True Intrinsics
Pretraining should not be the end of learning.
True Intrinsics is an early-stage AI company developing technologies that enable AI agents to continually learn, acquire new skills from minimal experience, and improve through use.
We are application-first, driven by fundamental research in continual learning.
AI agents are entering both work and personal life. Today's AI models have remarkable capabilities. Yet adapting them to new tasks remains expensive, and much of what agents learn from experience is not retained as lasting capability.
We believe AI systems must learn from experience, rapidly acquire new skills from just a few demonstrations, and become more capable through everyday use.
Making this possible efficiently is the fundamental challenge we are pursuing.
We want AI agents to learn from experience and get better at their work, much like human professionals do.
We're pursuing this through two types of models.
Continual Learning Models
Learns as it goes.
Kumo is a new family of AI models we're developing to keep learning from experience, even after deployment.
We're working toward models that can pick up new skills from just one or two demonstrations, retain what they've learned, and adapt to new tasks without expensive retraining.
Our goal is to make continual learning dramatically faster and cheaper, with both open-weight and proprietary models.
Agent Companion Models (Agent-CMs)
A little help, a lot of learning.
Mochi = A Native Continual-Learning Companion Model for AI Agents
Most AI agents today rely on foundation models whose weights developers cannot change. But that shouldn't stop these agents from learning.
Mochi is a family of learning models that works alongside existing AI agents, even when their underlying foundation model weights are inaccessible.
It learns from what agents actually do: the tasks they attempt, where they fail, how people correct them, and what eventually works.
Over time, Mochi models aim to build up reusable knowledge and skills across tasks and environments, helping agents learn faster, avoid repeating mistakes, and improve without changing their underlying foundation models.
RESEARCH
How can AI agents acquire new skills from just a few demonstrations, retain what they learn, and continuously improve toward human professional-level performance?
This is the fundamental research challenge behind True Intrinsics.
