An exact learning layer for the models you already use. Learn, forget, correct and undo in milliseconds, and always get the same model a full retrain would give you.
Most AI products put a trained layer on top of a foundation or embedding model: a classifier, a router, a ranking head. That layer has to be retrained every time something changes.
The Adaptive Plasticity Module replaces that layer. Instead of gradient descent, it keeps an exact closed-form solution that can be updated one example at a time. Add data, delete data, correct a label or undo a change, and the result matches retraining from scratch on the current data, to within 10−13.
Your foundation model stays exactly as it is. Nothing about how you embed data changes.
The Adaptive Plasticity Module works on the outputs of the embedding or foundation model you already use.
It solves for the best layer exactly, with no gradient descent. Under a second for 20,000+ examples on a GPU.
Add, forget, correct or undo in milliseconds. Always equal to a full retrain.
A new tool, intent or product goes live with a few examples. About 2 ms on a GPU, with no downtime.
Remove any data and the model is identical to one that never saw it. Retiring a tool with 219 examples took 5 ms.
Corrections fixed the target with a median of one example and changed only 0.16% of other predictions.
Every change is reversible, restoring the previous model to within 10−15.
Sites learn locally and merge exactly. A 5-site merge reached 93% accuracy without moving raw data.
About 2 ms per update on a GPU, and 30 to 300 ms on a basic 2-core CPU, against 71 s to retrain.
New tools go live instantly. A top-10 shortlist covers 93.6% of correct tools using just 1.1% of the catalogue's tokens.
→Add a new intent in 4 ms and keep accuracy on the existing ones, in every language you support.
→Honour deletion requests inside the model itself, with a delete you can verify.
→Learn from behaviour as it shifts, without nightly retraining jobs.
→Hospitals, banks or branches learn locally and merge exactly.
→Fix a bad prediction or roll back bad data, and nothing else moves.
→No. The Adaptive Plasticity Module sits on top of the model you already use and replaces only the final trained layer: the classifier, router or head.
In our benchmarks it is within about one point of a conventionally trained layer, and about ten points ahead of plain description search on tool routing. The difference is what happens after training.
After any sequence of additions, deletions and corrections, the model matches one retrained from scratch on the current data, to within floating-point precision (about 10−13).
No. A GPU brings updates to a couple of milliseconds, but it also runs on ordinary CPUs at tens to hundreds of milliseconds per update.
Nowhere. It runs inside your environment, in your cloud or on-prem.