Researchers from Stanford and the University of Washington have successfully trained an AI reasoning model called s1, which performs similarly to OpenAI’s o1 and DeepSeek’s R1 on math and coding. The s1 model was created using a method called distillation, where it was fine-tuned from a base model by training on its answers.
The team used the Google AI Studio platform to access pre-trained models, including the Gemini 2.0 Flash Thinking Experimental, which they distilled into their own model. This process allowed them to create an AI model with strong reasoning capabilities for significantly lower costs than traditional methods.
Compared to a recent creation by Berkeley researchers that cost $450, s1 was trained using just $50 in cloud computing credits. The team achieved this feat by creating a small dataset of 1,000 carefully curated questions and answers, which they used to fine-tune the model.
The researchers’ approach shows promise for re-creating AI models with reduced costs, but also raises questions about the commoditization of these models. By using distillation, experts can create similar models without having to train them from scratch, which has significant implications for the future of AI development and deployment.
Source: https://iblnews.org/researchers-at-stanford-and-the-university-of-washington-trained-a-model-similar-to-openais-o1-and-deepseeks-r1