AI News: Researchers Build Powerful AI Model for Under $50

AI researchers from Stanford and the University of Washington have created s1, a low-cost AI reasoning model, using less than $50 in cloud computing.

Key Points

• s1 performs similarly to top AI models like OpenAI’s o1 and DeepSeek’s R1 in math and coding tests.
• The s1 model is available on GitHub, along with the data and code used to train it.
• The researchers said s1 is distilled from Google’s Gemini 2.0 Flash Thinking Experimental and trained on a small dataset of 1,000 curated questions.
• s1 is built on Alibaba’s Qwen AI model, its training took under 30 minutes using 16 Nvidia H100 GPUs.
• A unique trick—adding the word “wait”—helped s1 improve its reasoning accuracy.
• s1 is available on GitHub,its code and data are open to developers.

Why OpenAI, Google, and DeepSeek Invest Heavily in AI Training

o1, Gemini 2.0, and DeepSeek R1 are original large models, trained from scratch with extensive computing resources, making them more powerful than distilled models like s1.

OpenAI’s o1, Google’s Gemini 2.0 Flash Thinking Experimental, and DeepSeek’s R1 are built using large-scale pretraining and fine-tuning methods, rather than the distillation approach used for s1.

Original LLMs are trained on massive datasets containing text, code, and reasoning-based tasks and learn patterns, logic, and contextual relationships through self-supervised learning on high-performance GPU/TPU clusters over weeks or months.

After pretraining, these models undergo fine-tuning and these models employ chain-of-thought (CoT) prompting and self-reflection techniques to improve reasoning.

They also leverage retrieval-augmented generation (RAG) for accessing external knowledge.

News Gist

AI researchers from Stanford and the University of Washington created s1, a low-cost reasoning model, for under $50 in cloud computing.

Distilled from Google’s Gemini 2.0, s1 performs similarly to OpenAI’s o1 and DeepSeek’s R1 in math and coding.

Built on Alibaba’s Qwen, it trained in 30 minutes and is available on GitHub.