Researchers Created an Open Rival to OpenAI’s o1 ‘Reasoning’ Model for Under $50
Breakthroughs in artificial intelligence are ever-evolving. Often, the breakthroughs need vast resources and extensive financial investment. However, researchers at Stanford and the University of Washington proved that innovation in AI development doesn't always have to come from deep pockets.
They managed to train an AI reasoning model called s1 under $50 worth of cloud compute credits. Develop with the s1 model. The s1 model competes with OpenAI's o1 and DeepSeek's R1 in reasoning, math, and coding performance, and it is openly available on GitHub.
This development raises critical questions about AI accessibility, commoditization, and broader implications of democratized AI development. Here, we dive into the technical depths of the s1 model, behind its creation, and the potential implications for the AI industry.
Development of s1
The team behind s1 began with an off-the-shelf base model from Qwen, a Chinese AI lab owned by Alibaba. Rather than training a new model from scratch, the researchers used a technique called distillation to fine-tune the model. Distillation is a method of training a smaller AI model using the outputs of a larger, more advanced AI model.
For s1, the group derived its reasoning ability from the Gemini 2.0 Flash Thinking Experimental model free on Google AI Studio, with daily usage limitations. By applying supervised fine-tuning (SFT), the group trained s1 on 1,000 handcrafted questions and answers and their associated reasons for why this was true that were sourced from Gemini 2.0 Flash Thinking Experimental.
Cost-Effective Training
The training of s1 was done cost-effectively:
Hardware Used: 16 Nvidia H100 GPUs
Duration: Within 30 minutes
Cost: Approximately $20: cloud computing resource cost
Comparative to investment value: AI giants invested heavily. For example, OpenAI, Meta, Google, and Microsoft all plan to spend hundreds of billions of dollars in 2025 as an investment in the future of AI infrastructure and next-generation models training.
Performance and Innovations
Despite the low-cost training process, s1 could still present sound performance on the key AI reasoning benchmarks. Researchers take a new approach that strengthens the model's accuracy by teaching the model to "wait" before deciding on a response, increasing its reasoning time thus yielding better outcomes. This simple prompt engineering approach shows how narrow tweaks can improve AI performance without much computational overhead.
Ethical and Legal Implications
The emergence of s1 highlights the rapid commoditization of AI reasoning models. If a team of researchers can replicate a multi-million-dollar AI model with minimal financial investment, it challenges the competitive edge of major AI companies.
Unsurprisingly, major AI labs are not pleased with this trend. OpenAI previously accused DeepSeek of using unauthorized API data for model distillation, raising concerns about intellectual property and data usage policies. Furthermore, Google's terms prohibit reverse-engineering its AI models for competitive purposes, which might place s1's developers in a precarious legal position.
The Future of AI Democratization
In conclusion, s1's success clearly indicates that with the critical development of AI technology, financial inputs are not enough to establish frontiers, whereas smaller teams still have much potential in creating impactful contributions by strategizing over established models and developing innovative training methods.
However, distillation alone is far from causing revolutionary leaps in AI capabilities. Instead, it can open avenues for wide accessibility of AI, providing room for smaller organizations and individual researchers to build competitive AI models at a fraction of the cost.
Frequently Asked Questions
What is the s1 model, and how does it compare to OpenAI's o1?
The s1 model is an AI reasoning model developed by researchers at Stanford and the University of Washington. It has shown to be similar in reasoning tasks, math, and coding benchmarks to OpenAI's o1 and DeepSeek's R1.
How was s1 trained for under $50?
The researchers used a process called distillation, leveraging Google's Gemini 2.0 Flash Thinking Experimental model. They trained s1 on a small dataset of 1,000 curated questions using supervised fine-tuning (SFT) in under 30 minutes with 16 Nvidia H100 GPUs.
What are the implications of s1 for the AI industry?
S1 demonstrates that high-performing AI models can be developed at low cost, challenging the dominance of major AI companies and raising questions about the commoditization of AI capabilities.
Are there any legal or ethical concerns surrounding s1?
Yes, OpenAI has previously accused other AI developers of using unauthorized API data for model distillation. Google's terms prohibit using its models to create competing services, which could lead to potential legal scrutiny for s1's researchers.
Can distillation lead to new AI breakthroughs?
Distillation is a good technique to replicate the current AI capabilities at lower costs but does not help in fundamental innovation. It is still large-scale investments that will be required for developing completely new AI architectures and capabilities.
The development of s1 proves to be an excellent example of the power of strategic AI model fine-tuning and cost-efficient training techniques. Although it won't revolutionize AI capabilities, it lowers the barrier to entry for AI research and makes more advanced reasoning models accessible to independent researchers and smaller organizations. In this regard, the continued democratization of AI will continue to challenge the status quo, provoking excitement as well as controversy in the industry.
