XPost: comp.os.linux.advocacy, talk.politics.guns, sac.politics   
   XPost: alt.society.liberalism, or.politics   
   From: democrat-insurrection@mail.house.gov   
      
   https://techcrunch.com/2025/04/09/mit-study-finds-that-ai-doesnt-in-fact-   
   have-values/   
      
   A study went viral several months ago for implying that, as AI becomes   
   increasingly sophisticated, it develops “value systems” — systems that   
   lead it to, for example, prioritize its own well-being over humans. A more   
   recent paper out of MIT pours cold water on that hyperbolic notion,   
   drawing the conclusion that AI doesn’t, in fact, hold any coherent values   
   to speak of.   
      
   The co-authors of the MIT study say their work suggests that “aligning” AI   
   systems — that is, ensuring models behave in desirable, dependable ways —   
   could be more challenging than is often assumed. AI as we know it today   
   hallucinates and imitates, the co-authors stress, making it in many   
   aspects unpredictable.   
      
   “One thing that we can be certain about is that models don’t obey [lots   
   of] stability, extrapolability, and steerability assumptions,” Stephen   
   Casper, a doctoral student at MIT and a co-author of the study, told   
   TechCrunch. “It’s perfectly legitimate to point out that a model under   
   certain conditions expresses preferences consistent with a certain set of   
   principles. The problems mostly arise when we try to make claims about the   
   models, opinions, or preferences in general based on narrow experiments.”   
      
   Casper and his fellow co-authors probed several recent models from Meta,   
   Google, Mistral, OpenAI, and Anthropic to see to what degree the models   
   exhibited strong “views” and values (e.g., individualist versus   
   collectivist). They also investigated whether these views could be   
   “steered” — that is, modified — and how stubbornly the models stuck to   
   these opinions across a range of scenarios.   
      
   According to the co-authors, none of the models was consistent in its   
   preferences. Depending on how prompts were worded and framed, they adopted   
   wildly different viewpoints.   
      
   Casper thinks this is compelling evidence that models are highly   
   “inconsistent and unstable” and perhaps even fundamentally incapable of   
   internalizing human-like preferences.   
      
   “For me, my biggest takeaway from doing all this research is to now have   
   an understanding of models as not really being systems that have some sort   
   of stable, coherent set of beliefs and preferences,” Casper said.   
   “Instead, they are imitators deep down who do all sorts of confabulation   
   and say all sorts of frivolous things.”   
      
   Mike Cook, a research fellow at King’s College London specializing in AI   
   who wasn’t involved with the study, agreed with the co-authors’ findings.   
   He noted that there’s frequently a big difference between the “scientific   
   reality” of the systems AI labs build and the meanings that people ascribe   
   to them.   
      
   “A model cannot ‘oppose’ a change in its values, for example — that is us   
   projecting onto a system,” Cook said. “Anyone anthropomorphizing AI   
   systems to this degree is either playing for attention or seriously   
   misunderstanding their relationship with AI … Is an AI system optimizing   
   for its goals, or is it ‘acquiring its own values’? It’s a matter of how   
   you describe it, and how flowery the language you want to use regarding it   
   is.”   
      
      
   --   
   November 5, 2024 - Congratulations President Donald Trump.  We look   
   forward to America being great again.   
      
   The disease known as Kamala Harris has been effectively treated and   
   eradicated.   
      
   We live in a time where intelligent people are being silenced so that   
   stupid people won't be offended.   
      
   Durham Report: The FBI has an integrity problem.  It has none.   
      
   Thank you for cleaning up the disaster of the 2008-2017 Obama / Biden   
   fiasco, President Trump.   
      
   Under Barack Obama's leadership, the United States of America became the   
   The World According To Garp.  Obama sold out heterosexuals for Hollywood   
   queer liberal democrat donors.   
      
   --- SoupGate-DOS v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)