Path: ...!weretis.net!feeder9.news.weretis.net!usenet.blueworldhosting.com!diablo1.usenet.blueworldhosting.com!nnrp.usenet.blueworldhosting.com!.POSTED!not-for-mail From: "Edward Rawde" Newsgroups: sci.electronics.design Subject: Re: smart people doing stupid things Date: Sat, 18 May 2024 19:32:07 -0400 Organization: BWH Usenet Archive (https://usenet.blueworldhosting.com) Lines: 206 Message-ID: References: Injection-Date: Sat, 18 May 2024 23:32:09 -0000 (UTC) Injection-Info: nnrp.usenet.blueworldhosting.com; logging-data="44215"; mail-complaints-to="usenet@blueworldhosting.com" Cancel-Lock: sha1:tdNG7iG1Ld9JFPsGHbXzUZd6Vlc= sha256:LYytr5ybIPtXqP50UbcR+qF8jGuesir5IMzsYI8dWno= sha1:vQyj+9u6aT86I6jIskIyO/gOkCs= sha256:m9rA3SoEm0Evp47i1Bq23NNy4WmC8BYCXnTEHycKuuw= X-MimeOLE: Produced By Microsoft MimeOLE V6.00.2900.6157 X-Newsreader: Microsoft Outlook Express 6.00.2900.5931 X-Priority: 3 X-MSMail-Priority: Normal X-RFC2646: Format=Flowed; Response Bytes: 9852 "Don Y" wrote in message news:v2baf7$308d7$1@dont-email.me... > On 5/18/2024 7:47 AM, Edward Rawde wrote: >>>>> But, as the genie is >>>>> out of the bottle, there is nothing to stop others from using/abusing >>>>> it >>>>> in ways that we might not consider palatable! (Do you really think an >>>>> adversary will follow YOUR rules for its use -- if they see a way to >>>>> achieve gains?) >>>>> >>>>> The risk from AI is that it makes decisions without being able to >>>>> articulate >>>>> a "reason" in a verifiable form. >>>> >>>> I know/have known plenty of people who can do that. >>> >>> But *you* can evaluate the "goodness" (correctness?) of their >>> decisions by an examination of their reasoning. >> >> But then the decision has already been made so why bother with such an >> examination? > > So you can update your assessment of the party's decision making > capabilities/strategies. But it is still the case that the decision has already been made. > > When a child is "learning", the parent is continually refining the > "knowledge" the child is accumulating; correcting faulty > "conclusions" that the child may have gleaned from its examination > of the "facts" it encounters. The quality of parenting varies a lot. > > In the early days of AI, inference engines were really slow; > forward chaining was an exhaustive process (before Rete). > So, it was not uncommon to WATCH the "conclusions" (new > knowledge) that the engine would derive from its existing > knowledge base. You would use this to "fix" poorly defined > "facts" so the AI wouldn't come to unwarranted conclusions. > > AND GATE THOSE INACCURATE CONCLUSIONS FROM ENTERING THE > KNOWLEDGE BASE! > > Women bear children. > The Abbess is a woman. > Great-great-grandmother Florence is a woman. > Therefore, the Abbess and Florence bear children. > > Now, better algorithms (Rete, et al.), faster processors, > SIMD/MIMD, cheap/fast memory make it possible to process > very large knowledge bases faster than an interactive "operator" > can validate the conclusions. > > Other technologies don't provide information to an "agency" > (operator) for validation; e.g., LLMs can't explain why they > produced their output whereas a Production System can ennumerate > the rules followed for your inspection (and CORRECTION). > >>> So, you can >>> opt to endorse their decision or reject it -- regardless of >>> THEIR opinion on the subject. >>> >>> E.g., if a manager makes stupid decisions regarding product >>> design, you can decide if you want to deal with the >>> inevitable (?) outcome from those decisions -- or "move on". >>> You aren't bound by his decision making process. >>> >>> With AIs making societal-scale decisions (directly or >>> indirectly), you get caught up in the side-effects of those. >> >> Certainly AI decisions will depend on their training, just as human >> decisions do. > > But human learning happens over years and often in a supervised context. > AIs "learn" so fast that only another AI would be productive at > refining its training. In that case how did AlphaZero manage to teach itself to play chess by playing against itself? > >> And you can still decide whether to be bound by that decision. >> Unless, of course, the AI has got itself into a position where it will >> see >> you do it anyway by persuasion, coercion, or force. > > Consider the mammogram example. The AI is telling you that this > sample indicates the presence -- or likelihood -- of cancer. > You have a decision to make... an ACTIVE choice: do you accept > its Dx or reject it? Each choice comes with a risk/cost. > If you ignore the recommendation, injury (death?) can result from > your "inaction" on the recommendation. If you take some remedial > action, injury (in the form of unnecessary procedures/surgery) > can result. > > Because the AI can't *explain* its "reasoning" to you, you have no way > of updating your assessment of its (likely) correctness -- esp in > THIS instance. I'm not sure I get why it's so essential to have AI explain its reasons. If I need some plumbing done I don't expect the plumber to give detailed reasons why a specific type of pipe was chosen. I just want it done. If I want to play chess with a computer I don't expect it to give detailed reasons why it made each move. I just expect it to win if it's set to much above beginner level. A human chess player may be able to give detailed reasons for making a specific move but would not usually be aske to do this. > >> Just like humans do. >> Human treatment of other animals tends not to be of the best, except in a >> minority of cases. >> How do we know that AI will treat us in a way we consider to be >> reasonable? > > The AI doesn't care about you, one way or the other. Any "bias" in > its conclusions has been baked in from the training data/process. Same with humans. > > Do you know what that data was? Can you assess its bias? Do the folks > who *compiled* the training data know? Can they "tease" the bias out > of the data -- or, are they oblivious to its presence? Humans have the same issue. You can't see into another person's brain to see what bias they may have. > > Lots of blacks in prison. Does that "fact" mean that blacks are > more criminally inclined? Or, that they are less skilled at evading > the consequences of their crimes? Or, that there is a bias in the > legal/enforcement system? I don't see how that's relevant to AI which I think is just as capable of bias as humans are. > > All sorts of "criminals" ("rapists", "drug dealers", etc) allegedly coming > into our (US) country. Or, is that just hyperbole ("illegal" immigrants > tend to commit FEWER crimes)? Will the audience be biased in its > acceptance/rejection of that "assertion"? Who knows, but whether it's human or AI it will have it's own personality and its own biases. That's why I started this with "One thing which bothers me about AI is that if it's like us but way more intelligent than us then..." > >> Human managers often don't. Sure you can make a decision to leave that >> job >> but it's not an option for many people. >> >> Actors had better watch out if this page is anything to go by: >> https://openai.com/index/sora/ >> >> I remember a discussion with a colleague many decades ago about where >> computers were going in the future. >> My view was that at some future time, human actors would no longer be >> needed. >> His view was that he didn't think that would ever be possible. > > If I was a "talking head" (news anchor, weather person), I would be VERY > afraid for my future livelihood. Setting up a CGI newsroom would be > a piece of cake. No need to pay for "personalities", "wardrobe", > "hair/makeup", etc. "Tune" voice and appearance to fit the preferences > of the viewership. Let viewers determine which PORTIONS of the WORLD > news they want to see/hear presented without incurring the need for > a larger staff (just feed the stories from the wire services to your > *CGI* talking heads!) > > And that's not even beginning to address other aspects of the > "presentation" (e.g., turn left girls). > > Real estate agents would likely be the next to go; much of their ========== REMAINDER OF ARTICLE TRUNCATED ==========