How one big model becomes many small and personal ones - without anyone starting from zero.
Nobody weaves their own cloth to buy a jacket. You buy off the rack and the tailor takes it in: a hem here, a shoulder there. Fine-tuning is exactly that - the ready model is the off-the-rack suit, your data is the measurement, and the result is a model that fits your work. Hours instead of months, thousands instead of billions.
Distillation is the other craft in the same atelier. The big model is the master: knows everything, but is expensive and slow. The small one is the apprentice: it watches how the master answers and learns to reach the same answers with far fewer means. That is why your phone can run a model that behaves surprisingly close to the big ones - it is a distillate, not a copy.
The fine print
A fine-tuned model inherits the flaws of its base - the weak spots arrive together with the knowledge. And distillation has an edge being argued over in disputes and lawsuits: if the apprentice learns from the answers of someone else's model, whose is the result? Some licenses forbid it explicitly. The technique is simple; the law around it is not.
Our read: these two words explain why the AI market is not just a handful of giants. On top of every big model grow thousands of small ones, tailored to one job each. The giants sell the cloth. The tailor shops are everything else.