Researchers from Stanford University and the University of Washington have achieved a remarkable breakthrough in artificial intelligence by training a new reasoning model, named s1, for less than $50 using cloud computing credits. This innovative development was detailed in a research paper released last Friday.
The s1 model exhibits performance on par with leading reasoning models, such as OpenAI’s o1 and DeepSeek’s R1, particularly in tests assessing mathematical and coding skills. The model, along with its training data and code, is now available on GitHub, providing an opportunity for further exploration and development by the AI community.
The team behind s1 began with a readily available base model and refined it through a technique known as distillation. This process involves extracting reasoning capabilities from an existing AI model by training the new model on the answers provided by the original. Specifically, s1 was distilled from Google’s Gemini 2.0 Flash Thinking Experimental model. This approach is similar to the method used by researchers at Berkeley, who developed a comparable AI reasoning model for approximately $450 just last month.
The ability of a small team of researchers to innovate in the AI field without substantial financial backing has generated excitement among industry observers. However, the emergence of s1 raises critical questions about the commoditization of AI models. If a multi-million-dollar model can be closely replicated for a fraction of the cost, it challenges traditional notions of competitive advantage in the AI landscape.
Major AI laboratories have expressed concern over this development. OpenAI has accused DeepSeek of improperly utilizing data from its API to facilitate model distillation, highlighting ongoing tensions within the industry.
The s1 research team aimed to identify the simplest method to achieve robust reasoning performance and enhance “test-time scaling,” which allows an AI model to engage in deeper thought before responding to queries. These objectives align with breakthroughs seen in OpenAI’s o1, which other AI labs, including DeepSeek, have sought to replicate through various methodologies.
The findings in the s1 paper indicate that reasoning models can be distilled using relatively small datasets through a process called supervised fine-tuning (SFT). This method instructs an AI model to imitate specific behaviors within a dataset and is generally more cost-effective than the large-scale reinforcement learning techniques employed by DeepSeek for its R1 model.
Google provides free access to its Gemini 2.0 Flash Thinking Experimental model, albeit with daily usage limits, through its Google AI Studio platform. However, Google’s terms prohibit reverse-engineering its models to create competing services, and the company has been contacted for further comment.
The s1 model is based on a compact, off-the-shelf AI model from Qwen, a Chinese AI lab owned by Alibaba, which is freely available for download. To train s1, the researchers compiled a dataset of just 1,000 meticulously selected questions, complete with answers and the reasoning process derived from Google’s Gemini model. The training process for s1 took less than 30 minutes using 16 Nvidia H100 GPUs, and the researchers noted that the necessary computing resources could be rented for about $20.
An interesting technique employed by the researchers involved instructing s1 to “wait” during its reasoning process, which resulted in slightly more accurate answers.
Looking ahead, major tech companies like Meta, Google, and Microsoft are set to invest hundreds of billions of dollars in AI infrastructure by 2025, which will support the development of next-generation AI models. While distillation has proven effective for recreating existing AI capabilities at a lower cost, it may not lead to the creation of significantly superior models compared to those currently available.
