I get the hallucination question a lot. At Dreamforce, I got it again. This time it was someone from a large GSI. They asked which models we use.
My answer was short. We have the context. Use a dumber model.
That answer sounds like discipline. It wasn’t. We didn’t start with the cheap model because we were smart. We started there because we were watching cost. The plan was to move up when we needed to. We haven’t needed to.
A cheap model can’t cover for you. Hand it a vague job, and you get a wrong answer, fast. We had to do the work first.
This is not only us. This is every team picking a model right now.
A model hallucinates when there is a gap. It doesn’t have the answer, so it writes something that sounds right. Every model does this. The big models guess better. That is why everyone wants one.
People are buying a better guesser. The other choice is to remove the guessing.
Removing the guessing is boring work. Find the data that belongs together. Decide what the model needs to read. Cut the rest. None of that needs a model. All of it used to be someone’s job.
By the time the model gets involved, the job left is small. Read this. Answer from it. A dumb model reads fine.
Give the model the smallest job left.
The big model made that work easy to skip. Paste everything in. Ask the question. Let the model sort out what matters. That is outsourcing the work. The model will take it. It won’t tell you where it guessed.
Budget hides this. A team watching tokens has to define the job. A team with a big budget never gets forced to. They get a hallucination and call it a model problem. Then they ask which model is safe.
There is a limit. Small models get worse on long, messy context. Some problems have no shape yet, and those need the big model. I know that from reading. I haven’t hit it.
I don’t have a test that proves the small model is enough. I have a design that leaves it very little to get wrong. Nobody has complained. That is what I have.
The person from the GSI wanted a model name. I gave them a list of work. They said that makes sense. Then they moved on.