Google charges Gemini models faster than artificial intelligence safety reports

More than two years later that Google was arrested flat by Openai’s release ChatgptThe company picked up a lot.

In late March, Google launched a male thinking model, Gemini 2.5 ProWhich – which Leading the industry In several criteria that measure the capabilities of coding and mathematics. This launch came just three months after the technology giant appeared another model, Gemini 2.0 FlashThis was on the latest cases at the time.

In an interview, Google and President of Gemini, TULSEE DOSHI, Techcrunch, told the Google Producer, the growth of the company’s model is part of a concerted effort to keep pace with the advanced artificial intelligence industry.

“We are still trying to find out the correct way to take these models – what is the right way to get comments,” Douche said.

But it seems that the timeline of the escape has come at a cost. Google Safety reports have not yet been published on its latest modelsIncluding Gemini 2.5 Pro and Gemini 2.0 Flash, which raises fears that the company gives speed priority on transparency.

Today, it is somewhat accustomed to the Frontier AI laboratories – including Openai, Noteropic, and META – to report safety test, performance assessments, and use cases whenever you make a new model. These reports, which are sometimes called “system cards” or “forms cards”, were proposed years ago by researchers in industry and academic circles. Google was actually one of the first to suggest models cards in A. Search paper 2019, He described them “an approach to responsible, transparent and accountability practices.”

Doshi Techcrunch told the company that the company has not published a model card for Gemini 2.5 Pro because it considers the form a “experimental” version. She said that the goal of these experimental versions is to develop the Amnesty International model in a limited way, obtain notes, and repeat the model before the production is launched.

Google intends to publish the Gemini 2.5 Pro model when making the model available in general, according to Doshi, adding that the company has already conducted a safety test and Red Team.

In a follow -up message, a Google Techcrunch spokesman told the company’s “maximum priority” for the company and that it plans to issue more documents on artificial intelligence models, including Guemi 2.0 Flash, and move forward. Guemini 2.0 Flash, which is generally available, lacks a model card. The last model card released by Google Gemini 1.5 ProWhich came out more than a year ago.

System cards and models cards provide useful – do not spoil, sometimes – information that companies always do not always announce artificial intelligence. For example, the system card that was released for its O1 thinking model revealed that the company The model has a tendency to “plan” against humans And follow the goals secretly on their own.

In general, the artificial intelligence community believes that these reports are making good efforts to support independent research and safety assessments, but reports have been of additional importance in recent years. like adapter Previously, Google told the US government in 2023 that it would publish safety reports for all “important” important “” “” important “versions within the scope. Company Similar commitment to Other governmentsAnd promising “providing public transparency”.

There were regulatory efforts at the federal and state level in the United States to create safety reporting criteria for artificial intelligence developers. However, they were met with limited adoption and success. It was one of the most prominent attempts California vetoed the SB 1047 billThe technology industry is severely opposed. The legislators also submitted legislation that would allow the American Salama Institute, which is the standards of setting standards in the United States, To create instructions for typical versions. However, the Safety Institute is now facing Possible discounts Under Trump’s management.

From all manifestations, Google falls behind some of its promises to report the examples test and at the same time charging models faster than ever. It is a bad precedent, many Experts He argues – especially because these models become more capable and developed.