The AI your team used this morning finished learning before any of you met it. It learned language, knowledge and reasoning in one long training run, from more writing than anyone could read. Then its training ended, it was released, and what it knows was fixed.
Since then, your team has taught it plenty. They have corrected its answers, walked it through the new returns policy and explained the exception for your biggest customer again and again. Each lesson lasted one conversation, and the model is exactly as it was on the day it was released.
Until now, every AI model’s learning has ended where its training ends. ACI begins there. It learns and acts at runtime, while your work happens, in tandem with the model you already use. That makes it a new category of intelligence, and Duplex is built on it.
A model learns the world in one long training run, before anyone uses it.
Every model stops learning when its training ends.
Models stop learning at release for a good reason. Teach a trained network something new, and it tends to overwrite what it already knew; researchers call it catastrophic forgetting. Hold it steady, and it keeps what it knows but stops learning. And because what it learned is spread through all of it, no single lesson can be corrected or removed without training it again.
Together, these make the trilemma of neural nets: a network can learn fast, keep what it knows or let you change what it learned, and it gets two at best. So every model is trained once, fixed and released. From then on, it gets better only when its maker trains the next version: on the maker’s schedule, for every customer at once, and from the world’s writing rather than your work.
The usual ways of teaching a model your work stay within that limit. A prompt, a document or a retrieved passage puts your latest correction in front of the model for one answer: the model reads it, uses it and learns nothing from it. Fine-tuning makes a lesson last, but each fine-tune is another training run on a cluster of accelerators, and the model it produces is fixed again on the day the run ends.
ACI learns and acts at runtime.
ACI starts with your first request. When your team confirms a result, the next request like it reflects what they decided. When a rule changes, ACI follows the new one from the next request. It needs no training run and no new version, and your model stays as it is.
Each result builds on the last, so the new doesn’t push out the old: a decision your team settled in March still applies in November, unless your team changes it. And your team can. Anything ACI has learned can be reviewed, corrected or removed on its own, and everything else stays as it was. A neural net gets two of the three at best; ACI holds all three, at runtime.
And it acts on what it learns. For each request, ACI decides the route: it answers what it has learned, brings in your model for what is new, briefed with what ACI knows about the case, or uses one of your tools, always within your rules. When the result comes back, ACI learns from that too.
A policy change puts all of this to the test. In ten runs, Duplex followed the new policy with 0% stale answers and 100% of its earlier work intact, while cache and retrieval went on serving the old answer 72% of the time.
It is intelligence, and it is yours.
ACI learns the pattern behind a result, and that is what makes it intelligence. When a request it knows comes back in new words, it recognises it. When a request only looks familiar, it tells the difference and hands it to your model. And when a piece of work keeps coming back, ACI acquires a skill of its own, proves it before use and removes it on request.
A cache or a search over past answers works differently. It finds what looks like the request, so a look-alike gets the old answer, and the same request in new words finds nothing close enough. That is useful for finding a document, and no substitute for knowing your work.
ACI learns your work: the decisions your team settles, the exceptions it makes, the way it does things. A model belongs to the lab that trained it. What ACI learns belongs to you, and for anything it has learned, you can see which agent it came from, with which model and when.
A new kind of intelligence, beside the one you have.
ACI works in tandem with the AI models you already use. The brain pairs two learners the same way: a slow one for general knowledge, like the neocortex, and a fast one for each new experience, like the hippocampus. Your AI model is the slow one: general intelligence, learned from the world. Duplex, built on ACI, is the fast one: runtime intelligence, learned from your work.
Each does what it does best. Your model brings breadth, and can take on almost any question. ACI brings depth: your rules, your decisions and your results, learned in detail. And adding an intelligence beside the model beats retraining the model itself. In six tests, on three public datasets and two model families, Duplex beat a fine-tuned adapter in all six, by 30.3 points on average, with no training run.
That pairing is what makes ACI a category of its own: a second intelligence, rather than a bigger model, a faster training run or a better place to keep documents. Years of research into the trilemma of neural nets became ACI: Vareon’s proprietary algorithms and architecture for an intelligence that learns from every result, builds on what it knows and stays in your hands.
Every model you use will still stop learning on the day it is released, and so will the next one. What changes is everything after that day. From your first request, ACI learns from your work, and every day it runs, the intelligence beside your model knows that work a little better. Training ends where it always has, and the learning goes on for as long as your work does.