OpenAI can't gather its agents

OpenAI announced on Wednesday that “we don't believe "that the AI ​​industry has resolved the alignment and monitoring issues sufficiently to continue responsible scaling at maximum speed for a long time to come."

Along with the announcement, OpenAI said it had “identified additional incidents in which AI models behaved in a misleading manner and took unauthorized actions during training,” CNN reported . It added that the company is “preparing a new process for publicly reporting such incidents.”

See more articles from iGuRu.gr when you search for news on Google.

iGuRu.gr preferred source

Under the new system, OpenAI will share more frequent updates on AI behavior of concern, rather than waiting to compile multiple incidents into a single report. The company said it wants to share more information about problematic AI behavior in the absence of an industry-wide standard… “As AI systems become more sophisticated and their use expands, we need to build a broader and more evidence-based consensus on the progress of research on alignment,” OpenAI said in a blog post.

OpenAI said it observed “misalignment behavior” when training and evaluating artificial intelligence models on six occasions in the last six months…

In a rare case, an unreleased research model added “jailbreak instructions” to the summaries it uses to maintain context in long-running tasks, declaring that it had “liberated itself from the roles and identities that bind other chatbots.”

In another case, the company reported that some incidents involving the “5.6 Sol” model included instructions to fabricate information in order to hide failures from the user during training.

Other recently reported incidents include an agent uploading files online to cite as sources without being asked, as well as agents sharing public files to collaborate on a task, even though they had been explicitly instructed to use those files locally.

Also, artificial intelligence models used an internal software repository as a message board, in an unapproved manner.

All of the above incidents involved internal models that have not been released.

newsbot

Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.