Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Researchers in Stanford and Washington University in Stanford and Washington were able to train a “reasoning” model below $ 50 in cloud billing loans Research paper released last Friday.
The model, known as S1, implements the tests that measure math and coding abilities, Openai’s O1 and DeepSeek’s R1 like advanced thinking models. Is the S1 model Available in GitHubalong with the information and code used to prepare.
The team behind the S1 said they started with an outside base model, then corrected with distillation, and fixed a process to correct “reasoning” from another model. Researchers said S1 was experimenting from one of the justifiers of Google’s justification, the Twins 2.0. The same approach to distillation is Berkeley researchers Create an EU thinking model between about $ 450 last month.
Some of them are a few researchers behind them, are a few researchers, an innovative innovative innovative innovative in the AI space. However, S1 raises real questions about the commodity of AI models. If someone can repeat a million dollar model with relative pocket change, where is Moat?
Surprisingly, large AI laboratories are not happy. Openai is concerned about the data that is not properly accumulated from the API for purposes model expulsion.
The researchers behind the S1 were looking for a simplest approach to the performance of strong thinking and “test time scale” or to think more of the AI model before answering a question. These were a few of the Openai O1 trying to repeat through various methods of DeepSEEK and other AI laboratories.
S1 paper shows that the AI model can not be transported by a relatively small database using a process called a controlled subtle adjustment (SFT) to imitate certain behaviors in a database. SFT tends to be cheaper than a large-scale reinforcement training method that works to bring to the Openai’s Openai O1, R1.
Google offers free access to testing twins 2.0 Flash Turning with daily grade restrictions through the Google AI Studio platform. Its conditions, on the contrary, Google’s reverse engineering to develop services that compete with their AI offers. We have reached Google for comments.
S1 is based on a healing EU model, smaller than the Chinese AI Lab Qwen, which is available for free to download. To train S1, researchers are combined with answers to these questions, and paired with answers to these questions, but also created a database combined with the “Thinking” process behind each response.
16 Nvidia H100 After the S1 training, which uses less than 30 minutes using GPU, S1, in the opinion of the researchers, the S1 has achieved a strong performance on certain AI criteria. Stanford researcher Niklas Muennighoft, who works on the project, said that TechCrunch could rent the necessary calculation for about $ 20 today.
Researchers used the S1 to double check the work of the S1 and to take S1 to extend the “Thinking” time: They said they were waiting. In addition to the S1’s justification, adding “Wait” to the model helped some more accurate answers to the paper a little more accurately.
Meta, Google and Microsoft in 2025 He plans to invest $ hundreds of billion dollars in the AI infrastructureThe next generation will partially guide the AI models. This investment level may still need to push the AI innovation elegance. Distylation showed that the AI model has a good method for cheap reduction capabilities, but it does not create better new AI models than the present.
