Sam Altman Co-founder and CEO of OpenAI speaks through the Italian Tech Week 2024 at OGR Officine Grandi Riparazioni on September 25, 2024 in Turin, Italy.
Stefano Guidi | Getty Photographs Information | Getty Photographs
OpenAI on Wednesday introduced a brand new “security evaluations hub,” a webpage the place it should publicly show artificial intelligence fashions’ security outcomes and the way they carry out on exams for hallucinations, jailbreaks and dangerous content material, equivalent to “hateful content material or illicit recommendation.”
OpenAI stated it used the protection evaluations “internally as one a part of our determination making about mannequin security and deployment,” and that whereas system playing cards launch security check outcomes when a mannequin is launched, OpenAI will any longer “share metrics on an ongoing foundation.”
“We are going to replace the hub periodically as a part of our ongoing company-wide effort to speak extra proactively about security,” OpenAI wrote on the webpage, including that the protection evaluations hub doesn’t mirror the complete security efforts and metrics and as an alternative exhibits a “snapshot.”
The information comes after CNBC reported earlier Wednesday that tech firms which can be main the way in which in synthetic intelligence are prioritizing merchandise over analysis, in line with trade consultants who’re sounding the alarm about safety.
CNBC reached out to OpenAI and different AI labs talked about within the story nicely earlier than it was printed.
OpenAI just lately sparked some on-line controversy for not operating sure security evaluations on the ultimate model of its o1 AI mannequin.
In a current interview with CNBC, Johannes Heidecke, OpenAI’s head of security techniques, stated the corporate ran its preparedness evaluations on near-final variations of the o1 mannequin, and that minor variations to the mannequin that passed off after these exams would not have contributed to vital jumps in its intelligence or reasoning and thus would not require extra evaluations.
Nonetheless, Heidecke acknowledged within the interview that OpenAI missed a possibility to extra clearly clarify the distinction.
Meta, which was additionally talked about in CNBC’s reporting on AI security and analysis, additionally made an announcement Wednesday.
The corporate’s Basic AI Analysis staff launched new joint analysis with the Rothschild Basis Hospital and an open dataset for advancing molecular discovery.
“By making our analysis extensively accessible, we purpose to offer easy accessibility for the AI group and assist foster an open ecosystem that accelerates progress, drives innovation, and advantages society as a complete, together with our nationwide analysis labs,” Meta wrote in a blog post asserting the analysis developments.