Google shifts Gemini models faster than its AI security reports

After a violent foot after a violent foot after Google TittleThe company has drastically taken the pace.

In late March, Google launched the model of AI reasoning, Gemini 2.5 Pro:that manages the industry On several benchmarks of coding and mathematics. This launch came to another Model of Tech Giant, Gemini 2.0 flashIt was once modern art.

Google Director and Gemini Manager Toulse Doshi said in an interview with TechCrunch that a supporter effort to maintain the company’s model-rising cat-industry is a supporter.

“We are still trying to find out what the right way to put these models is the right way to have a feedback,” said Doschie.

But the term of release seems to come at a price. Luck should still publish security reports for its latest modelsIncluding Gemini 2.5 Pro and Gemini 2.0 Flash, raising concerns that the company prioritizes the speed of transparency.

Today, it is quite standard for borderline AI laboratories, including Openai, Anthropic and Meta, in case of reporting security testing, performance assessments. These reports, sometimes called system cards or model cards, have been offered years before by industry and academy researchers. Google was actually one of the first to suggest model cards a 2019 Research paper, Calling them “Approach for the responsible, transparent and reporting practice of machine learning.”

Doschi told TechCrunch that the company did not publish for Model Card Gemini 2.5 Pro, as it considers the issue of “experimental”. The purpose of these experimental releases is to install ai model in a limited way, get feedback and repeat the production start, he said.

Google intends to publish a Gemini 2.5 Pro Card when it makes the model generally available, adding that the company has already made security testing and adversary red teams.

In the further message, Google’s spokesman told TechCrung that security continues to have a “priority” company, which provides more documents to release its AI models, including Gemini 2.0 Flash. Gemini 2.0 Flash, which is generally available, does not have a model card. Google Released Model Card was intended Gemini 1.5 Prowhich came out more than a year ago.

System cards and model cards provide useful and unattractive, sometimes information that companies are not always widely advertised on their AI. For example, the Openai system card card has been released for its O1 reasoning model, revealed that the company The model adds to a “scheme” manand secretly pursue self-purpose goals.

In general, the AI ​​community perceives these reports as a good faith effort to support independent research and security assessments, but in recent years the reports have given additional importance. As Transformer It was previously mentioned that in 2023, Google told the US government that it would publish security reports on all the “significant” public AI model. Company made a similar commitment to Other governments:promising to “provide public transparency.”

The US Secondary Service Standards has been implemented in the Federal and State Levels of Federal and State Levels. However, they met with limited adoption and success. Was one of the more remarkable attempts Vetoed California Bill SB 1047whose technology industry was categorically opposed. Legislators have also laid legislation that will allow the US Security Institute, the US Standard Regulatory Authority. To establish model releases guidesA number However the security institute now stands Possible cuts Under Trump’s management.

Google falls from all its performances to report some of its promises to report on exemplary testing at the same time shipping models faster than ever. It’s a bad precedent, many experts argue – in particular, as these models become more capable and complicated.