Test-time-training models with a transactional safety harness, and a study of what they can learn lastingly from experience. plastic studies models that learn while they read: their memory is a set of fast weights that take gradient steps on a self-supervised objective as tokens arrive, so context is compressed into weights instead of a growing cache. Three substrates share - PlasticCore, a from-scratch model of about 6M parameters: a selective state-space recurrence plus a delta-rule fast-weight memory, meta-trained end to end through its inner loop. The same block stack serves next-token prediction on text and a hidden-friction control task where the only way to predict well is to infer…
Independent publisher
David Montgomery
dmontgomery40
Models in Library1
Datasets in Library0
Models on Hugging Face3
Followers1
Compare the published evidence
This profile connects the 1 model and dataset record shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Reported licenses: other. Open a record to review its specific license and conditions.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.