“We are really” pushing for thinking, “says Jack Rae, a principal research scientist from a deep mind. Such models, which are logically designed to work through problems and spend more time to reach the answer, became important with the issuance of the Deepsk R1 model earlier this year. They are attractive to AI companies because they can improve a current model by training a problem to approach a problem. In this way, companies can avoid creating a new model from the beginning.
When the AI ​​model devotes more time (and energy) to a question, it costs more. Leaderboards Reasoning models show that a job can cost more than $ 200. The promise is that they perform better in dealing with extra -time and money -help models, such as analysis of the code or collecting information from many documents.
“The more you can follow the speculation and ideas,” says Korekokloo, chief technical officer of Google Deep Mind, “so much more” is looking for the right thing. “
However, this is not true in all cases. “Eliminates the model,” says Tulsi Dhoshi, leading the product team in Gemini. “The simple indicator, the model thinks excessively.”
When a model spends more time on a problem, it makes the model expensive to run for developers and destroys AI’s environmental image.
The embracing facial engineer Nathan Habib, who has studied the spread of such arguments, says he is too much to think. Habib says that in the rush to show the smart AI, companies are reaching for models for hammer models for models for hammer, even not even nail in the eyes. In fact, when Openi announced a new model in February, he said it would be the company’s last unprecedented model.
Habib says the benefit of performance is “undeniable” for some tasks, but not for many other people where people usually use AI. Even when the reasoning is used for the right problem, matters can be worse. Habib showed me an example of a well -known reasoning model asked to work on the issue of organic chemistry. Its start is fine, but the model’s response to the process of reasoning began to resemble the error: he “wait, but …” spreads hundreds of times. The end of this takes more time than an unmarried model that will cost a job. Google models can also be trapped in loops, says Kate Olizoska, who works on the Gemini model examination in the Deep Mind, says.
Google’s new “reasoning” dial is an attempt to solve this problem. For now, it is made not for the user version of Gemini but for developers who are making apps. Developers can determine how much computing power should be spent on a particular issue, the idea is that if the work should not be added to the work, then reject the dial. When the reasoning is turned on, the Six is ​​almost six times more expensive to produce the results from the model.