It is increasingly clear that AI models are a tool for the goods, as a sudden increase in open source sources such as Deepseek Show let alone be done together in a small budget. A young winner called S1 has also strengthened the idea, as researchers at Stanford and the University of Washington trained the “thinking” model using less than $ 50.
S1 is a direct competition of O1’s O1, called a consultation model because it produces the “thinking” comments of the related questions that can help you examine its work. For example, if the model is requested to explain how much money can replace all the UBER cars on the way of the Waymo steps, it can look how many Bers are on the street today, and it is very expensive Waymo Vehicle Manufacturing.
In accordance with TechcrunchS1 is based on the farm language model, which was taught to think by reading questions and answers from Google Model, Gemini 2.0 Severe thinking. Google model shows the thought process after each response is reinforcing the S1 of the small amount of training information – 1,000 cut carefully, and answers – and teaches you to imitate the Process of Geminema.
Some exciting information is how researchers are able to improve S1 performance using the easiest way:
Investigators used nifty trick to get S1 to look twice his work and expand their “thinking” time: They told them to wait. Add the word “Waiting” during S1’s thinking helped the model to reach more accurate steps, with paper.
This suggests that, despite concern that AI models stray the wall with skills, there are many fruit hanging. Some of the visual development at the computer science branch plundering the correct words of Incontation.
Openai reportedly cried for Wiles about the deepest Chinese group trained for its outgoing. The irony has not lost many people. ChatGPt and other large models are trained data from the web without permission, the issue remains prohibited in the courts such as companies such as companies such as New York Times want to protect their work to be used without compensation. Google too abundant prohibit rivals such as S1 from the training of Gemini exit.
Finally, the functioning of S1 is impressive, but it does not mean that a person can train a small model from the beginning of only $ 50. The model is actually a piggybackid off all gemini training, to get cheating sheet. The good comparison may be pressured in the picture. The default type of AI can compare with the JPEG of the image. Good, but yet lose. Also large models of language are resistant to many issues with accuracy, especially large models that require all the production of answers. But the same model can be useful in places such as the processing of features such as Apple Intelligence.
There was a lot of debate about what the cheapest source models may mean that the technical industry was written by a large sum of money. Is the openai tormented if their models can easily be copied by anyone? The company protectors claim that language models are always meant to be sold. MYAIAI, and Google and others, will succeed to build useful apps above models. More than 300 million people use Chatgpt each week, and the product has been similar to Chatbots and a new search method. Interface over Models, such as Opelai operator can navigate the web on the user, or unique data set as xai accessories in X (oversized to the highest partition.
Another thing to consider is that “anointing” is expected to be expensive. Recognition of the actual processing of each user question sent to the model. Since AI cheaper models and readily available, thinking is available, AI will contact all aspects of our lives, causing the greatest need for computer resources, not less. And a 1500 billion farm project of $ 500 billion will not be few. That is all of this hype around AI is not just a bubble.
Source link
