What We're Building
SenseAct is an AI research lab startup creating a new kind of intelligence rooted in true AI autonomy. Our goal is to achieve general-purpose AI that grows with time and experience by acting in the world. We are building artificial beings that live and grow in the world, learning and improving through real-world experience.
We focus exclusively on Agent AI—not Tool AI. Agent AI are autonomous systems that sense, reason, act, learn, and evolve new capabilities through ongoing experience. Tool AI, on the other hand, simply processes data or performs fixed tasks and does not improve after deployment. The majority of AI systems today—including those labeled as "agentic AI"—are actually Tool AI: they rely on large models that function as advanced tools but lack the ability to learn or adapt independently. Our mission is to create agents that interact with and learn from the real world, moving beyond passive tools.
We leverage continual learning, reinforcement learning, and sequence modeling to build adaptive, general-purpose intelligence. Our agents are not static models with extended context windows; instead, they are designed to learn and grow over time.
This approach unlocks new possibilities. For example, AI assistants or operating systems that genuinely adapt to each user, robots that get better through real-world experience, and artificial scientists that discover new knowledge by experimenting continuously.
Our Philosophy
Intelligence is fundamentally dynamic—it unfolds and grows through time. Our agents learn from direct experience (see The Era of Experience paper by David Silver and Richard Sutton), not from static, pre-collected datasets. We believe genuine capability emerges through continuous, autonomous interaction with the world, learning to predict, control, and adapt.
The distinction is fundamental: while current AI systems maintain fixed performance regardless of experience, our dynamic intelligence continuously improves through real-world interaction.
This points toward a new "scaling law" relating performance to experience—a frontier that remains largely unexplored at scale.
Research
Our approach is grounded in cutting-edge research in lifelong learning, streaming algorithms, and sequence modeling. These foundational papers demonstrate the core principles and architectures that power SenseAct:
- Streaming Deep Reinforcement Learning Finally Works — Introduces Stream-X algorithms that enable (for the first time) reinforcement learning agents to learn continuously from a steady stream of data, much like humans and animals. Watch this presentation video for an in-depth overview.
- Multi-stream Sequence Learning — Introduces Memora, a recurrent architecture designed to learn from multiple streams of data in their natural order, similar to what humans and animals do.
Contact
We're currently in stealth mode but welcome conversations with exceptional researchers, engineers, and early-stage investors who share our vision for truly autonomous intelligence.