Researchers build OpenAI rival for under $50, raising questions about AI’s future
Thora Flatley
▶February 8, 2025
Become smarter in just 5 minutes
Feeeed serve up the most interesting stories in AI, tech, business, science, and more—all in 5 minutes or less.
Booyah! The first newsletter will go out early next week.
Oops! Something went wrong while submitting the form.
Why it matters: A team from Stanford and the University of Washington trained a reasoning model, s1, for just $50 in cloud credits. The model rivals OpenAI’s o1 and DeepSeek’s R1, showing that cutting-edge AI may no longer require massive budgets.
The big picture:
- How they did it: Researchers fine-tuned an off-the-shelf model using a process called distillation, training it on answers from Google’s Gemini 2.0 Flash Thinking Experimental model.
- The cost: Training s1 took just 30 minutes on 16 Nvidia H100 GPUs, with compute costs estimated at $20–$50.
- Performance: s1 matches top reasoning models in math and coding benchmarks, despite being trained on a dataset of only 1,000 curated questions.
Between the lines:
- Open-source shift: s1’s code and data are available on GitHub, making it accessible to anyone.
- Industry disruption: If high-performing AI models can be replicated cheaply, it challenges the dominance of big labs like OpenAI, which spend millions on training.
The controversy:
- Data ethics: OpenAI has accused competitors like DeepSeek of improperly using its API data for distillation. Google’s terms also forbid reverse-engineering its models, raising questions about s1’s legality.
- Limitations: While distillation is cost-effective, it doesn’t create groundbreaking new models — it simply replicates existing ones.
What’s next:As AI giants like Meta, Google, and Microsoft pour billions into next-gen models, s1’s success highlights the growing accessibility of AI innovation. But legal battles over data use and intellectual property could shape the future of open AI development.
