The researchers trained the Openai competitor in half an hour with less than 50 dollars

To do this, researchers at Stanford University and Washington University used a method known as distillation – which allows smaller models to extract from the answers produced by large answers – to improve S1 using answers from the Google thinking model, Gemini 2.0 Flash Thinking. Google Conditions of service Note that you cannot use the Gemini Application interface “to develop models competing with” AI models for the company. freedom Contact Google with a request for comment but not heard immediately.

Researchers depend on the S1 on QWEN2.5, which is an open source model from Alibaba Cloud. They initially started with a range of 59,000 questions to train the model, but they found that the larger data set does not make “great gains” on a set of only 1000. The researchers say they trained the model on 16 NVIDIA H100 graphics processing units.

The S1 model also uses a technology called the test time, allowing the “thinking” model for a longer period of time before producing an answer. As observed in the paper, the researchers forced the model to continue thinking by adding “waiting” to the form of the form. “This can lead to the model leading to its answer, and it often determines the incorrect steps of thinking,” the paper says.

Openai’s O1 thinks a similar method, something DEPSEEK Soft Amnesty International It sought to repetition with the launch of his R1 model, which claims to be trained on a small part of the cost. Openai has been accused since then Deepseek In distillation of information from its models to build a competitor, he violates the conditions of the service. As for the S1, the researchers claim that the S1 “exceeds the O1-PREVIEW of mathematics questions in the competition by up to 27 %.”

It threatens the smaller and cheapest artificial intelligence models by raising the entire industry. They can prove that major companies such as Openai, Microsoft, Meta and Google do not need to spend Billion dollars in artificial intelligence trainingwhile Huge construction Data centers filled with thousands of NVIDIA graphics processing units.