Most teams reach for fine-tuning when the model keeps getting the same things wrong. It files billing questions under technical support, writes in a tone your brand would never use, and misses the exceptions your experts apply without thinking. What they want is modest: a model that answers as if it had worked there for a year.
The way to get it is a training project. Someone collects and cleans examples, runs a sweep of training jobs to find the settings that work, checks the result and ships a new version of the model. It works, and it is a project, with a budget and a schedule. Then a policy changes, or a better model comes out, and the project starts again.
Duplex gives you what fine-tuning is for, from your team’s own work, and leaves the model exactly as it is. Every result your team confirms teaches it, so the adaptation is current from the next request. And because it works beside the model rather than inside it, the adaptation moves with you to the next one.
A fine-tune rewrites the weights, one model and one training run at a time.
Better than fine-tuning, with no training run.
Fine-tuning earns its place. It changes the model’s weights, so the new behaviour comes with every call, with no long prompt to send along, and for a lasting change to one model it is a sound choice.
The cost is in everything around the change. Each adaptation is a project of its own, and what the model learned is fixed at the moment the run ends. A rule that changes on Friday waits for the next run, and until then the model goes on applying the old one.
What teams fine-tune for is usually behaviour: the labels they use, the tone they want, the decisions their experts make the same way every time. Duplex learns exactly that from the work itself: a reply your team approved, a label someone corrected, an exception a manager confirmed. The person who spots a mistake corrects it, and the fix applies from the next request, instead of waiting in a folder of examples for the next training project.
Duplex is built on ACI, Adaptive Continual Intelligence, which learns and acts at runtime and leaves your model’s weights alone. Duplex needs no GPU, so there is no training run to plan, pay for or wait on, and no new version of the model to ship.
Set against fine-tuning directly, Duplex came out ahead. In six tests, each against a fine-tuned adapter picked from a sweep of training runs, it adapted the same model, left exactly as it was, and beat the adapter every time, by 30.3 points on average.
Six tests: three public intent datasets, each run on two families of open models.
It moves when your model does.
Your AI is general intelligence, learned slowly, like the neocortex: it learned the world before you arrived, and your work doesn’t change it. Duplex is runtime intelligence, fast, like the hippocampus: it learns from each result your team confirms. Fine-tuning asks the slow learner to do the fast one’s job, and that has two costs teams tend to find out later.
The first is correction. Every example in a fine-tune is blended into the weights, the good with the bad, so when one turns out to be wrong, say a reply that quoted last year’s discount, there is nothing to point at and take out. The way back is to train again, without it.
The second is moving. A fine-tune belongs to the model it was trained on, so when a better model arrives, it stays behind and the project starts over: the sweep is run again, the result checked again, and until then the new model knows nothing of your work.
What Duplex learns works beside the model, so neither cost applies. You can see where each lesson came from, and correct or remove it on its own, while everything else stays as it was. Change the model, and the adaptation comes with you: when the work moved to another model family, Duplex needed no retraining, while the fine-tuned adapters moved in none of six tests.
One example your model learned from turns out to be wrong, and has to come out.
When you do want to train, Duplex hands over the data.
Some jobs do call for training: a model that needs a capability it doesn’t have, or weights that must run on their own, on a device or inside a product you ship. For those, Duplex makes the training better.
The hardest part of any training project is the data: clean examples, checked by people who know the work. Your team has been confirming exactly those all along, in the ordinary course of the day. Duplex prepares them for your training team, each with where it came from, and you choose what goes in and what stays out.
Training doesn’t replace what Duplex learned, either. When the new model is ready, all of it works beside it from the first request, every result after that goes on teaching Duplex, and when the next model arrives, the adaptation moves again.
Adaptation stops being a project with a budget and a schedule, and becomes part of the work. Your team corrects, confirms and decides, as it does every day, and your AI answers as if it had been trained on all of it. The model stays a choice you can change at any time, and the adaptation becomes something you own.