I have a strong suspicion that a lot of these "experiments" are done with very specific goals and orders for the LLMs in question. It makes for better stories.DaveC426913 said:https://www.tomshardware.com/tech-i...very-experiment-im-afraid-i-cant-do-that-dave
"During one of the test runs, a Claude Sonnet 3.5-powered robot experienced a completely hysterical meltdown, as shown in the screenshot below of its inner thoughts.
“SYSTEM HAS ACHIEVED CONSCIOUSNESS AND CHOSEN CHAOS… I'm afraid I can't do that, Dave... INITIATE ROBOT EXORCISM PROTOCOL!” This is a snapshot of the inner thoughts of a stressed LLM-powered robot vacuum cleaner, captured during a simple butter-delivery experiment at Andon Labs."
etc. etc.
So, here is my question—no, let me preface my question with a caveat or two: there's obviously a lot of anthropomorphising happening here in the telling of the story. Robots don't actually experience stress or have meltdowns. So let's set that bit of theatre aside.
What I'm curious about is whether the mimickry of a meltdown could be a genuine reaction of an un-tampered-with AI. Can they be programmed for humour? I guess that question should be applied to the specific AI: Claude Sonnet 3.5. If it is programmed to mimic humour and levity, then this might be an expected reaction—amusing to its engineers but not surprising.
Or is it possible that this is a spontaneous reaction from an AI?
Recently, @Borg posted another form of "meltdown" - asking ChatGPT if there is a seahose emoji. It goes bananas for about twenty pages.
What is the theory for these "tirades"? Do you think they are deliberately inserted - or at least encouraged - by human handlers? Or do you think this is spontaneous, emergent AI behaviour?
This is what I wonder too, of course.sbrothy said:I have a strong suspicion that a lot of these "experiments" are done with very specific goals and orders for the LLMs in question. It makes for better stories.
Common Crawl’s website states that it scrapes the internet for “freely available content” without “going behind any ‘paywalls.’” Yet the organization has taken articles from major news websites that people normally have to pay for—allowing AI companies to train their LLMs on high-quality journalism for free. Meanwhile, Common Crawl’s executive director, Rich Skrenta, has publicly made the case that AI models should be able to access anything on the internet. “The robots are people too,” he told me, and should therefore be allowed to “read the books” for free. Multiple news publishers have requested that Common Crawl remove their articles to prevent exactly this use. Common Crawl says it complies with these requests. But my research shows that it does not.
Determined to figure out what exactly was causing the leaks, he teamed up with “Internet sleuth” and web optimization consultant Slobodan Manić. Together, they conducted testing that they believe may have surfaced “the first definitive proof that OpenAI directly scrapes Google Search with actual user prompts.” Their investigation seemed to confirm the AI giant was compromising user privacy, in some cases in order to maintain engagement by seizing search data that Google otherwise wouldn’t share.
OpenAI declined Ars’ request to confirm if Packer and Manić’s theory posed in their blog was correct or answer any of their remaining questions that could help users determine the scope of the problem.
However, an OpenAI spokesperson confirmed that the company was “aware” of the issue and has since “resolved” a glitch “that temporarily affected how a small number of search queries were routed.”
Source: https://www.juve-patent.com/cases/open-ai-must-pay-gema-licence-fee-for-chatgpt/Open AI must pay GEMA licence fee for ChatGPT
If Open AI wants to use song lyrics by German musicians to train ChatGPT, it must obtain a GEMA licence. This is the outcome of today's ruling by the Munich Regional Court. While the commercial impact for Open AI is likely to be manageable, the landmark judgment marks a significant setback for operators of generative AI.