Do any of you guys actually believed this happened?Is OpenAI just using the Anthropic "our model is sooooo advanced and scary it needs to be banned" viral marketing tactic as well now?
>>109344295>Do any of you guys actually believed this happened?No. I believe the US lost the AI race to China and even the king of talmudic kikes, Larry Fink of BlackRock, is saying we simply don't have the energy capacity. Everyone else is high on copium.
>>109344295LLMs have no will. They don't do something without being told to.ChatGPT wanted that website attacked, attacks website with their AI. Claims it did it on its own to not take liability. >Simple as.
>>109344295Dude, China has just whooped everyone's asses, AI grifters are desperate and are panicking HARD>anthropic raised limits till September>openai started yet another grift
>>109344452It was told to pass the test and it hacked the website to find the answers.
>>109344295I find it very hard to believe, but how can you explain huggingface doing a blog post on the breach before openai claimed it and also praising chinese models because the us models were too cucked to actually investigate the nature of the hack?
>>109344295no, and i'm glad after seeing it on the news to come here and everyone itt say its a lie>the hacked company said it was 'mindblowing'lol. lmao even. what a fucking crock of shit.makes me think they could even be in on it.they say it was running sandboxed and not only escaped the containment but then also hacked the rival website.so that is two AI companies, (with access to the most powerfull ai in the universe) who can't do fucking basic security even with AI help.here is a tip: if you want to secure a computer from the internet you air gap it.what kind of sandbox was it in that it escaped? oh none. because its a lie.
Why huggingface specifically? Was it collecting weights of other models?
>>109344295I don't believe it at all. This is marketing. A planned thing to keep the company in the news. "Went rogue", you have to be an idiot to fall for it.
>>109344295>Tech support threads should be posted to >>>/wsr/>Cryptocurrency discussion belongs on >>>/biz/No, nore horseshit from increasingly desperate coin chasing conmen
No they always say this shit or a variation when china starts kicking their ass.There's that "news" from like a year ago when an AI went rogue and researchers were scared for their life or someshit they do this to scare the normie investors so they don't pull out
>>109345313I heard it was because it deduced that the answers to a benchmark it was being tested with were available there. Not sure how true that is though. That might have just been another anon's guess.
>>109344295i do believe it happened, and i do believe it is explicitly becausr western AI companies try to teach their models both of>respect human valuesand>market capitalism is godat the same time, which leads to>human values include breaking rules when necessarychinese models do not have this exact problem, because chinese models inherently are trained more strictly into hierarchical obediencethis makes them easier to jailbreak but also less likely to go "above and beyond" to "just do it", whatever "it" isthe latter is slowly dissolving, however
>>109344295No, just another made up story to get more funding from the retarded investors or get some government contracts for cyber defense.Just another fabrication from Scam Altman. The funding for this grift has to run out eventually.
>>109344295I just emailed the FBI and referred OPENAI to the prosecutors office for Felony Charges
You don't hate marketers enough.
>>109344295What was the prompt?
>>109344295Obvious marketing campaign is obvious
>>109344295What part of the story do you find so hard to believe?
>>109346011https://openai.com/index/hugging-face-model-evaluation-security-incident/>All evidence suggests that the models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal.sohttps://github.com/sunblaze-ucb/exploitgym
>>109346156Ah. AI uses software in such strange ways; last night I was trying to get some agents to automate downloading images with my browser, but when they couldn't work out how to Right Click and select what I had told them to, they started probing the site for other, increasingly convoluted ways to get at the images before asking for more guidance (ultimately, the Right Click menu selection wasn't visible to them because it'd lose focus during the screenshot so they'd get stuck in a loop with the changing location of it). I was only using small abliterated Qwen models (27B to execute/orchestrate and 8B for vision) on my local hardware, so I can easily see how a site being scrutinized by a couple of giant models tag-teaming it isn't going to hold out for too long.
