[ExI] Language models are like mirrors

spike at rainier66.com spike at rainier66.com
Mon Apr 3 00:30:38 UTC 2023



From: extropy-chat <extropy-chat-bounces at lists.extropy.org> On Behalf Of Gordon Swobe via extropy-chat


>…OpenAI claims major improvements in a parameter it calls “conciseness” in version 4. It would seem that LaMDA could use similar improvements if conciseness includes for example the capacity to distinguish between the symbol “definition” and the symbol “meaning.”





Idea: consider Eliza, see if we can find the source code of the 1970s version of that.  I saw it once a long time ago, kinda figured out what it is doing, which really is nothing magic.  It has language tricks where it finds the verb, the predicate, the subject, and uses a fairly simple algorithm, really nothing tricky at all, and a limited database of canned responses.  


Eliza was just a toy, but a fun one.  If you played with it, you could easily figure out some things, such as if you said “Eliza I sometimes worry that {yakkity yak and bla bla.}  It would come back with a canned response, such as: Anxiety is a well-known problem in our ever-increasingly complex world, as we all try to find meaning in rapidly changing times.  Why do you think you are concerned that {yakkity yak and bla bla}?


Eliza would subtly get the user talking to themselves.  But it was a cool illusion really, even after you figured out how it works.


What we are seeing with ChatGPT appears to me to be Eliza plus fifty years, with access to the internet for a database.  Its grammar is perfect, vocabulary is erudite, it sure sounds like it knows what it is doing.  Until… you ask it about stuff you really really know a lot about.  Then… the spell is broken.  It appears to me that ChatGPT is really good at looking up stuff on the internet, then forming an answer with perfect grammar and advanced vocabulary.


So… try it yourself.  Take some area which you know more about that anyone, some technical niche perhaps, something on which you are a qualified judge, hand it to ChatGPT.  It will take a shot at it and it will really sound like it knows from shinola.  


OK then, does it?







-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230402/237a4a6f/attachment.htm>

More information about the extropy-chat mailing list