This Startup Wants to Spark a US DeepSeek Moment (www.wired.com)

Prime Intellect, a startup focused on decentralized AI, is training a new frontier LLM called INTELLECT-3 using a novel distributed reinforcement learning (RL) approach for fine-tuning. Rather than relying on a single cloud provider or massive centralized infrastructure, the company stitches together heterogeneous hardware across locations and lets models improve post‑pretraining by practicing in task‑specific RL environments. Prime Intellect has released a framework so researchers and community contributors can build custom environments (the team demoed a Wordle solver) and combine the best environments to push capability growth. This matters because modern frontier models no longer advance solely by scaling data and compute—RL fine‑tuning is the key to teaching models skills like reasoning, math, or domain expertise, and today that work is mostly closed behind big AI labs. By open‑sourcing distributed RL tooling and demonstrating that calculations can be partitioned and recombined (INTELLECT‑1 was a 10B distributed model; INTELLECT‑2 used distributed RL to boost reasoning), Prime Intellect aims to democratize capability tuning and enable startups to build specialized agents without huge staff/expertise. Endorsements from figures like Andrej Karpathy and the momentum behind Chinese open models (DeepSeek, Qwen, etc.) underscore why distributed RL environments could be the next inflection point for open AI research and productization.