AI Assistant

[iampremo]

Brings an AI Assistant to GameMaker, ask it a question about GameMaker and it will work its machine brain to find you an answer. Initial version will be trained on the GM manual so will know about all GM functions.

[Glim888]

I am wondering what technology you will use?

[henrylkirk]

Would recommend taking a look at this GMEdit plugin my buddy made: https://github.com/jolionlands/talk-codebase

It's super new and experimental, but it's a good POC of vectorizing an entire GameMaker project and then being able to ask the AI questions about it

[Glim888]

Would recommend taking a look at this GMEdit plugin my buddy made: https://github.com/jolionlands/talk-codebase

It's super new and experimental, but it's a good POC of vectorizing an entire GameMaker project and then being able to ask the AI questions about it

I can ask it about my own project? This would help a lot by refactoring etc. Great

[W1llth0t]

.

[evantor]

I'm very hesitant on any sort of AI being shoehorned into a program I love to use. Very much against the current AI 'boom' which is more like a fad that excites shareholders. I hope this doesn't replace a well maintained and integrated manual, and isn't taking resources away from much more needed and basic features, as I don't see much of a use for this beyond 'asking' what a function or reference does.

[backYard321]

Companies like twitter, bluesky, etc are baking it into their ToS that you MUST agree to let your posts to be used by AI in order to use their service. Until you can somehow guarantee that users' will NEVER be obligated to allow their work to be used by AI to continue to have access to gamemaker, I'm never going to be comfortable with GM using any AI/machine learning whatsoever.

[gnysek]

This is not that kind of AI. It's rather Clippy from Office. You can enter question, and it would serve answers from manual and official tutorials or helpdesk articles. Probably nothing more.

It won't generate code for you, it won't analyse your projects.

[CpazR]

It would be nice to have some future-proofed assurances (maybe just a reminder if it's burried in the license somewhere) that no code is analyzed for this ai in the future without express permission.

On paper, this seems like a neat idea. But I've also seen too many companies as of late blindsiding users with ai trained on content without consent (i see sites like deviantart and twitter as exhibit A)

[backYard321]

This is not that kind of AI. It's rather Clippy from Office. You can enter question, and it would serve answers from manual and official tutorials or helpdesk articles. Probably nothing more.

It won't generate code for you, it won't analyse your projects.

I understand that; I'm talking about my concerns about how AI could be used by GM later down the line. Like @CpazR said, it would be nice to have assurances in some form.

[Hyomoto]

I'm inclined to agree with @evantor here. I misread their concerns because they opened with "the AI boom," but their follow-up concerns are realistic. Feather is not where it should be. The room editor is not where it should be. There are functions missing in GM Visual. The short version is a novelty AI-based tool that answers questions about the manual is nifty, but if that is all it does, then it is also of extremely niche use. I have a dedicated Bigsby button on my phone. I have never pressed it on purpose because it isn't more accessible or better than just pushing buttons. In fact, it's often worse.

So, the concern that a reasonably prominent feature is being added when other chunks of the program still lag behind is one that I can share: it sounds more like adding a bullet point to an investor report than a tool that can assist users or even onboard new ones. After all, it will be compared against more mature AI tools: if Chat GPT is better at this than the built-in version, it will be ignored. That lack of presence will lead to a lack of development. GM's source control is a fantastic example of this.

[mistletoe]

I've done a number of things with generative AI in early 2023, including writing a fancy-ish UI for ChatGPT we used to attempt to get it to answer certain types of targeted questions.

While I understand that "incorporate LLM into IDE" was one of the big buzzwords of 2023... I think this comes with significant risks for YYG and I'd suggest caution here.

1. A Clippy-style helper sounds great, although the pitfalls are now obvious. It really depends on the scope, and it's non-trivial to implement. Writing a UI to "talk to" a LLM isn't really hard (I implemented something including a UI in less than a month), but then the problems start showing up as you start trying to make it useful. StackOverflow and many other coder resources have had to either forbid or put extreme fencing around using generative AI, because, when it comes to code, LLMs tend to do an absolutely terrible job; "hallucinations" can make an AI trained on an IDE or codebase go sideways rather rapidly and provide users with answers that sound useful, but really aren't.

2. If it's just another way to search through the IDE's functions, how does this improve the workload at YYG? I don't think it will help much, and will probably hurt more than you think, in terms of maintenance going forward. For this kind of tech to be roughly useful, you'd still need human staff writing code documentation and examples, and they have to be written correctly and succinctly. While an LLM might be vaguely useful for writing an outline for a paper on a well-understood subject or a contract, it's utterly useless for describing a feature in a piece of software.

Please remember, this tech doesn't "think" and doesn't "understand" anything; it relies on statistical convergence concepts and attempts to find the most-likely word. Asking it to describe a feature for you accurately will take as much time and trouble as writing the actual documentation will.

And then there's the maintenance costs. I pretty much guarantee you that the first version will look OK-ish in testing, then go out to users and, well... you know how that goes, lol.

If I were you, I'd spend the resources improving Feather's utility instead (seriously, making a "make Feather fully-recurse a project" function, so that it's not bombarding us with false-positives, would take a day, tops; making it accessible to end-users through the UI would be most of the work).

Anyhow, as somebody who's already been down this rabbit-hole a bit... that's my take: caveat emptor. This isn't magic, and it doesn't work as well in technical cases as a lot of people think it will.

3. A "I think you're trying to do X, here's a suggestion" helper is vaguely plausible, if trained on enough data... but will definitely suffer from the occasional severe hallucination. Moreover, it'd have to be given explicit access to our work, which is rather concerning. I do not want a "helper" sending any part of my project across the Internet without my explicit consent. The bigger IDEs have a much-vaster resource base to build upon, and they're still having these problems.

4. The overhead from having this thing run locally isn't minor. Even a small LLM will occasionally eat GPU, and if you folks don't write its state machine exceptionally carefully, it'll be a real problem. As it is, Feather's often wasting CPU over here, because it's identifying "errors" that aren't actually errors, but represent recursion-depth problems with Feather. It keeps Feather's overhead "cheap enough", but in a bad way.

5. If there's even a hint in the EULA or the behavior of GMS going forward that suggests that you're planning to feed our work into an LLM to "improve our model" or whatever, I'm gone, and so is anybody else with a lick of sense. I'd eat the year it would take to port my project to another engine in a heartbeat if I thought my work was getting fed to a third party.

I'm sorry if this comes across as "paranoid" or rude, but this is 2024... 2023 happened, the Unity scandal happened... and nobody in their right mind has a lot of faith in companies doing the right thing re: copyrights and proprietary information when this technology is concerned and data-mining your clients is the new hotness. Don't burn us, please. The money won't be worth losing us and your reputation is all you've got.

[03jordan11]

I am very excited for this feature. I already use AI a ton when working with gamemaker, but the models aren't the best when it comes to gamemaker versus other code

[ShadowTheDragon]

Really hope this doesn't ever get added tbh, for stuff like this there is already a big existing community of GameMaker users + YouTube + the documentation. If it ever does get added, can there at least be a setting to turn it off across all projects for users who dont want or need it?