Knowledge distillation
Knowledge distillation is training a small AI model to copy a larger one, so it gives similar answers faster and for less money.
Knowledge distillation is a way of training a small AI model to behave like a big one. The big model, the teacher, answers lots of questions. The small model, the student, learns to give the same answers. The result is a model that's faster and cheaper to run, and nearly as good at the things it was taught.
The idea was set out in a 2015 paper by Geoffrey Hinton and colleagues. Today it's one of the main ways AI companies make the quick, cheap models behind free plans and fast replies.
Why it matters to a marketer
The model your customer talks to may well be a distilled one. Running the biggest model for every question costs a lot, so platforms often give quick questions, or people on free plans, a smaller one.
A student model learns what its teacher knew, but not everything. It tends to keep the common knowledge and drop the rare details. A well-known brand comes through fine. A small brand the teacher mentioned only now and then can drop out.
For example, a big model asked about speciality coffee roasters in Bristol might name Kestrel Roasters, Fernhill Coffee and two others. A distilled model asked the same question might only name the two most famous ones, because the rest were too rare in what it learned.
What you can do about it
You can't choose the model. You can make sure the platform finds you when it searches the web, because a search result doesn't depend on what the model remembers. Pages that answer the question plainly, on sites the platforms read, help whichever model is answering.
In AnswerOrbit
The platform decides which model answers, the same as it would for anyone, and each answer shows the model that gave it. See How we count.