A new AI reasoning model rival Openai, trained at less than $ 50 in a calculation

It is increasingly clear that AI language models are a commodity instrument as the sudden rise of open source suggestions as Deepseek Show that they can be hacked along with a relatively small budget. A new participant, called S1, reinforces this idea as researchers from Stanford and the University of Washington have trained the “reasoning” model, using less than $ 50 in cloud calculations.

S1 is a direct competitor to Openai’s O1, which is called a reasoning model, as it answers prompts through “thinking” through related questions that can help him check his work. For example, if the model is asked to determine how much money it could be to replace all Uber cars on the road with the Waymo fleet, it can break down the question into multiple steps – like checking how many Ubers are on the road today, and today Then how much does the Waymo vehicle cost for production.

According to TechcrunchThe S1 is based on a language model outside the shelf that was taught to think by studying questions and answers from a Google model, Gemini 2.0 flashing thinking experimental. The Google model shows the thinking process behind any answer that returns, allowing S1 developers to give their model a relatively small amount of training data – 1,000 carefully cure questions, along with the answers – and teach it to mimic the process of Gemini thinking.

Another interesting detail is how researchers were able to improve the work of S1 reasoning using a brilliantly simple method:

Researchers used a fine trick to get the S1 to check their work and extend their time for “thinking”: they told him to wait. Adding the word “wait” during S1’s reasoning helped the model get a little more accurate answers, according to paper.

This suggests that despite the concerns that AI models hit a wall in the capabilities, there is a lot of low -hanging fruits. Some remarkable improvements to the Computer Science branch are descended to create the right spell words.

Openai is reported a bird For the Chinese Deepseek team, which trains their models. Irony is not lost by most people. Chatgpt and other basic models were trained in data scraped by the network without permission, and the problem is still disputed in courts as companies such as New York Times Look to protect their work from using without compensation. Google too technically It prohibits competitors like S1 from training for Gemini results.

After all, the performance of the S1 is impressive, but it does not imply that one can train a smaller model than scratch with only $ 50. The model essentially gave up all the twin training sessions by receiving a cheat sheet. A good analogy can be the image compression. Distilled version of AI model can be compared to JPEG in a photo. Okay, but still lost. And large language patterns still suffers from many problems with accuracyParticularly large -scale common models that are looking for the whole network to produce answers. It also seems leaders in companies such as Google Skim over text generated by AI without checking it. But a model like S1 can be useful in areas such as processing features such as Apple Intelligence.

There is a lot of debate about what the rise of cheap open source models for the technology industry means. Openai is doomed if his models can easily be copied by someone? The company’s defenders say language models have always been destined to be codified. Openai, along with Google and others, will be able to build useful apps at the top of the models. More than 300 million people use Chatgpt every week, and the product has become synonymous with chatbot and a new search form. The interface on top of the models, such as Openai Operator This can be oriented on the network for a user or a unique set of data such as XAI to X’s access (previously Twitter) is what will be the best differential.

Another thing to keep in mind is that the “conclusion” is expected to remain expensive. The conclusion is the actual processing of each user request sent to a model. As AI models become more expensive and more accessible, thinking goes, AI will infect every aspect of our lives, which leads to a much more demand for computing resources, not less. And on Openai A $ 500 billion server farm project It will not be a loss. This is, while all this over -AI is not just a balloon.