The US finally may have found an answer to China’s DeepSeek technology. For months, China DeepSeek AI model has been giving sleepless nights to its counterparts in the AI race, reflecting that high-level reasoning AI model developments are not just exclusive to Western tech bosses like OpenAI or Google. But now, American researchers have retorted to this with an AI breakthrough of their own.
A team of researchers from Stanford University and the University of Washington have developed a new reasoning model, dubbed s1, that could beat OpenAI’s o1 and DeepSeek’s R1, all for less than $50 in cloud computing costs.
Breaking the myth of expensive AI
For years, developing a cutting-edge AI has been believed to involve billion-dollar investments and massive network of GPUs. OpenAI, Google, and Meta have been breaking the bank tl train and develop their AI models making it almost impossible for smaller teams to race against them.
The new s1 model take a leaf out of Berkeley researchers’ book who developed a similar model last month for approximately $450 a month. For the development of s1, the researchers have leveraged a process called distillation, which allows a smaller model to absorb the reasoning capabilities of a more advanced AI.
In the case of s1, it has been trained on responses from Google’s Gemini 2.0 Flash Thinking Experimental model. The result was an AI that performs on par with much larger, more expensive systems.
How s1 is built for only $50
The researchers behind s1 chose to go with a carefully curated set of 1,000 high-quality reasoning problems instead of falling back on massive datasets. It was observed that a smaller and thoughtfully curated dataset matched the performance of larger models, thus, eliminating the need for computationally expensive resources. The s1 model is available on GitHub for use along with its data and code used to train it.
China Vs US.: The AI battle goes a step-up
China’s DeepSeek R1 has quickly shot to fame due to its affordability and cheap computational costs. With s1, US researchers have a chance to demonstrate its reasoning model which is on par with DeepSeek and OpenAI with a minimal computational cost applied.
Ethical and legal questions
The s1 project, despite its success, raises some ethical concerns. The model was trained on Google’s Gemini 2.0 responses which potentially violates the latter’s terms of service. Google explicitly prohibits users from using their API to train competing models.
