Retraining a generative AI model’s weights with specialized examples until it “speaks” like the right kind of organization or answers questions focused on a particular domain.
"The model was pretty good, but it kept returning examples in US English, so we fine-tuned it to use UK English instead."
In the beginning, the default method for customization was fine-tuning. This meant retraining the model’s weights with specialized examples until it “spoke” like your domain or organization. While powerful, it ended up being costly, brittle, and static—more like building a machine that only knows one trick than teaching a student who can keep learning and adapting. In other words, once fine-tuned, the model is stuck with what you gave it, and updating it means starting over. Imagine taking a generalist AI, and training it only on sock-related info—eventually, the overabundance of sock info will displace the more generalized information.
Today, we’ve mostly moved away from fine-tuning an entire model because it locks you into a static version of a fast-moving field. Instead, lighter, more dynamic methods are usually better. Still, fine-tuning can be the right choice in scenarios where compliance requires outputs in a very specific format, where internet access isn’t possible, or where latency must be minimized by baking knowledge directly into the model.
