.

there are several ways to reduce the cost. better hardware can do more work for the same energy. better software can use that hardware more efficiently. a smaller model can be trained for a narrower job. none of these requires every model to improve at everything.

price is only one constraint, though. an inexpensive answer that needs constant correction can cost more in attention than it saves in money. a fast answer is not useful if it arrives before the model has enough information to be right.

this is why i care about verification. a tool that can check its work can make use of cheaper attempts. a tool that can’t check its work may just produce mistakes faster.

the question isn’t only how good the next model will be. it’s which useful tasks become affordable with the models we already have, and how we will know when they have actually finished them.