2024-12-01 02:17:09 <KipIngram   > I completely agree with this - I've found online AI to be quite poor enough of the time to call its value into serious question.  Using it only when you're sure you can vet the answer is absolutely the way to go.  That can include TESTING the answer - it doesn't necessarily have to be something you can tell by looking is right.
2024-12-01 02:18:24 <KipIngram   > There's a book series I enjoy a lot and have read several times.  I've asked ChatGPT about it a couple of times, and it'll spew a bunch of stuff very authroitatively, and MOST of it is right, but then BAM - there will be some huge egregious error right there in the middle of it.  Something just... beyond wrong.
2024-12-01 02:18:38 <KipIngram   > I've also caught it out making algebra errors.
2024-12-01 02:19:01 <KipIngram   > That I don't have to catch by checking - I can see it's wrong by some "basic test" of some kind.
2024-12-01 02:19:36 <KipIngram   > I assume all of this stems from the fact that all it can really do is compute "most probable next words" - it doesn't really KNOW anything at all.
2024-12-01 02:19:37 <thrig       > electrical dialectical
2024-12-01 02:20:23 <KipIngram   > And re: the book series, I have to wonder if its "training" there came from things written online ABOUT the series rather than from digesting the series itself.
2024-12-01 02:20:44 <KipIngram   > And a good portion of that online commentary may be wrong in and of itself.
2024-12-01 02:21:14 <Zarutian_iPad> LargeLanguageModels are just T9 kind of crap on hyperiods
2024-12-01 02:21:36 <KipIngram   > Yes - just enormous mountains of data fed into probabilistic algorithms.
2024-12-01 02:22:09 <KipIngram   > It really undermines the "singularity" argument.
2024-12-01 02:22:47 <KipIngram   > Which I think is a crock anyway, but it's an overt indication of how implausible it is.
2024-12-01 02:22:50 <Zarutian_iPad> however as lot of ?reading comprehension? and other such excercises were part of the training set, you can do some nifty things with it
2024-12-01 02:23:13 <Zarutian_iPad> hmm... you know what a cortical column is?
2024-12-01 02:24:15 <Zarutian_iPad> a big llm is pretty much the equiv of about half a dozen in the Broca or Wittinger area of the human brain iirc
2024-12-01 02:25:20 <thrig       > cortical is the column in Excel just after corticak
2024-12-01 02:25:54 <Zarutian_iPad> it is when you use Retrival Augmented Generation plus say expert system with bayes inference engine that you might start get somewhere with this
2024-12-01 02:26:44 <Zarutian_iPad> llms are not the end all of ai research and construction.
2024-12-01 02:27:33 <Zarutian_iPad> I think Moravec had the glimmering of the right idea
2024-12-01 02:28:07 <thrig       > remember when ai was graph search, or mccarthy looking for some grant money?
2024-12-01 02:29:07 <Zarutian_iPad> nope
2024-12-01 02:31:36 <thrig       > https://www.youtube.com/watch?v=03p2CADwGF8&t=2758s
2024-12-01 04:59:41 <smlckz      > How to portably implement +TO for incrementing VALUEs, if it's possible?
2024-12-01 05:10:02 <zelgomer    > congratulations, you found why i don't like values
2024-12-01 06:17:45 <smlckz      > ah
2024-12-01 10:26:29 <lilo_booter > https://asciinema.org/a/tmbgVUkorUwThpMatRJxDdNMu <- finally working the way i wanted it :D
2024-12-01 10:33:55 <lilo_booter > totally agree on with that stuff on ai too KipIngram - similar experiences when getting it to summarise novels which i know well - trying to correct it is pointless - it just spirals away from the stuff it got initially correct
2024-12-01 13:32:05 <lilo_booter > hmm - looking again at my use of evaluate - it seems to me that this should be allowed: : y=(( noname: ]] {: F: x :} [[ ; but gforth errors on the content of the braces (F: is not defined, or x is not defined if i remove F: )
2024-12-01 13:34:32 <lilo_booter > replacing the local with an fvalue is still an option to remove that altogether, but it still feels like what i'm trying to do should be valid?
2024-12-01 16:04:48 <xentrac     > smlckz: I don't think +to can be implemented portably because postpone to is specifically prohibited
2024-12-01 17:14:07 <zelgomer    > plus you need to reference the value twice. so even if you could postpone to, the end result would look like   6 +to foo foo
2024-12-01 17:14:41 <zelgomer    > 
ACTION immediately regrets not using choo as a fake value name

2024-12-01 18:29:08 <veltas      > https://www.youtube.com/watch?v=bAq4760h5ZQ
2024-12-01 18:32:20 <veltas      > EuroForth talk by M Anton Ertl (one of the main gforth contributors, not sure if he's the maintainer)
2024-12-01 18:38:43 <veltas      > What's interesting is (spoiler) he reveals that on modern computers it's faster to implement DEFER like a classic Forth would have, with a variable representing the thing to jump to, rather than patching a dummy jump instruction
2024-12-01 18:58:22 <xentrac     > zelgomer: oh, good point
2024-12-01 18:59:05 <xentrac     > so you really need a C-like & in order to implement a C-like += (or =+) for values
2024-12-01 18:59:56 <xentrac     > if you were willing to choo choo you could choo 6 + to choo instead of 6 +to choo choo
2024-12-01 19:00:55 <xentrac     > but if you had a postponeable & you could do something like : =+ postpone & dup @ rot + swap ! ;  6 =+ choo
2024-12-01 21:11:17 <TactfulCitrus> How is everyone?