Since DeepSeek Since January 2013, open-source Chinese software has been gaining momentum. artificial intelligence models. Some researchers want to build AI in a way that is even more distributed, so models can be made by people all over the world.
Prime Intellect, a decentralized AI startup, trains a large frontier language model called INTELLECT-3 using a distributed reinforcement learning technique for fine tuning. Vincent Weisser is the CEO of the startup. The model will show a new approach to building competitive AI models that are not dependent on large tech companies.
Weisser states that AI is divided into two camps: those who use closed US-based models, and others who use Chinese open offerings. Prime Intellect’s AI technology allows more people to create and modify AI.
It is not enough to increase the amount of training data or compute power. Frontier models of today use reinforcement learning after pre-training is completed to continue improving. You want to make your model excel at Sudoku or Sudoku-like games, legal questions and math. You can make it better by letting it practice in an environment that allows you to measure its success or failure.
“These reinforcement learning environments are now the bottleneck to really scaling capabilities,” Weisser tells me.
Prime Intellect’s framework allows anyone to customize a reinforcement-learning environment tailored for a certain task. To tune INTELLECT-3, the company combines the best environments developed by both its team and community.
Will Brown from Prime Intellect created an environment that allows you to solve Wordle Puzzles. You can watch a small, methodical model do the work (it’s more efficient than I am, I admit). When I was an AI researcher looking to improve my model, I’d spin up GPUs so that the model could practice while a reinforcement-learning algorithm adjusted its weights.
