Neuro-inspired plasticity in large models
draftExploring ways to let models change slowly over time while preserving core capabilities, inspired by biological learning and consolidation.
Primary focus: long-term memory, stability vs. adaptation.
SHV Developers · Research
Research directions
The model stack is anchored in a few specific research lines, not a vague idea of "AGI". This page tracks those lines and how they shape the developer API.
Emotional signals without role-play
draftDesigning Velin to output simple, honest signals about user state and conversational risk instead of pretending the system has feelings.
Primary focus: user wellbeing, safety, attachment.
Sparse adaptation with MG-SGLoRA
draftUsing sparse adapters and structured updates so that tenants can adjust behaviour for their domain without training a new base model.
Primary focus: multi-tenant safety, infra, reproducibility.
How this shows up in the API
The research is not separate from the product. If a research idea can't be expressed as a clear observable or control in the API, it isn't ready. When we design endpoints, we start with:
- • What state is the model tracking internally?
- • Which parts of that state should a developer see?
- • Which dials can be safely exposed and which must stay internal?