Post #318190
2026-02-17 13:51 UTC
Replies (26)
-
@MachineLordZero@mastodon.social 2026-02-17 13:53
@Njord Worse than useless. A description of entropy I've heard is that if you drop a spoonful of sewage into a barrel of wine, you get a whole barrel full of sewage. However, if you drop a spoonful of wine into a barrel of sewage, it remains a barrel of sewage.
-
@len@nyan.network 2026-02-17 13:59
@Njord I don't use them for the same reason I don't use AWS and all that, but coworkers keep using them for things that I 100% know there are templates for that do the exact same thing! Because I've used those! So it just ends up being "use a thing that replaces a thing that also exists and does this same job but with a guaranteed working outcome". I guess that falls under worse than useless?? The few things that people seem to use that "win" anything fall under "let a thing run while you do other stuff", and then you still need to double triple check things... except they don't, they let it write a bullshit test suite and consider it done if it passes all the tests. Tests only mean something if they've been crafted with intent!
-
@psilocervine@gamedev.lgbt 2026-02-17 14:00
@Njord "basically just a toy" sums it up. for practical use cases like code, you do so much work checking things more complex than the most basic of boilerplate that it's a time sink and if the boilerplate is that basic you can usually just automate it. sometimes it makes funny voices. it's just... a toy, but like one that runs on coal for some reason so it poisons you while you play
-
@gbargoud@masto.nyc 2026-02-17 14:05
@Njord A hard part about this question is that the options are not mutually exclusive. Useful in limited cases can easily also mean "worse than useless in other cases"
-
@lu_leipzig@troet.cafe 2026-02-17 14:06
@Njord Worse than useless because they can reinforce false assumptions and thus be dangerous.
-
@keydelk@fosstodon.org 2026-02-17 14:06
@Njord if you ignore the ethical, environmental, and social problems of LLMs, I think they can be useful in limited situations (taking care to remember that it is often wrong and keeping skeptical of its replies). But once you factor in the ethical, environmental, and social problems, they are worse than useless. The limited benefits in a few situations are vastly outweighed by the massive harms .
-
@JeremiahFieldhaven@mastodon.gamedev.place 2026-02-17 14:11
@Njord Are you running a local model trained on specific data you control to do a constrained task? Then they can be useful. If you're talking about the ChatGPTs/etc of the world? They're not only worse than useless, they're actively harmful.
-
@unchartedworlds@scicomm.xyz 2026-02-17 14:13
@Njord Temporarily setting aside the bigger picture... I think "useful in limited cases", such as first drafts of writing which will subsequently be checked, AND "worse than useless" in other situations, such as when advising people to eat poisonous mushrooms, or making it more difficult to find accurate information, or encouraging unmaintainable code. I'd rather do my own writing as well, but not everyone's like that.
-
@RC_1290@mastodon.gamedev.place 2026-02-17 14:17
@Njord I compare the output from a local LLM (local because why pay the AI companies when they've already violated copyright on a massive scale), with random answers on random Q&A sites. You can't trust any of it. With luck it helps you find the terms and context you need to find and understand the actual documentation, but you probably shouldn't copy paste any of the code. They're decent at: Rephrasing things Reverse image search / classification("what is this") Rubber ducking
-
@gkrnours@mastodon.gamedev.place 2026-02-17 14:24
@Njord I mean, if you remove all issues from somethings, it's either neutral because there is nothing left or it's good. Is the x4 on memory and SSD prices a very serious ethical issue? Is the automated harassment of open source maintener one? Is the massive influx of security bug one? Is the poisoning of truth o the internet one? Is the rolling DDoS crawling wave one? LLM seems good at fluffying up info and summing up info. Aka giving work to itself. I put that into very serious ethical issue
-
@tiredbun@akko.wtf 2026-02-17 14:24
@Njord Ethical issues are one thing, but there are also issues that aren't ethical per say, but come from the same "big bag of issues": - Requires a lot of compute for minimum usability. Because of that, it's always a service, one that could be taken away, become worse, become unavailable temporarily or even forever. - Factual mistakes it makes may be harmful in cases where they look believable and someone relies on LLMs. - It still can and will be used to generate stuff of questionable quality/factuality in bulk, adding even more noise to social media and web as whole, both of which were already justifiably called "a big junkyard" before LLM and image generators were popular. If solutions for those issues are somehow included in theoretical solution for "ethical issues", I would choose between "Basically just a toy" and "Useful in limited cases". But I think that while those issues are connected to ethical ones, they are still separate, so I'll say "Worse than useless".
-
@laurie@hachyderm.io 2026-02-17 14:25
@Njord I put "useful in limited cases", but I'm aware that at least some of these cases are themselves caused by the rise of LLMs. For example as web search gets worse and worse, chat bots become relatively better in comparison, but the end result could still be worse than when we had working web search. Similarly the more AI generated garbage I'm presented with to read, the more need I have of a chat bot to summarise it for me.
-
@ShinNoNoir@mastodon.social 2026-02-17 14:29
@Njord I found them to be useful in a few, but limited cases, e.g.: - Knowing a certain function should exist in a language you haven't used in a while, but can't remember the name of (but you do remember its behavior). - Generating a non-trivial test example. I once wanted an example plane A and point P in 3D space with the following properties: A should have integer coefficients. P should have integer coordinates. The projection of P on A should also have integer coordinates.
-
@blaue_Fledermaus@mstdn.io 2026-02-17 14:48
@Njord If used as what they are, Media Transformers, they can be useful. Like converting scripts between bash, bat, and powershell.
-
@radicalabacus@hachyderm.io 2026-02-17 14:57
@Njord honestly, I'm conflicted: useful in limited cases and worse than useless. If you know what you're doing they can produce useful results, but they are misused in ways that are so wildly destructive that I think we'd be better off without them. Many of the things I see people use them to do can be more efficiently done some other way. Sometimes you need some bullshit and I don't think human effort should be wasted producing it; but in those situations I think it's probably wiser to take a step back and figure out what you're doing wrong that you need bullshit in the first place.
-
@stylus@social.afront.org 2026-02-17 15:25
@Njord I wish there was an ethically trained, locally usable diffusion image model. A couple of years ago I was very much into using SD to generate images to use in TTRPG, mainly character portraits. It was nice that I could do this locally, and using no more power than if I fired up a video game on my PC. In reflecting on that today, I think what was actually useful about it may have been the effort of writing the prompts. Seeing the image served as some kind of confirmation that what I was seeing in my mind's eye was a lot like what others at my table would imagine, even if I didn't share the image. Things that made it just a toy include that you couldn't do things like "now put character A and character B into setting X" and get recognizably the same people as in their original portraits. Heck, you couldn't even get the same portrait twice. Too much randomness. An LLM-adjacent tech I wonder about is cosine-similarity search. Would I benefit from this as a complicated alternative to full-text search for local code searching? Finally, I have found ente.io's local image classification ("magic search") to be quite satisfactory, once I accept that it's just a best-effort search. It makes all kinds of errors ("dog" gets you all sorts of non-dog images) and there's no way of knowing if it's comprehensive (these can't really be ALL my "macro" photos here, can they?) but it does let me quickly find relevant photos from my large collection using obvious terms. I don't know anything about the origin or training of this model and I haven't thought deeply about how flaws in the model would have real world impact. But for instance if you told me that this model turned out to have a racial bias when searching for occupation names I would not be at all surprised...
-
@witix@mk.absturztau.be 2026-02-17 15:31
@Njord@mastodon.gamedev.place It's somewhat useful for purposes it wouldn't need all the ethical breach for. For example when looking for a new headphones recently, I was out of options and didn't want to interact with the ✨ audio community ✨, so I gave my requirements to it and it spit out a few options one of which is what I'm using today and yeah I can't say I regret it. The thing is, even ignoring the fact that other options would be possible for this purpose (we should make a https://distrochooser.de for everything tbh), looking at online shop offers doesn't require stealing copyrighted material
-
@admin@mastodon.slightlycyberpunk.com 2026-02-17 15:34
@Njord As the guy who often gets roped into "fixing" (deleting and rewriting) the LLM slop of my coworkers...definitely worse than useless.
-
@iagondiscord@wetdry.world 2026-02-17 15:47
@Njord LLMs are pretty good at predicting text I wouldn't trust them to write anything of quality, let alone write anything that is actually true, but they are technically not a useless technology But this is also the absolute best case scenario; the way they are actually used and implemented in almost all real-life cases makes them toys at best and harmful at worst
-
@sidereal@kolektiva.social 2026-02-17 16:10
@Njord it’s been years and I have never been impressed by a single use case. I don’t see how people can trust any of its output now, but whatever.
-
@curtmack@floss.social 2026-02-17 17:15
@Njord From my experimenting, I would say that AI is "good" at taking something that already exists and using it to produce something of similar length or shorter (e.g. writing documentation or simple unit tests for existing code, translating or proofreading text, etc.). It's "bad" at creating new things that are larger than the input prompt (e.g. "vibe coding," art, video). And both of these are just likely outcomes, so you can get bad outcomes out of the tasks it's "good" at and vice-versa.
-
@VANTABlack2000@mastodon.gamedev.place 2026-02-17 17:24
@Njord Somewhere between a toy and limited cases. I can see some good uses when limiting the output to numbers or booleans or something like that, but GenAI is just a toy.
-
@Nawer_Rapter@mastodon.social 2026-02-17 18:01
@Njord If you could add not only ethical but also enviormental concepts to the "just imagine they don't" list, they could be maybe barely a toy. A inspiration tool, perhaps. Maybe even a somewhat acceptable...No for real the fact that it works less than 60% (generous numbers) of the time means it's pointless to even try to use it on any serious regard. Image generators MAYBE can be used as shitty assets when you're too lazy to try and find those shitty assets for free. And that's being generous.
-
@yon@sakurajima.moe 2026-02-17 20:57
@Njord I think that in some cases there could be better tools to use instead, but LLMs using their brute force method (with the lack of reliable output!) seems to have ground any such initiatives to a halt. Most of what I can (and want) to do with a code LLM I’d rather do with a better IDE that worked on an AST level. I truly think refactoring and templates was never taken to a point where they could shine.
-
@aburka@hachyderm.io 2026-02-18 03:40
@Njord I have to reject the premise of the question, the issues are simply not ignorable
-
@Voline@kolektiva.social 2026-02-24 14:11
@Njord Yesterday, I fed an audio file into a machine learning (ML) transcript program. It spit out an accurate text file quickly, with little effort from me. The ML image enhancement feature in my image editor does an excellent job. You can keep your chatbots, though. I won't tolerate their lies and resource consumption.