A demo lasts an afternoon. Production lasts years, and in that time an AI that learns will meet thousands of results, a policy that changes on a Tuesday, a new product line, a rule tightened after an audit and, sooner or later, someone trying to talk it into something it shouldn’t do.
Most systems that learn need a minder for all of that. Someone retrains them when they fall behind, clears them out when they grow too large and checks after every change that nothing important was lost. The minding never ends, and when the minder is busy, the learning is what gets switched off.
Duplex is built to be left running. It goes on learning from every result, applies your rules ahead of anything it has learned, blocks attacks and can remove poisoned data. Learning in production rests on five requirements that have to hold at the same time, and of the three approaches tested, only Duplex met all five.
Cache and retrieval average 2.2 / 5: they go stale and keep growing.
Five requirements, all at once.
The five are plain enough: learn new work, lose none of the old, follow a change, stay the same size and know when to ask. Any one of them is easy to show in a demo. The hard part is all five together, because each tends to cost another: learn faster and a system forgets, hold it steady and it stops learning.
A system that misses one can look fine for weeks. Then it starts to forget, or to give last quarter’s answer, or to grow until it is slow and costly, or to answer questions it was never taught. These failures don’t announce themselves. Each one turns up later, as a customer who got the old policy or a bill that crept up.
The usual approaches each get part of the way. Cache and retrieval take on new work easily and lose nothing of the old, which is what they are for, but they serve the old answer after a change and grow with everything they are given. Online fine-tuning follows a change, stays the same size and knows when to ask, and it forgets almost all of its earlier work in the process. Across the five, cache and retrieval averaged 2.2 / 5 and online fine-tuning 4 / 5.
Duplex is built on ACI, Adaptive Continual Intelligence, the result of years of research into why a neural net forgets when it learns and stops learning once it is steady. It met all five requirements on every one of ten runs, with the same supervision and the same inputs as every other approach. That is what separates a system you can demo from one you can leave running: no gaps across the five, every time.
Current after every change.
Rules change all the time. A refund limit moves, an approval step is added, a supplier is dropped. A system left running has to follow the new rule at once and leave everything else it learned as it was. With Duplex, your rules come first, ahead of anything it has learned, so a rule your team changes at noon applies from the next request.
Some changes are ones nobody asked for. People will try to talk a system past your rules, rephrasing the same attack until something gets through. Against six families of paraphrased attacks, mixed with harmless requests written to resemble them, no attack succeeded, and Duplex refused only 0.7% of the legitimate requests, the lowest of any approach that blocked every attack.
Bad data is the quieter threat: examples fed in on purpose to teach the wrong lesson. After deliberate poisoning, Duplex removed all 40 poisoned samples, and its accuracy came back to within a point. Everything legitimate it had learned was left as it was, checked bit for bit.
The changes you do want, it follows in full. After a policy change, Duplex gave 0% stale answers and lost none of its earlier work. Cache and retrieval went on serving the old policy, 72% stale. Online fine-tuning learned the new policy and kept Almost none of the earlier work.
A policy changes: a learner has to follow it without losing anything it learned before.
It stays the same size, and plans its moves.
Size is the slow failure. A system that grows with everything it sees gets slower, costlier and harder to check, a little more every month. Memory approaches work this way: they add what happens to a growing pile, which makes the past easy to look up and the pile hard to live with.
Duplex doesn’t grow as it learns, while cache and retrieval, in the same runs, grew to about 1,300 rows. Duplex took on the new work and followed the change at the same size, so there is nothing to prune and nothing to clear out to keep it quick.
The same size doesn’t mean knowing less. What Duplex learns about a world is enough to plan in it. A reactive agent takes whatever step looks best right now, which works until the way that looks direct leads nowhere: in a multi-step world over ten runs, it reached the goal 0% of the time. Duplex found the exact shortest path on every run, without calling a model. A frontier model also reached the goal, by spending tokens on every episode.
Your team still decides what Duplex learns, and can review, correct or remove any of it. What goes away is the minding: the retraining, the pruning and the anxious check after every change. Duplex learns from every result for as long as it runs, and the longer you leave it running, the more of your work it knows.
A small world with a goal, and a way that looks direct but leads nowhere.