Researchers from Stanford and the University of Washington successfully trained an AI “reasoning” model for less than $50 using cloud computing credits, as detailed in a recent research paper published last Friday.
The model, referred to as s1, exhibits performance comparable to leading reasoning models like OpenAI’s o1 and DeepSeek’s R1 on assessments that gauge mathematical and programming skills. You can find the s1 model on GitHub, along with the associated training data and code.
The s1 development team started with a standard base model and refined it using a process known as distillation, which extracts reasoning abilities from another AI model by training on its responses.
The researchers disclosed that the s1 model was distilled from one of Google’s reasoning models, named Gemini 2.0 Flash Thinking Experimental. This method of distillation mirrors the technique employed by Berkeley researchers to develop an AI reasoning model for approximately $450 last month.
The notion that a small group of researchers can make significant advancements in AI without substantial funding is invigorating to many. However, s1 raises pertinent questions regarding the commodification of AI models.
What safeguards exist if an individual can replicate a multi-million-dollar model for a fraction of the cost?
Not surprisingly, major AI laboratories are displeased. OpenAI has accused DeepSeek of illegally mining data from its API for model distillation.
The s1 research team aimed to discover the most straightforward method to attain impressive reasoning capabilities and “test-time scaling,” which allows an AI model to engage in deeper thinking before responding. These were among the key innovations in OpenAI’s o1, which DeepSeek and other AI labs have attempted to replicate through various means.
According to the s1 paper, reasoning models can be distilled using a relatively small dataset through a technique known as supervised fine-tuning (SFT), which instructs an AI model to imitate specific behaviors in a dataset.
SFT generally proves to be less expensive than the large-scale reinforcement learning approach that DeepSeek used to develop R1, its competitor model to OpenAI’s o1.
Google provides free access to Gemini 2.0 Flash Thinking Experimental, albeit with daily usage limits, through its Google AI Studio platform.
However, Google’s terms prohibit reverse-engineering its models to create competing AI services. We have reached out to Google for additional comments.
S1 is based on a compact, readily available AI model from Qwen, a Chinese AI lab owned by Alibaba, which can be downloaded for free. To train s1, the researchers compiled a dataset containing just 1,000 meticulously selected questions, accompanied by their answers and the reasoning processes derived from Google’s Gemini 2.0 Flash Thinking Experimental.
Following the training of s1, which took less than 30 minutes using 16 Nvidia H100 GPUs, it demonstrated impressive results on certain AI benchmarks, as stated by the researchers. Niklas Muennighoff, a Stanford researcher involved in the project, mentioned to TechCrunch that he could lease the necessary computing resources today for approximately $20.
The researchers employed a clever technique to enable s1 to verify its outputs and prolong its “thinking” duration: they instructed it to “wait.” Incorporating the word “wait” during s1’s reasoning process aided the model in producing slightly more accurate results, according to the paper.
In 2025, tech giants like Meta, Google, and Microsoft are set to invest hundreds of billions of dollars into AI infrastructure, a portion of which will be allocated toward training advanced AI models.
This level of funding may still be essential to advance AI innovation further. While distillation has proven to be an effective way to recreate an AI model’s capabilities at a lower cost, it doesn’t lead to the creation of entirely new models that significantly outperform current technologies.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
