Ever since DeepSeek burst onto the scene in January, momentum has grown around open source Chinese artificial intelligence models. Some researchers are pushing for an even more open approach to building AI that allows model-making to be distributed across the globe.
Prime Intellect, a startup specializing in decentralized AI, is currently training a frontier large language model, called INTELLECT-3, using a new kind of distributed reinforcement learning for fine-tuning. The model will demonstrate a new way to build competitive open AI models using a range of hardware in different locations in a way that does not rely on big tech companies, says Vincent Weisser, the company’s CEO.
Weisser says that the AI world is currently divided between those who rely on closed US models and those who use open Chinese offerings. The technology Prime Intellect is developing democratizes AI by letting more people build and modify advanced AI for themselves.
Improving AI models is no longer a matter of just ramping up training data and compute. Today’s frontier models use reinforcement learning to improve after the pre-training process is complete. Want your model to excel at math, answer legal questions, or play Sudoku? Have it improve itself by practicing in an environment where you can measure success and failure.
“These reinforcement learning environments are now the bottleneck to really scaling capabilities,” Weisser tells me.
Prime Intellect has created a framework that lets anyone create a reinforcement learning environment customized for a particular task. The company is combining the best environments created by its own team and the community to tune INTELLECT-3.
I tried running an environment for solving Wordle puzzles, created by Prime Intellect researcher, Will Brown, watching as a small model solved Wordle puzzles (it was more methodical than me, to be honest). If I were an AI researcher trying to improve a model, I would spin up a bunch of GPUs and have the model practice over and over while a reinforcement learning algorithm modified its weights, thus turning the model into a Wordle master.
Read the full article here
