More judgment in the model.
Teams should spend less time scripting every step. We want the model to carry more of the work: noticing what needs doing, acting and knowing when to ask.
We’re building smaller, cheaper models that finish tasks quickly, know the person and reuse what works.
Request model accessOur thesis
A capable assistant should take less supervision and less work to build.
Fast execution
Repeating context and directing every step turns getting help into another job.
Our goal: find the relevant information quickly, connect it to the request and move the task forward.
Personal judgment
Remembering a fact means little if it never changes the next decision.
Use preferences, past decisions and current context to anticipate needs, with judgment about what matters, when to interrupt and when to ask.
Reliable tool use
A successful tool call can still return the wrong result. Someone has to notice.
Choose the right tool, check the answer and recover from mistakes. The aim is a correct outcome with fewer corrections from you.
Learning through memory
Every task shouldn’t start from scratch. Every correction shouldn’t need repeating.
Our first focus is remembering corrections and reusing successful workflows. The goal: faster, more personal help as experience carries forward.
Research objectives illustrated above. Not live model demonstrations.
The direction
Teams should spend less time scripting every step. We want the model to carry more of the work: noticing what needs doing, acting and knowing when to ask.
Personal help should be available to everyone. Our goal is small models cheap enough to run all day, and eventually on your own device.
Longer-term research Lightweight model updates shaped by personal memory and preferences. Related research ↗
Personal help, for everyone.
For teams building personal assistants. Request early model access.
For model access and release updates. Privacy