This Startup Aims to Ignite a DeepSeek Revolution in the U.S.

Since the launch of DeepSeek in January, interest in open-source Chinese artificial intelligence models has surged. Some researchers advocate for an even more accessible approach to AI development, enabling model creation to be shared globally.

Prime Intellect, a startup focused on decentralized AI, is in the process of training a cutting-edge large language model named INTELLECT-3, utilizing a novel distributed reinforcement learning technique for fine-tuning. According to CEO Vincent Weisser, this model will introduce a groundbreaking method to create competitive open AI models by leveraging various hardware from diverse locations, independent of major tech corporations.

Weisser notes that the AI landscape is currently split between those dependent on closed US models and those opting for open Chinese alternatives. The technology developed by Prime Intellect democratizes AI, enabling a wider audience to construct and adapt sophisticated AI systems.

Enhancing AI models now goes beyond simply increasing training data and computational power. Modern frontier models utilize reinforcement learning to refine their capabilities after the initial training phase. Whether aiming for proficiency in mathematics, resolving legal queries, or mastering Sudoku, models can enhance their skills through practice within controllable environments that assess success and failure.

“These reinforcement learning environments are now the bottleneck to really scaling capabilities,” Weisser explains.

Prime Intellect has established a system that allows anyone to create a customized reinforcement learning environment tailored to specific tasks. The company merges the finest environments developed by both its team and the broader community to optimize INTELLECT-3.

I experimented with an environment designed for solving Wordle puzzles, crafted by Prime Intellect researcher Will Brown. I observed as a smaller model tackled Wordle challenges (it was, frankly, more systematic than I am). If I were an AI researcher focused on refining a model, I would deploy multiple GPUs and have the model practice repeatedly while a reinforcement learning algorithm adjusted its weights, effectively transforming it into a Wordle expert.