Researchers trained an OpenAI rival in half an hour for less than $50
SMRTR summary
Researchers developed s1, a cost-effective AI reasoning model, in 26 minutes using distillation. Based on Alibaba Cloud's Qwen2.5 and trained with 16 Nvidia H100 GPUs, s1 was refined using 1,000 questions for under $50. It employs test-time scaling and reportedly outperforms OpenAI's o1-preview on math questions by up to 27%. This achievement suggests that creating effective AI models may not require massive investments from major tech companies.
SMRTR provides this summary for quick context. The original article belongs to The Verge.
Read the original article