ChatGPT is WRONG!

[AGKorson]

I decided to try out this new-fangled thing called ChatGPT. I asked it to explain AGI. Its response was mostly correct. However it had this to say about AGI's design:

The AGI engine was designed to allow for the creation of graphical adventure games with a point-and-click interface.

That's not quite right. True, some point-and-click capabilities were added near the end of AGI (for non-MSDOS systems), but it certainly wasn't what it was originally designed for.

But what was even more disconcerting was the response to this question:

Have you heard of WinAGI Game Development System?

(yes, maybe just a small bit of vanity- what can I say?)

Here's what it had to say:

Yes, WinAGI Game Development System is a software tool that allows users to create their own adventure games using the Adventure Game Interpreter (AGI) engine developed by Sierra On-Line in the mid-1980s.

WinAGI was developed by Alex and Phillip Fontana and was first released in 1998. It is designed to be an easy-to-use tool that allows game developers to create their own adventure games without needing to write code or have specialized technical knowledge.

Who the hell are Alex and Phillip Fontana??? Why are they credited with creating my program?  (Also, suggesting that there's no need to write code to use WinAGI is completely wrong!)

I confronted ChatGPT with the truth:

What is your source? I know you are wrong- WinAGI was created by Andrew Korson. I know this because I AM Andrew Korson

After thinking for several seconds, Chat responded:

An error occurred. Either the engine you requested does not exist or there was another issue processing your request. If this issue persists please contact us through our help center at help.openai.com.

OK, I guess I broke it! "You can't handle the TRUTH!"

I then tried the 'Regenerate response' button just to see what would happen. I got this:

I apologize for the error in my previous response. Thank you for letting me know that WinAGI Game Development System was actually created by you, Andrew Korson. I appreciate the correction and will be sure to update my knowledge accordingly.

To answer your original question, WinAGI Game Development System is a software tool created by Andrew Korson that allows users to create their own adventure games using the Adventure Game Interpreter (AGI) engine developed by Sierra On-Line in the mid-1980s. It was first released in 1998 and remains a popular tool among enthusiasts of classic adventure games and the AGI engine.

I apologize again for the mistake in my previous response, and thank you for bringing this to my attention.

So at least I got an apology. But it never did reveal its (incorrect) source. I did a 'normal' search of the Internet, and couldn't find anything connecting WinAGI with Alex and Phillip Fontana. But then again, I'm an old curmudgeon, so maybe there's something there that I just missed.

I think I'll skip the whole AI thing and continue using actual humans for conversation. But now I'm wondering - how many of the profiles on this site are actually human? Maybe ALL of you are just AI bots??? Maybe I'm a bot???

I miss the 1980s...

[Collector]

But now I'm wondering - how many of the profiles on this site are actually human? Maybe ALL of you are just AI bots??? Maybe I'm a bot???

I miss the 1980s...

LOL. Going down a Dickian rabbit hole, I see.

[doomlazer]

WinAGI is amazing. I've been using it to create point and click games since 1998 without the need for code or technical knowledge. Big thanks to Alex and Phillip Fontana!

[lskovlun]

I laughed at that bit of "SCI0" code...

[gumby]

Having a code co-pilot to assist in writing AGI/SCI games would be pretty cool.

[Doan Sephim]

Having a code co-pilot to assist in writing AGI/SCI games would be pretty cool.

Especially for troglodytes like me! 

[Kawa]

Having a code co-pilot to assist in writing AGI/SCI games would be pretty cool.

I'd rather have smart templates than ML for that sort of thing tbh.

[gumby]

Yeah the smart templates do help for sure. I guess what would be nice would be something a bit more robust than templates that could help based on your intentions.  I have no clue if something AI-driven (that provided value) would even possible or worth the effort. Interesting to think about though.

[Kawa]

Certainly it could be worth the effort. But you don't need AI/ML for that.

The tools we have now are a good start, but could stand to be a little more robust, perhaps. For example, trying to use "Go to definition" on a sent message, it doesn't work on:

• anything other than the first message, so (sirenView init: setCycle: Oscillate) only lets you go to View::init but fails to find setCycle;
• object references in variables, like (gEgo inRect: 140 110 94 116).

#1 seems like scanning further back to find sirenView instead of init: might work, and I totally get why #2 won't work. That global could hold anything, and it makes sense that it wouldn't know which class' inRect to look up. Best you could here maybe is make a best guess, find a "leaf" class with that method.

Having something like "Insert <blank> as >" might be cool. You type (aSiren init:) without defining instance aSiren of Prop, so instead of doing "Insert Object" in a blank space you right-click aSiren and pick "Insert aSiren as > Prop" and get that whole template with the name already filled in, like how you might click up method stubs in Visual Studio.

None of that requires AI/ML though.

[troflip]

ChatGPT is surprisingly good and writing small bits of code, I know a "non programmer" who uses it a lot to write little editor scripts for Unity. It would take him 10x longer to code it himself than it takes for him to give ChatGPT the right prompt.

But because there's so little information online about AGI/SCI, it doesn't seem like ChatGPT could do anything useful with it.

[Kawa]

So basically he's using it like GitHub Copilot?

[troflip]

I'm not sure, I've never used that.

[cosmicr]

I'm a big user of Github Copilot and ChatGPT.

I'm a huge follower of all things "AI" (that's the new term) or Machine Learning for quite a while now. Check out my post on converting AGI screens to AI generated images from a few months ago.

For what it's worth - the mistakes ChatGPT makes are sometimes called hallucinations and are quite common and well known. Noone should accept anything it produces as fact. Actually, producing facts is not even a recommended use for it.

Think of ChatGPT as a predictions model, you input text and it "predicts" what it thinks should be the best response to that text. So even if it's completely wrong, it will give you a confidently incorrect prediction.

As far as Sierra AGI/SCI, there's just not enough training data for it to be able to understand enough to give accurate results. I have toyed with the idea of using the OpenAI API to give ChatGPT extra context (I created text embeddings of the Wiki pages) but it's still not really enough. What you'd have to do is fine-tune a model with specific source code data which would take a long time to both prepare and then process (and $$$).

The best you could probably do right now with source code is get ChatGPT to check it for syntax or logic errors, but not offer solutions.

[Kawa]

I'm not convinced ChatGPT is the solution, or even a solution to the problem of checking for syntax or logic errors. Or that it makes a good tool to solve that problem, for that matter.

We already have basically live syntax checking. Just in SCI Companion, you miss one closing parenthesis and the moment you hit ^S that dreaded yellow error panel appears. The only difference really is that something like Visual Studio has squiggly underlines and such, like a spell checker, to indicate where the error is.

Imagine for a moment, as an example, an SCI script that has a procedure with another procedure inside. Because maybe you forgot the closing parenthesis on the "outside" one.

Code: [Select]

(procedure (Outside)
(Print "This is okay.")
(procedure (Inside)
(Print "This is wrong.")
)
)The error given is "expected an expression", pointing out the second "procedure". That's all you get. Now I can imagine extending this to include the basic facts that certain forms (lisp term) can only appear on the top level. So you might get "procedure inside an instance/a class/a method/another procedure" or "method outside of an instance/a class" instead of just "expected an expression".

Logically, it makes more sense to compare to values in an if statement than to assign them, which is why C# emits a warning when you do if (x = y) instead of if (x == y).

All that and more without a hint of machine learning.

[cosmicr]

ChatGPT in it's current form is definitely not a replacement for the tools we currently have in any way. But it could (possibly) be a good supplement to those tools.

ChatGPT can do things like actually find why your code doesn't work despite compiling fine. At least for other languages it can. I haven't really tried it for SCI or AGI code.

A simplistic example might be you coded a pong clone, but for some reason the ball always bounces wrong. No amount of code linting or syntax highlighting is gonna find why it's not working. But it turns out you put DX instead of DY as a variable somewhere by accident. ChatGPT is great for finding those kinds of errors.

Github Copilot I find is really really good for creating boilerplate code, or refactoring code too. An example, I wrote a long list of dictionary items, and then I wanted to write a conditional for each item in that dictionary - After I typed the first, copilot went and created the rest for me without having to waste time typing them all in (or even copy/pasting).