A good new hire learns all the time. In their second week someone mentions that one client is billed quarterly, and that client never gets a monthly invoice again. By the end of the year they know hundreds of small things like it, and learning the latest has cost them none of the others.
Ask the same of an AI, and what every good colleague does becomes a research problem. Your model learned the world in one long training run, and then what it knew was fixed. To teach it something new you train it again, and training on the new tends to wear away the old. So teams correct the same answers week after week, or wait for the next version and hope.
Duplex learns the way that new hire does: from every result, while the work happens, and each new lesson costs it none of the old ones. It also adapts as the work changes, which is harder still.
Continual learning means taking in the new without losing the old.
The new and the old pull apart.
Your work keeps moving long after your model was trained. A fix lands on Monday, a rule changes on Wednesday, and every day your team confirms results the model has never seen. Learning from each of them as it happens is hard, because it asks for things that pull in opposite directions.
Learn eagerly, and the new pushes out the old. Researchers call it catastrophic forgetting, and it is why a model trained on this month’s corrections can start getting last month’s answers wrong. Hold on tightly instead, and what it knows stays put while your work moves on, so the answers go stale. And whatever it learns has to stay open to your team, so that a wrong lesson can be corrected or removed without disturbing the rest.
Together these make the trilemma of neural nets: plasticity, stability and editability. A neural net gets two at best. Learning in production asks for more than any two: a system has to follow a change, keep its earlier work, take a correction, stay fast as it grows and know what it has not learned, all at once and on live work. Duplex meets all five, on every run.
Retraining makes every fix wait.
Most of what is sold as continual learning is continual retraining. The system gathers what changed, retrains the model, tests the new version and swaps it in. Then it waits for the next run.
Between runs, the model stays as it was. The fix your team made on Monday waits for the run, and until then the same mistake goes out to every customer who asks. The answers are only ever as current as the last run, and your work doesn’t wait for anyone’s schedule.
Each run is a project in itself: a training pipeline, a cluster of accelerators and a team to run them. And every run makes a new model, which can lose some of what the last one knew.
Online fine-tuning, which retrains as it goes, runs the same risk. After a policy change, it learned the change and kept Almost none of its earlier work. Duplex followed the same change with 100% of its earlier work intact.*
Your team fixes a wrong answer on Monday.
Duplex learns, decides and acts at runtime.
Duplex learns from each result as your team confirms it: a test passed, a review approved, a correction accepted. The next request already reflects it, and Duplex decides and acts on what it learned, within your rules: it answers, hands over or declines. There is no training run to wait for, no new version to test and no model to swap in.
It learns beside your AI, rather than inside it. Your model’s weights stay as they are, and Duplex needs no GPU. You can review, correct or remove anything it has learned, and everything else stays as it was. That third need, editability, is the one retraining finds hardest, because in a retrained model no one can point to a single lesson and change it.
Duplex is built on ACI, Adaptive Continual Intelligence: Vareon’s proprietary algorithms and architecture for intelligence that learns, decides and acts continually at runtime. ACI is the result of years of research into that trilemma, and it is what lets Duplex take in the new without giving up the old.
Learning at runtime also keeps the answers current. After a policy change, 0% of Duplex’s answers were stale. Cache and retrieval, still serving what they had found before the change, gave stale answers 72% of the time.*
Adapting is harder still.
Learning a fix is half the work. The other half is adapting: knowing when what was learned applies, when it does not, and what to do when the work changes. A system that applies every lesson everywhere is as unhelpful as one that learns nothing, and more dangerous, because it sounds sure of itself.
Three moments show it. The same request arrives in new words, and Duplex knows it and answers. A request looks familiar but is new, like a refund on a gift card, and Duplex tells it apart and brings your AI in, briefed with what it knows about the case. A rule changes, and Duplex follows the new rule from the next request, while the rest of what it knows carries on unchanged.
On a stream of 3,000 decisions in which the rules changed five times, Duplex was right on 92.5% of the decisions those changes affected. The model on its own, with no way to learn the new rules, was right on 2.3%.*
Duplex learns from every confirmed result, and builds on it.
Every result makes the next one better.
A model gets better when its maker trains the next one. That happens on the lab’s schedule, for everyone at once, and the new model arrives knowing nothing about your work.
We believe Agents will get better with a second intelligence beside them, in a recursive loop, at runtime. It is why we built Duplex. An agent with Duplex beside it gets better every day it works, from its own results: it decides and acts, the result comes back, Duplex learns from it, and the next decision starts from more than the last one did. Better decisions bring better results, and better results teach Duplex more.
The models will keep improving, and you will want the best of them. Your work will keep changing too, every hour of every day. With Duplex beside your AI, learning is no longer something that happens to a model now and then. It becomes part of the work itself: every result teaches Duplex something, each lesson builds on the last, and all of it comes with you to the next model.