Since then DeepSeek burst onto the scene in January, and momentum has been building around China’s open-source AI models. Some researchers advocate an even more open approach to building artificial intelligence that would enable model creation to be distributed around the world.
Supreme Intellectstartup specializing in decentralized artificial intelligence, is currently training a pioneering huge language model called INTELLECT-3, using a modern type of distributed reinforcement learning for fine-tuning. The model will demonstrate a modern way to build competitive open AI models using a range of hardware in different locations in a way that is independent of huge technology companies, says Vincent Weisser, the company’s CEO.
Weisser says the AI world is currently divided between those relying on closed American models and those using open Chinese offerings. The technology being developed by Prime Intellect democratizes AI, enabling more people to create and modify advanced AI for themselves.
Improving AI models is no longer just about increasing the amount of training data and computation. Today’s pioneering models apply reinforcement learning to improve after the initial training process. Want your model to excel in math, answer legal questions, or play Sudoku? Let him improve by practicing in an environment where you can measure success and failure.
“These reinforcement learning environments are currently the bottleneck in truly scalable opportunities,” Weisser says.
Prime Intellect has created a framework that allows anyone to create a reinforcement learning environment tailored to a specific task. The company combines the best environments created by its own team and community to tune INTELLECT-3.
I tried running the Wordle puzzle-solving environment created by Prime Intellect researcher Will Brown, watching a tiny model solve Wordle puzzles (honestly, it was more methodical than I was). If I were an AI researcher trying to improve a model, I would fire up a few GPUs and train the model over and over while the reinforcement learning algorithm modified its weights, thus turning the model into a Wordle master.