Since Depseek The eruption of the January scene in January, the momentum was grown around open Chinese origin artificial intelligence models. Some researchers push for the most open method of building AI that allows the making model to distribute worldwide.

Prime wisdomA starting specialization of decentralized AI, currently trains a larger language model, which is called intellectual-3, using a new distribution classification classification for good tuning. The model will demonstrate a new way to establish competitive AI models using different locations in a way that does not trust the big tech companies, as Vincent Weiser, the company’s CEO.

Wiserer says the AI ​​world is currently divided between those who depend on closed US models and those who use open offerings in China. The main technology technology develops Democrats AI by allowing many people to build and change advanced AI for themselves.

Improvement of AI models is no longer a matter of rampage of training data and computation. The joint models ahead now use reinforcing encouragement to improve after the pre-training process. Want your model to exceed mathematics, answer legal questions, or play Sudoku? Press it to himself by practicing an environment where you can measure success and failure.

“These learning environments today are the bottleneck that actually scales the capabilities,” Wiser told me.

The main wisdom has generated a framework that allows anyone to produce a circle of coming on to reinforce for a particular task. The company combines the best environment that is self-held and in the community to thaw intellectual-3.

I have tried to run for the word puzzles, made by the principal research, to look as a small model to resolve word puzzles (more methods than me, to be faithful). If I was a researcher AI trying to improve a model, I will heal a set of GPUs and maintained the modeling while its algorithm algorithm changed the weights of an experienced master.