New AI Reasoning Model Trained on Less Than $50 Takes on OpenAI

The emergence of affordable AI language models is becoming increasingly evident, driven by the rise of open-source alternatives such as DeepSeek. A recent development in this landscape is the introduction of a new model, S1, created by researchers at Stanford and the University of Washington, who trained it using less than $50 in cloud computing resources.

S1 has positioned itself as a competitor to OpenAI's o1 model, noted for its reasoning capabilities. The S1 model provides responses to prompts by analyzing related questions to validate its answers. For example, if asked about the potential cost of replacing all Uber vehicles with a fleet from Waymo, S1 breaks the question into smaller interrogations—such as determining the current number of Ubers on the road and the manufacturing cost of Waymo vehicles.

According to reports from TechCrunch, S1 is built on a pre-existing language model and was trained to reason by learning from a dataset obtained from Google's Gemini 2.0 Flashing Thinking Experimental model. This model demonstrates the thought process behind its answers, which allowed the S1 team to utilize a minimal training set of 1,000 carefully selected questions and answers to immitate Gemini's reasoning capabilities.

One interesting aspect of how S1 enhances its reasoning skills is through a simple but effective technique: the researchers added the instruction "wait" during S1's reasoning process. Incorporating this directive helps the model to reassess its answers, resulting in slightly improved accuracy, as highlighted in their findings.

This discovery indicates that while some experts express concerns about AI reaching its limits in terms of capabilities, there are still many easily accessible improvements available. Sometimes, achieving advancements in this field is simply a matter of using the right commands.

OpenAI has raised concerns regarding the Chinese DeepSeek team's use of its model outputs for training. The situation presents an interesting irony, considering that models like ChatGPT were also formed using data collected from across the internet without consent, a practice currently under legal scrutiny. Additionally, Google has stated restrictions against competitors like S1 leveraging outputs from its Gemini model for training.

While the performance of S1 is noteworthy, it's important to note that training a distinct model from scratch for merely $50 isn't feasible. Instead, S1 draws upon the extensive knowledge base of the Gemini model, in a manner akin to receiving an educational 'cheat sheet.' In this comparison, the summary version of an AI model can be likened to a JPEG image: it may capture essential aspects, but certain details are inevitably lost. Large language models often grapple with significant accuracy challenges, particularly when they sift through vast web resources for answers. Even major firms, like Google, may overlook checking the correctness of AI-generated text.

The S1 model shows potential applications in specific areas, such as on-device processing capabilities for technologies like Apple Intelligence.

As the availability of inexpensive, open-source models increases, there are ongoing discussions regarding the implications for the broader technology industry. There are questions about the future of OpenAI if its models can be easily replicated. Supporters argue that language models are destined to become commoditized. Companies like OpenAI and Google are likely to continue thriving by developing innovative applications founded on these models. Currently, ChatGPT enjoys substantial popularity, attracting over 300 million users weekly, making it synonymous with chatbots and modern search tools. The unique interfaces that top companies create over these models, such as OpenAI's Operator which can navigate the web on behalf of users, will ultimately serve as key differentiators in the marketplace.

Another factor to consider is that the cost of inference—processing user inquiries posed to a model—is still expected to remain high. As AI systems become more affordable and accessible, the demand for computing resources is likely to escalate rather than decline. This suggests that OpenAI's ambitious plans for a $500 billion server infrastructure might not be in vain, assuming the current AI hype does not prove to be a passing trend.

AI, Reasoning, S1, OpenAI, Training