From: gaussianblue@tilde.pink   
      
   In <87a5iyi4n3.fsf@enoch.nodomain.nowhere> Mike Spencer  writes:   
      
      
   >gaussianblue  writes:   
      
   >> harley@yazzy.com writes:   
   >>   
   >>> As for this AI stuff, what the hell is it?  I can find on the Web   
   >>> how to do Python, C++ and other languages. Where is the page   
   >>> showing what in the hell this AI stuff is?  I'm rather inclined to   
   >>> believe the whole AI thing is a hoax.   
      
   >To get a clear picture of how the underlying computations work, have a   
   >look at the "Parallel Distributed  Processing" books (Rumelhart,   
   >McClelland et al, MIT Press, 1986).   
      
   >For even further background, look at the "expanded edition" of   
   >"Perceptrons", (Minsky & Papert, MIT Press, 1988) wherein the authors   
   >confess to having gotten the future wrong in the 1969 1st edition.   
      
   >Doing this stuff in 1986 was hard because the inherent notions far   
   >exceeded couputing power of the era.  We/They have come a long way.   
   >Now they're building the vast computation into dedicated hardware, so   
   >vast that they're finding energy and heat nore critical than RAM   
   >capacity and CPU speed.   
      
   >That said, I understand the basic stuff from the 80s & 90s but I don't   
   >inderstand at the same level how the current generation of   
   >"generative" AI works, on top of that underlying tech/theory to   
   >produce what spuriously appears to be creative, thoughtfull text,   
   >audio & video and well-constructed natural language.  I don't know of   
   >a similarly accessible book that explains this most recent advance.   
      
     Interesting. Have you tried searching for scientific papers on   
   webofknowledge or scholar.google.com?   
      
   >But the lesson to take away is that the current, putaive AI is no more   
   >"intelligent", in any meaningful sense, than the models found in   
   >Rumelhart & McClelland 40 years ago.  It's the same pattern   
   >recognition (a great success) overlain with a humongusly deceptive   
   >trick to produce convincing, natural-sounding language.   
      
     I'm familiar with GNU. The idea that when dealing with software the users   
   must be able to:   
    - Run the software   
    - Share it with others (or copy)   
    - Analyze how it works (or be able to read and analyze the source code)   
    - Make changes to it (or modify the source code)   
      
     AI is still a program because it runs on a computer. I wonder how the   
   demands we are making to software when advocating for free software look   
   like when that software is AI. Especially analyzing how the   
   software works. AI is composed of a program that runs on a computer   
   and of the training data. AI is becoming more and more popular.   
   GPT can be accessed for free on the web. Some   
   of us have access to paid versions as well. Many program parts of AIs   
   are closed source. Others are free software. An example of an AI I know   
   where at least the program part is free software is llamar by superkuh:   
      
   http://superkuh.com/llamar.html   
      
     Still for many AIs the training data is closed source. But even if it   
   was free, training data is huge. You don't analyze   
   how AI works by looking at the training data the way you would look   
   at the source code of a program.   
      
   (snip)   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)