Researchers from Stanford and the University of Washington have developed an AI model called s1 that rivals the capabilities of expensive commercial AI systems while costing less than $50 in computing resources to train. The model, which was created through a process called distillation using Google’s Gemini 2.0 Flash Thinking Experimental model, demonstrates similar performance to OpenAI’s o1 and DeepSeek’s R1 in mathematical and coding tasks.
The research team achieved this breakthrough by using a carefully curated dataset of just 1,000 examples and an off-the-shelf base model from Chinese AI lab Qwen. The training process took only 30 minutes using 16 Nvidia H100 GPUs, with the researchers implementing a simple but effective technique of inserting the word “wait” to extend the model’s reasoning time and improve accuracy. This approach to controlling the model’s “thinking time” allows it to double-check its work before providing final answers.
The development of s1 raises significant questions about the necessity of massive investments in AI infrastructure, with major tech companies planning to spend hundreds of billions on AI development. While the s1 project demonstrates that smaller teams can achieve comparable results with minimal resources, industry leaders like OpenAI have expressed concerns about unauthorized model distillation, which they consider a violation of their terms of service. The researchers have made s1’s code and training data publicly available on GitHub, promoting transparency in AI development.
Soruces: Tim Kellogg, TechCrunch