In a refreshing move towards open and accessible AI research, a group of researchers has announced the successful training of S1, a new reasoning model that prioritizes both speed and transparency. What’s truly remarkable is the lightning-fast training time: just 30 minutes using a modest setup of 16 Nvidia GPUs!
How did they achieve this impressive feat? The team leveraged Supervised Fine-Tuning (SFT), a method they describe as “faster and less time-consuming” compared to more complex approaches like Reinforcement Learning (RL) used by models like DeepSeek’s r1. SFT essentially guides the model with carefully curated examples, rather than forcing it to learn solely through trial and error, as is often the case with RL.
But the ingenuity doesn’t stop at efficient training. The foundation of S1’s reasoning prowess is built upon Gemini 2.0’s responses. The researchers didn’t just feed the model answers; they incorporated the reasoning behind those answers as well. This holistic approach, using 1,000 meticulously selected question-answer pairs enriched with Gemini 2.0’s insightful rationale, forms the bedrock of S1’s abilities.
What truly sets S1 apart is its commitment to openness and transparency. As the research group eloquently states:
“Recent advances in reasoning, such as OpenAI’s o1 and DeepSeek’s r1, lack transparency, limiting broader research progress. Our work aims to push the frontier of reasoning in a fully open manner, fostering innovation and collaboration to accelerate advancements that ultimately benefit society.”
This statement underlines a critical point. While models like OpenAI’s o1 and DeepSeek’s r1 represent significant leaps in AI reasoning, their inner workings are often opaque “black boxes.” This lack of transparency hinders the broader research community. S1, in contrast, is designed to be a beacon of open research, encouraging collaboration and accelerating the pace of innovation.
The choice of SFT over RL further contributes to this goal of accessibility. RL, while powerful, can be computationally intensive and time-consuming, often requiring significant resources and expertise. SFT offers a more streamlined and efficient alternative, making advanced AI model training more attainable for a wider range of researchers and developers.
The implications of S1 are significant. By demonstrating that a powerful reasoning model can be trained quickly, efficiently, and transparently, this research paves the way for:
Faster AI Development Cycles: Reduced training times mean quicker experimentation and iteration in AI research.
Democratized AI Research: Lower resource requirements open the door for more researchers and institutions to contribute to cutting-edge AI development.
Increased Transparency and Understanding: Open models facilitate deeper analysis and understanding of how AI reasoning works, leading to safer and more reliable systems.
Accelerated Societal Benefit: By fostering collaboration and rapid innovation, S1 and similar open approaches can expedite the development of AI solutions that address real-world challenges.
The S1 model is more than just a technical achievement; it’s a statement of intent. It represents a powerful push towards a future of AI research that is open, collaborative, and ultimately, more beneficial for everyone. It will be exciting to see how the research community embraces and builds upon this groundbreaking work.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →