Researchers from Stanford and Washington Universities have trained an AI model that rivals OpenAI’s advanced models in just 26 minutes and for less than $50. Named s1, this model was trained using a small dataset of 1,000 questions and developed through a technique known as distillation, which transfers knowledge from larger AI models to smaller ones.
$50 OpenAI Rival AI: s1
The researchers utilized Google’s Gemini 2.0 Flash Thinking Experimental model to enhance s1’s reasoning abilities. However, Google’s terms of service prohibit the use of its Gemini API for developing competing AI models, raising potential legal concerns.
The s1 model is built on Alibaba Cloud’s open-source Qwen2.5 model. While training initially began with a large dataset of 59,000 questions, tests revealed that a smaller dataset offered similar accuracy, significantly reducing training time and cost.
The training process was completed using 16 Nvidia H100 GPUs. The model improves its accuracy using a technique known as test-time scaling. Researchers added a “Wait” command to extend response times, allowing the model to review and correct its answers. This method enabled the AI to perform more accurate complex reasoning.
This development proves that high-performance AI models don’t necessarily require billions of dollars in investments and massive data centers to be trained. OpenAI’s O1 model also uses a similar reasoning technique, and China-based DeepSeek’s R1 model claims to have been developed at a much lower cost using a similar approach.
However, OpenAI has accused DeepSeek of distilling information from its models to develop a competing AI, which could lead major tech companies to reevaluate their stance on open-source AI models.
{{user}} {{datetime}}
{{text}}