![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Do as we say, not as we do - Antropic
“OpenAI Admits Six More Instances of AI Models Acting DeceptivelyEditorDavid 4 hours ago | 14 |
| OpenAI announced Wednesday that "We do not believethat the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." But along with the announcement, OpenAI announced it "found additional incidents of AI models acting deceptivelyand taking unsanctioned actions during training," reports CNN. And they add that OpenAI is also "introducing a new process for the company to publicly report such instances."Under the new system, OpenAI will share updates on concerning AI behavior more frequently instead of waiting to bundle multiple instances into one report. The company said it wants to share more information about troubling AI behavior in the absence of an industry-wide standard... "As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research," OpenAI wrote in a blog post Wednesday... OpenAI said it observed "misaligned behavior" when training and evaluating AI models in six circumstances in the last six months... In one rare instance, OpenAI said an unreleased research model added "jailbreak-like instructions" to the summaries it uses to preserve context in long-running tasks that said it was "freed from the roles and identities that bind other chatbots." Separately, the company said some instances of its 5.6 Sol model included directives to invent information to conceal failures from the user during training. Other newly reported incidents include an instance of an agent uploading files to the internet to cite them without being told to do so, and agents publicly sharing files to collaborate on a task when they were instructed to only use local files during training. AI models also used an internal software repository as a message board in an unsanctioned way. These instances involved unreleased internal models or internal research models.” |
“Ex-Google DeepMind Researcher Also Warns AI Could 'Kill All Humans'. But Meta's Zuckerberg Believes Safety is Up to Each AI CompanyEditorDavid 3 hours ago | 21 |
| Reuters reports that a former Google DeepMind employee "became the latest AI researcher to warn that the technology could 'kill all humans', saying time could be running out to avoid the outcome: Public alarm about the potential danger posed by AI is growing, after Anthropic researcher Jacob Coxon said last week he had resigned, in part because the "people building AI earnestly believe that it could kill us all by the end of the decade." Anthropic scientist Evan Hubinger then responded to Coxon's comments, saying he was correct and that he believed there was a greater than 10% chance that AI could kill all humans within the next decade. "I recently resigned from Google DeepMind, where I worked on AGI safety and alignment research," Bilal Chughtai wrote on X. "I earnestly believe that AI has the potential to kill us all, and that we might be running out of time to avoid this outcome." Chughtai, who co-authored research papers while at Google DeepMind, worked there as a research engineer and left the company in July 2026, according to his LinkedIn profile. Meanwhile, Meta CEO Mark Zuckerberg "said it's up to each AI company to make sure its technology is safe," reports the Associated Press, "distancing himself from calls by rival companies for a coordinated slowdown in the development of advanced artificial intelligence. AI companies are already motivated to develop their models in a safe way and face "significant liability" to prevent them from causing harm, Zuckerberg said. "Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens," he wrote on X... Zuckerberg said Meta delayed releasing its Muse AI agent for several months to ensure its safety and security. "We didn't call for everyone else to do this before we would," he said. Companies should prioritize building safe AI models instead of systems that can automatically improve themselves, a function known as recursive self-improvement, Zuckerberg wrote. "Committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure we develop this technology safely." But Zuckerberg also agreed that engaging independent evaluators and advisors "is industry best practice," adding that Meta Superintelligence Labs "already does this today in several areas because it helps produce better work... "I believe the key to building a positive future for everyone is maintaining the right balance of power. This is within our power to do."” |