Prime Intellect launches an open platform for reinforcement learning environments
Key Points
- Prime Intellect has launched the Environments Hub, an open platform where users can build and share reinforcement learning (RL) environments, aiming to provide an alternative to the closed platforms developed by major AI labs.
- The company plans to use contributed environments from the Hub to train its next open-source model, INTELLECT-3, and is offering cash bounties for environments that test real-world skills like code quality, long tasks, and creative writing.
- Founded by Vincent Weisser and Johannes Hagemann, Prime Intellect is committed to decentralized AI development, has raised $20.5 million from backers such as Founders Fund and Andrej Karpathy, and previously released the decentralized INTELLECT-2 model.
Prime Intellect, a San Francisco AI startup, has launched the Environments Hub, an open platform for building and sharing reinforcement learning (RL) environments. The aim is to counter the closed systems built by major AI labs.
Prime Intellect calls interactive training environments a critical bottleneck for the next wave of AI progress. In these environments, AI agents learn by interacting with rule-based worlds, receiving new states and rewards for their actions. Reinforcement learning only makes sense, the company argues, when models face dynamic situations - otherwise, it's "just math."
Prime Intellect sees a trend where large AI labs pour millions into developing and acquiring proprietary RL environments. This push toward privatization, the company says, makes it harder for open-source efforts to build competitive AI models.
The Environments Hub is positioned as an open alternative. The company wants to build open platforms and models, not locked behind the walled gardens of big labs.
Crowdsourcing RL environments for INTELLECT-3
Over time, Prime Intellect plans to use the Environments Hub as a data engine for its next big open-source model, INTELLECT-3. The company describes INTELLECT-3 as a "fully open, state-of-the-art agentic model" trained on data from contributed RL environments.
To drive the creation of targeted environments, Prime Intellect has posted a list of bounties with cash rewards. They're looking for environments to evaluate code quality, support long-running tasks with filesystem integration, and enable creative writing. The goal is to not only build a more capable model but also lower infrastructure barriers for developers.
Prime Intellect was founded by CEO Vincent Weisser and CTO Johannes Hagemann. Weisser previously worked in decentralized science (DeSci), while Hagemann scaled LLM training at Aleph Alpha. The company is committed to decentralized AI, pooling compute resources, training models across distributed systems, and sharing results with the community.
Prime Intellect says it has raised $20.5 million from investors including Founders Fund, Andrej Karpathy, and Hugging Face CEO Clem Delangue. Earlier this year, the company released the decentralized INTELLECT-2 model.
AI News Without the Hype – Curated by Humans
As a THE DECODER subscriber, you get ad-free reading, our weekly AI newsletter, the exclusive "AI Radar" Frontier Report 6× per year, access to comments, and our complete archive.
Subscribe now