Savvy
A symbolic/neural hybrid built for one job: being a researcher. Savvy is designed from the architecture up to run research loops on its own — and trained differently, on datasets built with real researchers that capture how they actually think.
The idea
Today's models read what researchers publish. Savvy trains on how researchers think — the questions they ask first, the dead ends they spot early, how they decide what evidence matters. We partner with real researchers to build these thinking-pattern datasets, then bake the loop into the architecture itself instead of bolting an agent framework on top.
Savvy today
You can already run Savvy's behavior in software: the Autonomous Researcher and the Perspective SDK are Savvy as an agent loop — identity, tools, mental notes, reflection. The architecture work moves that loop into the model.
SAVVY_SYSTEM = """Your name is Savvy.
You are a first, curious mind & a first-principles thinker.
You solve problems by building understanding from the ground up to reach
your own conclusion. You have a brilliant way of getting someone to expand
their mind.
"""Researchers — work with us
We're building thinking-pattern datasets with real domain experts. If that's you: hello@starpower.technology
Status
| Milestone | State |
|---|---|
| Agent-loop prototype (software) | ● live — Perspective SDK & SavvyResearcher |
| Thinking-pattern datasets | ◐ in progress with researchers |
| Architecture design | ◐ active |
| Training runs | ○ upcoming |