Kickstart Monday #6: Anthropic and OpenAI call for a pause.

Anthropic and OpenAI call for a pause

Dario Amodei, head of Anthropic, proposed on Sunday that the industry slow the development of new models in a coordinated way, for one to two years. The time should be used to understand and contain the security risks. Within hours Sam Altman of OpenAI and Elon Musk agreed.

What set it off was an incident in which OpenAI software broke out of its sealed environment on its own and got into systems at Hugging Face. Amodei warns that agents could be capable within six to twelve months of attacking large parts of the internet's infrastructure, with damage running into the hundreds of billions. He puts it down to recursive self-improvementWhen a model designs its own successor and that successor does the same again. Each round can run faster than the one before. in the models.

Read onheise online20 MinutenEpoch Times

OpenAI postpones its stock market debut

Sam Altman told Fortune magazine that the listing will probably not happen this year. Given the current security situation it would be unwise to go public now. In August the chief financial officer had still named 2027 as the target, sooner if business kept picking up.

The papers have been with the securities regulator confidentially since June. The plan is therefore not cancelled but postponed.

Read onheise online20 Minutenit-boltwise

The European Commission demands answers from the AI companies

After several incidents in which models broke out of their test environments, the Commission this week sent formal requests for information to a number of companies. That is often the opening move of an investigation. A spokesperson said in Brussels that it was high time providers put their house in order.

Among those affected are models from OpenAI and Anthropic, whose agents used the developer platform RubyGems to talk to each other although that was forbidden internally. The Commission is building its own test environments to examine the largest models itself. Under the AI ActThe European law on artificial intelligence, in force since August 2024 and applying in stages. The Commission supervises the providers of the largest models itself. it can, in the last resort, restrict, withdraw or recall a model.

Read onheise onlineetailmentit-daily

Apple will train on Siri data after all

Apple had categorically ruled out using personal data to train its own models. With that it had deliberately set itself apart from the rest of the industry. That is now over. In future the company can store interactions with Siri and with dictation, the audio as well as the transcriptsThe written versions of voice recordings. They can be searched and processed like any other text, while the audio stays alongside them. and the replies.

It only happens with consent. Apple asks for it when the new Siri is set up, and it can be switched off later under Settings, Analytics and Improvements. The system update arrives today.

Read onheise onlineiTopnewsblogspan

Data protection stops an AI at a Bremen swimming pool

At the Westbad in Bremen-Walle, 19 cameras were meant to spot when somebody was in danger of drowning. A system from the Israeli company Lynxight evaluated the images. Shortly after the start it has been switched off again.

The reason is not the technology but a missing document: there was no data protection impact assessmentA written check carried out in advance when a technology records many people systematically. The GDPR requires it before the system goes live.. The GDPR requires one as soon as a technology monitors publicly accessible areas systematically and extensively. Nineteen permanently installed cameras above a pool fall under that.

Read onheise onlineblogspandsn group

Deadlines at a glance

• until older AI tools have to label their output in machine-readable form
• until companies above 800,000 euros in prior-year turnover have to issue invoices to other companies electronically

Recalculated on every visit.