Post B4pboRqvHfeNpgYEFs by rl_dane@polymaths.social
 (DIR) More posts by rl_dane@polymaths.social
 (DIR) Post #B4pSJswXc2pZHvWxI8 by HeptaSean@social.tchncs.de
       0 likes, 1 repeats
       
       Even after a couple of years, it still feels odd to me that the AI people configure their AIs by _prompts_, by actually trying to _tell_ them to do no harm, not lie, …These things are still just statistical parrots trying to find a likely continuation of a text/dialogue. And talking to them as if they actually understand what they are doing is the most appropriate/efficient way to tailor them?That's all such a mess.
       
 (DIR) Post #B4pT0iQSalJl7Ywq24 by rl_dane@polymaths.social
       0 likes, 0 repeats
       
       @HeptaSeanThey could have been designed to be ethical and truthful from the very beginning.Instead, we got lying, golden-throated digital yes-men. :Pcc: @mirabilos
       
 (DIR) Post #B4pT0ig3elnJtwFILQ by mirabilos@toot.mirbsd.org
       0 likes, 0 repeats
       
       @rl_dane @HeptaSean they really could not
       
 (DIR) Post #B4pboRqvHfeNpgYEFs by rl_dane@polymaths.social
       0 likes, 0 repeats
       
       @mirabilos @HeptaSeanSure they could. They just wouldn't have been as powerful and money-magnety.Example:https://en.wikipedia.org/wiki/Cyc#OpenCycDisclaimer: I'm talking about "A.I." in general and not specifically LLMs.LLMs could be ethical if they were trained and run ethically, and always came with a warning that said, "This is a chatbot, it doesn't actually know anything," or something like that.
       
 (DIR) Post #B4pboS2caB0YPy1ZUO by HeptaSean@social.tchncs.de
       0 likes, 0 repeats
       
       @rl_dane @mirabilos On the one hand, that is the completely different _other_ strand of “AI” research and doesn't have that much to do with the currently hyped systems.(In fact, when I studied computer science in the early 2000s, we had two AI chairs, one doing knowledge-based systems, the other neural networks. Seems as if the neural networks which evolved to ML and then to LLMs won.)On the other hand, even such a knowledge-based system does not have _inherent_ concepts of ethics and truth. They have to be given from the outside. And it's hard or impossible to protect them against someone else then using and poisoning them with something that the original creator would consider factually or morally wrong.
       
 (DIR) Post #B4pboSH9i8dN92pB8y by rl_dane@polymaths.social
       0 likes, 0 repeats
       
       @HeptaSean @mirabilosI mean, if the claude LLMs could be trained to catch the REFUSAL magic string, couldn't all LLMs be trained to see "Are you a bot?" as a magic string and disclose that they are, indeed, a bot?That wouldn't fix all of the ethical problems, of course, but it would go a long way to fixing that particular one (lying to users, being used to dupe people, etc.)
       
 (DIR) Post #B4pboSVgq6GBs7cmnY by mirabilos@toot.mirbsd.org
       0 likes, 0 repeats
       
       @rl_dane @HeptaSean pretty sure it’s not the LLM that’s doing the stopping there