Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
QB64 GPT Just Rewrote My Code
#11
(05-28-2024, 10:32 PM)Kernelpanic Wrote: Wait a minute - there's no beef ham?  Rolleyes
Not unless your beef is made out of pork.   Big Grin

https://en.wikipedia.org/wiki/Ham

Quote: Ham is pork from a leg cut that has been preserved by wet or dry curing, with or without smoking.

It's almost like saying, "A chicken leg made from beef!"   Or even as crazy as saying "almond juice" is the same thing as "almond milk."  It ain't milk, unless it comes from a titty!  Tongue
Reply
#12
(05-23-2024, 07:09 PM)SpriggsySpriggs Wrote: To be fair, I had to tell it "ByRef" was invalid and a couple of other things. I also had to declare "y" each time it was used. But the last iteration only required me to declare "y". I think that is a decent enough result. Too bad I can't get it to be this good every time.
Pretty neat! Coupla questions:

  1. This is the paid subscription for GPT4? What do they charge and how many magic question units do you get per month/week/day/unit or whatever? I'm just curious how the fee structure works.
  2. Could you maybe share the whole session, ie. what exactly did you prompt it, what did it output, what did you say next, what did it say next, etc.? It sounds like you got some useful output but the process is kinda vague.
  3. Is there a limit to the input/output length with the paid version? Back when I started playing with a free GPT3, it didn't limit how many questions you could ask, but if you told it "write me a Pong game in QBasic 4", it would output maybe 10 or 20 lines of code, then you had to say "show me the rest" and it would output the next 10 or 20 lines, and so on. And a big problem was, the 2nd block of code was part of a whole solution that might be written totally differently than the whole solution the 1st block of lines came from, so you couldn't simply ask it to tell you the rest over and over til you got the whole thing then paste it all together, because it didn't match up. I expect a paid version using the newer GPT4 version would be more complete in this respect but I don't know so am asking. 

Interesting stuff, thanks for sharing!
Reply
#13
1) This is the paid version of GPT4. I am on the plus plan, so whatever that one has.
2) I think I deleted the session. Sorry. I only used it for as long as I needed it.
3) I don't know what the hard limit is. It's in "tokens", which I have no idea how those are calculated. I got a pretty large source code from one output and it can take a lot of input. I would just say it can handle quite a bit.

The GPT I used was one I trained using the Wiki, sample code, etc. At the time, it used GPT4. Custom GPTs now use 4o. I will probably need to republish it to take advantage of 4o for it.
Tread on those who tread on you

Reply
#14
(05-30-2024, 11:35 AM)SpriggsySpriggs Wrote: 1) This is the paid version of GPT4. I am on the plus plan, so whatever that one has.
2) I think I deleted the session. Sorry. I only used it for as long as I needed it.
3) I don't know what the hard limit is. It's in "tokens", which I have no idea how those are calculated. I got a pretty large source code from one output and it can take a lot of input. I would just say it can handle quite a bit.

The GPT I used was one I trained using the Wiki, sample code, etc. At the time, it used GPT4. Custom GPTs now use 4o. I will probably need to republish it to take advantage of 4o for it.
Thanks for sharing - how does this training work anyway? Do you just feed it one or more URLs and say "study this", or upload files? What can it learn from? PDFs, image files, etc. or does it need plain text? Does training use up tokens? Does it remember what it learned across sessions? 

Sorry if these are annoying questions, I'm genuinely curious.

Another thing I'm wondering is, would it be possible to set your GPT up with an account at QBJS (assuming QBJS has a way for users to legitimately "register" their personal GPT account) so that if you ask your GPT to write code, it can actually "play test" the code it writes for you, and tweak the code until it sees output that indicates it got it right. (I haven't played with qbjs yet, so maybe you can run a local copy at 127.0.0.1 and direct your GPT to that?)
Reply
#15
I guess training is the wrong word. A custom GPT has a "knowledge base". You can have a maximum of 20 files. It can use those files to create an answer. Even a zip folder can be used. It will basically only use the knowledge base when specifically asked. Otherwise, it is using whatever it already had in its model. As for testing code and such, you can create "actions" for your GPT that allow it to do things outside of ChatGPT, including REST API. So if dbox ever made a REST API for QBJS, you could definitely have it write QBJS code and then ask it to run it.
Tread on those who tread on you

Reply
#16
Does it make a Large Language Model that can be saved/shared?  The are some iPad apps Ive seen that load custom LLM’s. Would be cool to have an offline QB64 code generator to play with.

- Dav

Find my programs here in Dav's QB64 Corner
Reply
#17
No. It does not create an offline LLM.
Tread on those who tread on you

Reply
#18
(05-30-2024, 05:43 PM)SpriggsySpriggs Wrote: I guess training is the wrong word. A custom GPT has a "knowledge base". You can have a maximum of 20 files. It can use those files to create an answer. Even a zip folder can be used. It will basically only use the knowledge base when specifically asked. Otherwise, it is using whatever it already had in its model. As for testing code and such, you can create "actions" for your GPT that allow it to do things outside of ChatGPT, including REST API. So if dbox ever made a REST API for QBJS, you could definitely have it write QBJS code and then ask it to run it.
Interesting - thanks for explaining. 

I don't know if QBJS has a REST API, maybe? (If not, maybe one could write a primitive wrapper for it?)

In any case, I don't suppose you want to throw the problem of reading the keyboard via Raw Input at it? Tongue 

(Just kidding!)
Reply




Users browsing this thread: 1 Guest(s)