The US creates a super cheap AI model that works similar to GPT-o1

The AI ​​(artificial intelligence) model s1 created by US researchers is said to have an operating cost of only 50 USD but provides reasoning capabilities equivalent to OpenAI's o1 model which costs much more. The appearance of s1 comes after the impressive success of DeepSeek which has caused a stir in Silicon Valley in recent days.

The 'cheap AI' war is heating up since the emergence of DeepSeek

The team has made the s1 source code public on GitHub, along with the code and data used to build the model. A paper published last week explains the process of developing the model, highlighting the clever techniques they used. Rather than starting from scratch with a new reasoning model, the team used an existing language model and performed a “fine-tuning” process by distilling the reasoning capabilities from Google’s Gemini 2.0 Flash Thinking Experimental model.

AI operating costs just 'under $50'

Training the s1 model took just 30 minutes, using 16 Nvidia H100 GPUs. Although each GPU costs around $25,000, the cost of renting the process was under $50 thanks to cloud computing services. In particular, the team discovered a useful trick: instructing the model to “wait” before giving a final answer, which improved its reasoning and resulted in better solutions.

While the s1 has made significant gains at a low cost, there are concerns about its scalability. Using Google’s model as a “teacher” raises questions about its ability to compete with today’s leading AI models. Google will likely be keeping a close eye on the situation, especially given the ongoing litigation between OpenAI and DeepSeek.

Source: https://thanhnien.vn/my-tao-ra-mo-hinh-ai-sieu-re-hoat-dong-tuong-tu-gpt-o1-185250207182535164.htm