There is a point brought up in the #nvda add-on mailing list on AI slop being added to the store
-
There is a point brought up in the #nvda add-on mailing list on AI slop being added to the store.
The concern can have its merit, but all add-ons are hand-reviewed by folks at @NVAccess and I do not understand how substandard add-on can be there at the first place.
My personal philosophy is that so long as the author can be responsible to their code why need to reprimand them for using AI as their code assistance? -
There is a point brought up in the #nvda add-on mailing list on AI slop being added to the store.
The concern can have its merit, but all add-ons are hand-reviewed by folks at @NVAccess and I do not understand how substandard add-on can be there at the first place.
My personal philosophy is that so long as the author can be responsible to their code why need to reprimand them for using AI as their code assistance?My philosophy is not important here, @NVAccess should urgently write a guideline on AI assisted pr submition, issue and code contribution an Add-on authorship.
It's no longer valid to have yes or no binary stand here. People are truly using LLM to make their personal workflow for NvDA, and it's functional for them at least.
Pushing people away would not be a good strategy, the best we can do is to educate and empower on how they can use tools better.
cc @joslee @Tamasg @FreakyFwoof @ppatel -
My philosophy is not important here, @NVAccess should urgently write a guideline on AI assisted pr submition, issue and code contribution an Add-on authorship.
It's no longer valid to have yes or no binary stand here. People are truly using LLM to make their personal workflow for NvDA, and it's functional for them at least.
Pushing people away would not be a good strategy, the best we can do is to educate and empower on how they can use tools better.
cc @joslee @Tamasg @FreakyFwoof @ppatel@kaveinthran @NVAccess @joslee @Tamasg @FreakyFwoof What classify's as AI slop? If they don't do testing via PR and just let it into the store, fine. But if the addons are working and fully functional, I don't see how that classify's as Ai slop.
-
@kaveinthran @NVAccess @joslee @Tamasg @FreakyFwoof What classify's as AI slop? If they don't do testing via PR and just let it into the store, fine. But if the addons are working and fully functional, I don't see how that classify's as Ai slop.
@Orinks @NVAccess @joslee @Tamasg @FreakyFwoof I totally agree with you, ultimately is the responsibility of the author to make high-quality contribution to the ecosystem, no matter what tools they use
-
@Orinks @NVAccess @joslee @Tamasg @FreakyFwoof I totally agree with you, ultimately is the responsibility of the author to make high-quality contribution to the ecosystem, no matter what tools they use
@kaveinthran @Orinks @NVAccess @Tamasg @FreakyFwoof Some come from vibe coding, others come from more traditional AI assisted coding. Remember that AI generated code gets part of their "power" from data (or for LLM's, a large corpus) wihch includes example code scattered around the net and the world.
-
@Orinks @NVAccess @joslee @Tamasg @FreakyFwoof I totally agree with you, ultimately is the responsibility of the author to make high-quality contribution to the ecosystem, no matter what tools they use
@kaveinthran @Orinks @NVAccess @joslee @FreakyFwoof I'd say, AI slop are add-ons that the author doesn't "dogfood" themselves a bit before releasing, if they just carelessly layer on features commit after commit without tuning or fixing problems in it that could have crept in, than yeah, that's a problem. However, I think a full review process for an org as small as NVAccess may not be feasible. I've learned that what guards App Store releases is that you get to really do the internal and external testing before releasing, something that a lot of these add-ons don't even go through. So maybe there's room to say for NVDA store, if you're releasing there, you *MUST* have taken your product through cycles of testing before it is published. That removes some liability from NVaccess while keeping add-on authors responsible.
-
R relay@relay.publicsquare.global shared this topic
-
@kaveinthran @Orinks @NVAccess @Tamasg @FreakyFwoof Some come from vibe coding, others come from more traditional AI assisted coding. Remember that AI generated code gets part of their "power" from data (or for LLM's, a large corpus) wihch includes example code scattered around the net and the world.
@joslee @kaveinthran @NVAccess @Tamasg @FreakyFwoof I know how AI gets its training data.
-
@joslee @kaveinthran @NVAccess @Tamasg @FreakyFwoof I know how AI gets its training data.
@Orinks @kaveinthran @NVAccess @Tamasg @FreakyFwoof So, what can happen is that the add-on author accepts whatever is generated without realizing that: 1. someone already did it before (there are ad-dons with duplicate functionality but written quite differently), or 2. the code isn't quite formatted according to expectations from NVDA/add-on comunity (tabs for indentation versus spaces, for example).
-
@Orinks @kaveinthran @NVAccess @Tamasg @FreakyFwoof So, what can happen is that the add-on author accepts whatever is generated without realizing that: 1. someone already did it before (there are ad-dons with duplicate functionality but written quite differently), or 2. the code isn't quite formatted according to expectations from NVDA/add-on comunity (tabs for indentation versus spaces, for example).
@joslee @kaveinthran @NVAccess @Tamasg @FreakyFwoof Ah, I see. They need to start linting then, and configure the linter for what NVDA community expects re: formatting. Then make a pre-commit hook to make sure it happens on commit. That's how I have my code set up and I don't really have any special requirements, other than Python standards.
-
@joslee @kaveinthran @NVAccess @Tamasg @FreakyFwoof Ah, I see. They need to start linting then, and configure the linter for what NVDA community expects re: formatting. Then make a pre-commit hook to make sure it happens on commit. That's how I have my code set up and I don't really have any special requirements, other than Python standards.
@Orinks @joslee @kaveinthran @NVAccess @FreakyFwoof yes to precommit hooks. I use them at work all the time and they sure help kill headaches, if there are things like prettier rules, NVAccess could provide those as well so there's a bar of expectation. What I love about NVDA's sourcecode is how well commented it is, unless a new feature is in development it's very clear to get what does where, and why, so that's just extending the high standards built into the core of the source already. I believe they can do it. Along with maybe claude.md or other files that clearly instruct anyone working with an AI along with the source repo if using it for add-on development especially.