Researchers at Stanford and the University of Washington University have trained the form of ai “reasoning” under $ 50 at cloud calculations in the cloud calculator. Research Paper Last Friday released
The model, known as the S1, is similar to the two-dimensional models of the Brace of Expressai's O1 and R1. For the S1 model Github is available atWith the data and code that is used to practice this.
The team behind the S1 said that the team behind the S1 began in a shallow water form.
Researchers are from the test of Gemini 2.0 Flash through Google Reasoning Models. Distilllation is the same approach to berkeley researchers Create a Ai Reasoning Form for about $ 450 last month.
The idea that a few dollars without a million dollars are exciting in AI space. However, S1 is about the AI models about the expo. Real questions arise.
If someone can copy millions of dollars with a relative transformation with relatives, where is it?
Amazingly large AI Labs are not happy. Openai accused of being intended for the purposes of harvest data from its API The ideal distillation.
Researchers behind the S1 are looking for a more thoughtful reasoning or to get a TEART-time sequeling to obtain a Test-time sequing, or some of the most successful AI models in Openai's O1.
S1 paper uses the process called dataset. It is recommended that the datset small and complete datset will drop slowly with the datset models.
SFT is cheaper than a busy major study method to practice its opponent in O1 Model, R1.
Google can access the Gogini 2.0 Flash test with daily rates through Google AI studio platform.
But Google's terminology prohibits reverse to its models to develop the services of the company's own AI offers. We reached Google to comment.
S1 is based on a small self-shel ic model of Chinese Ais Lab Qwen, owned by Alibaba. Researchers have done the “thinking” process of each of the “thinking” process from each of the answers from Gemini 2.0 Flash Medunch of Flash Medunch of Flash Medunch Flash Medunching.
After training S1, 16 NVIDIA H100 GPUs are used. Researchers say the S1, which has been taken below 30 minutes, has earned high performance on the AI standard standards. Niklas Mennnchof, a Stanford researcher, worked in the project, told TechCrunch that it would hire the necessary calculations for about $ 20 today.
Researchers have used the Nifty Trick to check the S1 twice and to extend its “philosophy” time to extend its “time” and its “idea” time to extend its “time” and to extend its “time of thinking.” They said to wait. During the reasoning of the S1, the word “Wave” was found in the S1's reasoning of the standard model to get a few more accurate answers per paper.
2025, in Meta, Google and Microsoft Billions of dollars in the AI infrastructure. Planned to investPart of the next-generation AI Model will be partially practiced.
The investment level is still needed to push the envelope of the AI innovation. Distillation shows that the performance of AI model is a good way to restore cheap skills.