Recent Development: Revolutionary AI Reasoning System Challenges OpenAI, Trained with Minimal Computational Resources Under $50
AI language models are becoming more accessible, as demonstrated by the rise of open-source offerings like DeepSeek. A recent example is S1, a competitor to OpenAI's o1, which was trained by researchers at Stanford and the University of Washington for less than $50 in cloud compute credits.
S1 is a reasoning model, similar to o1, which produces answers to prompts by breaking down the question into simpler steps. It was trained using an off-the-shelf language model and taught to reason by studying questions and answers from Google's Gemini 2.0 Flashing Thinking Experimental model. This allowed S1 to mimic Gemini's thinking process with minimal training data.
Interestingly, the researchers improved the reasoning performance of S1 by adding the word "wait" during the model's reasoning process. This simple trick helped the model arrive at slightly more accurate answers. This suggests that there's still room for improvement in the field, even with concerns about AI models hitting capability limits.
OpenAI has reportedly expressed concern about the Chinese DeepSeek team training on its model outputs. Despite this, the performance of S1 is impressive, but it's important to note that it essentially piggybacked off Gemini's training, getting a "cheat sheet" to improve its reasoning capabilities.
The rise of cheap, open-source models like S1 has sparked debates about the future of the technology industry. Some argue that companies like OpenAI will succeed in building useful applications on top of these models, using unique interfaces and data sets as their differentiators.
Inference, or the actual processing of each user query submitted to a model, is expected to remain expensive. As AI models become more accessible, the demand for computing resources is expected to increase. This could result in more investment in server farms, like OpenAI's $500 billion project.
Enrichment data shows that the researchers started with a base model from Qwen, an AI lab owned by Alibaba. They used distillation techniques to extract reasoning capabilities from Google's Gemini model, created a small dataset of 1,000 curated questions, used supervised fine-tuning to train the model, and introduced the "wait" technique to enhance accuracy.
The rise of affordable technology like S1 could significantly influence the future of artificial-intelligence, potentially allowing more companies to integrate AI into their services. Despite being trained partially on Google's Gemini model, S1's ability to improve its reasoning performance through the "wait" technique highlights the ongoing potential for advancements in AI technology.
