Post Au17caOCFsgIEAtcFU by jfrench@cupoftea.social
(DIR) More posts by jfrench@cupoftea.social
(DIR) Post #Au16p9Jc7uqW4OBB4q by futurebird@sauropods.win
2025-05-12T11:51:32Z
0 likes, 0 repeats
Which of the following are “appropriate uses” for an LLM?:1. Brainstorming a list of titles for a short story. 2. Translating a product description. 3. Fixing grammar, spelling and formatting. 4. Making a list of hotels near a given address. 5. Checking a solution to statistics homework. 6. Listing places to eat in a small town that do not serve peanuts, or places that are Kosher 7. Listing photo editors that include a certain feature. 8. Making a complaint email sound “less harsh”
(DIR) Post #Au16wch1IffvD3guLA by statsguy@mas.to
2025-05-12T11:52:51Z
0 likes, 0 repeats
@futurebird That's a trick question, right? Surely the answer is "none of them"?
(DIR) Post #Au16xbDiGFTevpRrd2 by benh@mastodon.scot
2025-05-12T11:52:52Z
0 likes, 0 repeats
@futurebird Maybe 1) brainstorming Would not trust it with any of the others
(DIR) Post #Au17A2xmwFRqRzw2zI by esvrld@normal.style
2025-05-12T11:55:17Z
0 likes, 0 repeats
@futurebird well, it can't do any of them, so whichever you care about the least, i suppose
(DIR) Post #Au17CMQRY8KzOHRS0O by chris_evelyn@troet.cafe
2025-05-12T11:55:41Z
0 likes, 0 repeats
@futurebird Some of those are great and appropriate uses for Machine Learning.For an LLM? Maybe 1. if you don't care about the ethical problems.
(DIR) Post #Au17SGPTgPQfMoZbOK by Lyle@cville.online
2025-05-12T11:58:34Z
0 likes, 0 repeats
@futurebird My reply was too long. Briefly: it depends.
(DIR) Post #Au17T6dRVjUtLBVUNU by futurebird@sauropods.win
2025-05-12T11:58:46Z
0 likes, 0 repeats
@chris_evelyn In this context I don’t I’m just asking where t can be effective. My answer is 2 and 8 ONLY. it’s not really effective or could lead to catastrophic errors in all other cases. 3 might be OK too. If you have help.
(DIR) Post #Au17caOCFsgIEAtcFU by jfrench@cupoftea.social
2025-05-12T12:00:21Z
0 likes, 0 repeats
@futurebird I think only 1 and 8.Most of the others you are asking it fact based things that you'd need to double check before use. 1 and 8 are specifically language based queries that use your own thoughts and words as a basis.
(DIR) Post #Au17eHkUWxu6zj8P4q by RogerBW@discordian.social
2025-05-12T12:00:45Z
0 likes, 0 repeats
@futurebird 1 if it ever has a lower carbon footprint than the duce you could use already without paying rent to fascists.
(DIR) Post #Au17fAK9caW6FyEeA4 by futurebird@sauropods.win
2025-05-12T12:00:56Z
0 likes, 0 repeats
@benh #1 seems innocent enough, but when LLMs create titles, they often recycle well-known and copywritten material. The suggestions are kind of useless and worse they tend to be bland.
(DIR) Post #Au17hbSBj1fGyc9h9E by Gord1i@fosstodon.org
2025-05-12T12:01:10Z
0 likes, 0 repeats
@futurebird my take would be 1, 2 (if you know the language), 3 and 8. All under supervision, of course.Credit to @simon for the heuristic of using it for things that are considerably cheaper to check than to do. I also have a more personal one of not using them for stuff that involves explicit judgement calls or factual recall
(DIR) Post #Au17igPglqTYeGV93Q by chris_evelyn@troet.cafe
2025-05-12T12:01:21Z
0 likes, 0 repeats
@futurebird 2 & 8 can suffer from hallucinations, too, that's why I excluded them - if you aren't able to do the task yourself already, you won't be able to check the outcome thoroughly enough is what I think.
(DIR) Post #Au17jytZXZukG768jw by benh@mastodon.scot
2025-05-12T12:01:47Z
0 likes, 0 repeats
@futurebird Can you turn up the randomness of the suggestions?
(DIR) Post #Au17sUkyxqltON8jei by vanderZwan@vis.social
2025-05-12T12:03:18Z
0 likes, 0 repeats
@futurebird Going to side-step the question by answering a slightly different one.If the answer is supposed to represent authoritative truth, an LLM is not a reliable tool. Which excludes points two to eight, because it involved blindly trusting the answers. Points one and eight are suggestions that one may agree with or not.Which leaves all the other considerations implied by "appropriate" (e.g. the ethics of using an LLM) that I don't want to engage with.
(DIR) Post #Au1853CshpMViCB09g by benh@mastodon.scot
2025-05-12T12:05:36Z
0 likes, 0 repeats
@futurebird @chris_evelyn Why would you want a complaint email to be less harsh? Wouldn't that just water down what you need to say?As I get older (And morph into a grumpy old git) I value clarity over politeness. Too many misunderstanding arise because we are afraid to say what we mean
(DIR) Post #Au18CKp5bitwNIWNpA by tadghcr@social.vivaldi.net
2025-05-12T12:06:52Z
0 likes, 0 repeats
@futurebird they're so untrustworthy with anything fact-based, that rules out all but the first and the last. Brainstorming titles? It's hard for me to say "ethics aside," but that's so trivial it feels like if I'm that stumped I could highlight phrases at random if there's really no way to highlight a plot point, theme, or anything else myself or from anyone I show the draft. The last is literally the only use case I've heard unqualified success for. LLMs can change the tone of anything and they seem not to alter details after reading back through. This is just what they do and have done the past few years. Wish we lived in a world that, if it devalues reading so much, at least didn't force performative reading. (Thinking about that thread where people fed LLMs bullet points for it to make a long email for the recipient to feed the long email to an LLM for bullet points. Let's just cut out the water-destroying plagiarism machine.)
(DIR) Post #Au18GKI1X3oDsYxTiS by Catarina@masto.pt
2025-05-12T12:07:36Z
0 likes, 0 repeats
@futurebird 1. To get started, sure, brainstorming is brainstorming 2. If it is only for yourself, on low stakes stuff 3. Maybe, with double checking 4-7. Nope 8. Maybe, with double checking
(DIR) Post #Au18K7u5WctEcfZhAm by sabik@rants.au
2025-05-12T12:01:39Z
0 likes, 0 repeats
@benh @futurebird @chris_evelynBrainstorming with the help of an LLM has been tested, the results are worse than without
(DIR) Post #Au18K8gec3jp3IITs8 by futurebird@sauropods.win
2025-05-12T12:08:19Z
0 likes, 1 repeats
@sabik @benh @chris_evelyn I have tried to have it give me titles suggestions and the results are so basic, bland and unaware of connotations in the most unhelpful way I’m worried I won’t ever be able to pick a good title now because it’s poisoned me like some kind of anti-matter for creativity. Like it’s negated the creative spark through sheer force of boringness.
(DIR) Post #Au18KzETFZga1EWDnk by benh@mastodon.scot
2025-05-12T12:08:20Z
0 likes, 0 repeats
@futurebird Mind you, my human brain will do of those things
(DIR) Post #Au18NRRVKuQY5fmmDg by farah@beige.party
2025-05-12T12:08:54Z
0 likes, 0 repeats
@futurebird The restaurant is absolutely no. Have been personally affected by something similar
(DIR) Post #Au18Q9ClbGSeOcBNZY by Tock@corteximplant.com
2025-05-12T12:09:22Z
0 likes, 0 repeats
@futurebird In my experience:1. Non-LLM "Random Title Generators" are just as effective.2 and 3 No argument against; this is the root of LLM, it is what it is made out of. Language Translation and Correction engines. Only instead of being specialized, it's trained on massive amounts of random text and in the cloud.4. For now Map applications do a better job. (All search engines are seeking to replace prior results with only AI.)5. Mathway still exists, I believe it is not AI. At best, some free help, at worst, $14.95 a month for "I suck at this, show me everything I'm doing wrong."6. Nothing online can do this. (Restaurants are notorious for "set it and forget it" websites, even in the age of DoorDash.) I wouldn't trust an LLM to make such a list. Especially if we're talking allergies, it could kill someone.7. Wikipedia has geeks who love putting comparison lists together of many, many topics. Photo editors being among them. https://en.m.wikipedia.org/wiki/Comparison_of_raster_graphics_editors8. I'll concede this to LLM. Tone writing again being part of the same spell check and grammar engine.What is depressing: most of these answers in the next few years will be "LLM" whether it is true or false, because static websites and free info resources are under attack from alt-right and US politicians to serve them or shut down.I'll encourage people to use free info until it's all gone. Because throwing away decades of work for "billionaire speak and spell" to serve it up instead is madness to me. It's like taking the References of an academic paper offline after publishing it.
(DIR) Post #Au18RfWsu6PGwEJlyq by benh@mastodon.scot
2025-05-12T12:09:42Z
0 likes, 0 repeats
@futurebird @sabik @chris_evelynWe are here for your brain storming you know.All we need is a prompt
(DIR) Post #Au18TqqjmpVActe2AS by noodlemaz@med-mastodon.com
2025-05-12T12:10:03Z
0 likes, 0 repeats
@futurebird @chris_evelyn use a specific translation product for 2? Mass-use LLMs aren't built as translators, and while no translation model is perfect (humans are also better at this as they can understand context, subtlety and cultural differences), it's better than using an LLM that's more likely to give you errors.
(DIR) Post #Au18UuP23sYXw5Bq9A by Tock@corteximplant.com
2025-05-12T12:10:14Z
0 likes, 0 repeats
@futurebird In my experience:1. Non-LLM "Random Title Generators" are just as effective.2 and 3 No argument against; this is the root of LLM, it is what it is made out of. Language Translation and Correction engines. Only instead of being specialized, it's trained on massive amounts of random text and in the cloud.4. For now Map applications do a better job.5. Mathway still exists, I believe it is not AI. At best, some free help, at worst, $14.95 a month for "I suck at this, show me everything I'm doing wrong."6. Nothing online can do this. (Restaurants are notorious for "set it and forget it" websites, even in the age of DoorDash.) I wouldn't trust an LLM to make such a list. Especially if we're talking allergies, it could kill someone.7. Wikipedia has geeks who love putting comparison lists together of many, many topics. Photo editors being among them. https://en.m.wikipedia.org/wiki/Comparison_of_raster_graphics_editors8. I'll concede this to LLM. Tone writing again being part of the same spell check and grammar engine.What is depressing: most of these answers in the next few years will be "LLM" whether it is true or false, because static websites and free info resources are under attack from alt-right and US politicians to serve them or shut down.I'll encourage people to use free info until it's all gone. Because throwing away decades of work for "billionaire speak and spell" to serve it up instead is madness to me. It's like taking the References of an academic paper offline after publishing it.
(DIR) Post #Au1939MyvBMiCVYgCG by benh@mastodon.scot
2025-05-12T12:16:27Z
0 likes, 0 repeats
@futurebird @sabik @chris_evelyn Maybe you really need one of those password pickers that juxtapose random words in the hope of triggering a memorable image
(DIR) Post #Au1948chBNBEAuNaeu by mojala@mastodon.online
2025-05-12T12:12:54Z
0 likes, 0 repeats
@chris_evelyn @futurebird Especially for a second language, Number 8 is great stuff. Also 2 to get a bit further if needed. For 1 - no, but for project acronyms? Excellent stuff because triteness and trivialism is essential, the opposite of creative writing.
(DIR) Post #Au1949PGGo1obX6NMG by futurebird@sauropods.win
2025-05-12T12:16:30Z
0 likes, 0 repeats
@mojala @chris_evelyn This is a fair point it would probably be good for something like a course title where you want it to be similar to the titles of other courses doing the same thing. It very good at erasing the unexpected or unusual. And that’s what you want for things like spelling.
(DIR) Post #Au197Mnavwjd0d17Pk by alec@perkins.pub
2025-05-12T12:17:09Z
0 likes, 0 repeats
@futurebird @sabik @benh @chris_evelyn Sounds exactly right for a pool of averaged text sludge
(DIR) Post #Au198oPxjVJ8xOeW4e by LJ@zirk.us
2025-05-12T12:17:28Z
0 likes, 0 repeats
@futurebird maybe #2. If I had no other option & needed the translation urgently.
(DIR) Post #Au19GUGAZ0IykrjPtY by WeirdWriter@caneandable.social
2025-05-12T12:18:44Z
0 likes, 0 repeats
@futurebird Since there isn’t a none option, if I had to choose any of them, I would say number one. All the rest seem so it would be more trouble than it’s worth checking and triple checking and making sure the translation is correct, and all those other things. And maybe number eight if I wanted to sound like a corporate drone and couldn’t be bothered to learn how to write in corporate language, but even that would take more time an effort then it’s worth. All the rest would require some sort of fact checking, editing checking, to the point where it would be a literal waste of my time to use the LLM. Number one really seems to be the only one where I feel like it could maybe save me some time? But then I’d need to check and make sure I’m not plagiarizing anything so I don’t know
(DIR) Post #Au19ME9lLei2Y5TKKm by noodlemaz@med-mastodon.com
2025-05-12T12:11:36Z
0 likes, 0 repeats
@benh @futurebird @chris_evelyn I have a friend who uses it to rewrite their emails because as an autistic person, they don't like people's reactions to their directness.I would argue that's a societal problem and one of ableism, and continuing to mask through 'toning down' just perpetuates the problem for all of us.
(DIR) Post #Au19MFFTHurzw59Os4 by chris_evelyn@troet.cafe
2025-05-12T12:14:42Z
0 likes, 0 repeats
@noodlemaz @benh @futurebird But how does your friend know that the rewritten email has the right tone?
(DIR) Post #Au19MGDNhAnAvtBFFg by futurebird@sauropods.win
2025-05-12T12:19:53Z
0 likes, 0 repeats
@chris_evelyn @noodlemaz @benh You read it and see if it’s better. It is difficult to write in a polite tone if you don’t feel very polite. I don’t agree that it’s good for those working in a second language without confidence- it will probably be very transparent that it isn’t your own voice (though there are cases where this might not matter)
(DIR) Post #Au19NNZloQhS3JBbEG by Myoldpiano@mastodon.online
2025-05-12T12:20:07Z
0 likes, 0 repeats
@futurebird @sabik @benh @chris_evelyn “Sheer Force of Boringness” could be a brilliant title for the right sort of reader.
(DIR) Post #Au19YpgNeXrCPPju6a by mojala@mastodon.online
2025-05-12T12:22:10Z
0 likes, 0 repeats
@futurebird @chris_evelyn I use 3, too. I record myself blasting about something, push the transcript to a llm and prompt it to edit it to article format of (insert channel, style and character limit).
(DIR) Post #Au19bo1gGgE08l7Vqq by RedRobyn@mastodon.nz
2025-05-12T12:22:38Z
0 likes, 0 repeats
@futurebird 9. Summarizing a scientific paper in easy to understand language to help explain it to a patient
(DIR) Post #Au19dDwPP0IOntKXqq by PizzaDemon@mastodon.online
2025-05-12T12:23:00Z
0 likes, 0 repeats
@futurebird one and eight
(DIR) Post #Au19f8d27oVMuGBFfk by noodlemaz@med-mastodon.com
2025-05-12T12:23:20Z
0 likes, 0 repeats
@futurebird @chris_evelyn @benh but that's the point, if you can't judge a good tone when you write what you think, how are you judging what's 'better'?I don't agree with this usage either. I don't think it helps anyone.
(DIR) Post #Au19qmWG9W5PoKZ2Lw by suetanvil@freeradical.zone
2025-05-12T12:25:23Z
0 likes, 0 repeats
@futurebird Search engines will always do a better job than LLMs, even including the new intentionally-broken Google.
(DIR) Post #Au1ABOqfM5m6ThwNai by BLTpizza@mastodon.social
2025-05-12T12:29:08Z
0 likes, 0 repeats
@futurebird 2, 4, 8
(DIR) Post #Au1BL2iQbEXmgUFWSG by driftingThoughts@mastodon.online
2025-05-12T12:42:05Z
0 likes, 0 repeats
@futurebird 2, 3, 4, 8...as per me.Not saying that they are appropriate but they are the ones that I might use it for.If I already know a thing and I don't mind slight inaccuracies I will go for an LLM.If it is something important or very new to me I prefer doing my own searches.
(DIR) Post #Au1BpEpa7SZkwlzfVI by llewelly@sauropods.win
2025-05-12T12:47:34Z
0 likes, 0 repeats
@futurebird a decade ago I'd have said translation and making an email less harsh were the only appropriate uses.Now, I feel even those two are a lot like using an SUV to commute to work. It might seem normal and comforting, but it's horrible for society and for the environment.
(DIR) Post #Au1BrXwc1kASnzNnoe by bruce@darkmoon.social
2025-05-12T12:47:59Z
0 likes, 0 repeats
@futurebird #6 could get someone killed
(DIR) Post #Au1BtcNIWfZL7oqvBI by tshirtman@mas.to
2025-05-12T12:48:21Z
0 likes, 0 repeats
@futurebird to me 3 seems the easiest to argue for, you should still proofread to ensure it didn't change the meaning, but it's also great to force you to read rather than glancing through. I've been thinking about making a pre-commit hook to pass over my commit messages using a local model.2 i'd say some translation is usually better than no translation, but make it clear it's generated, in the UI, if it's not verified by a human.
(DIR) Post #Au1DadLIu3hEX6zlWi by Burn_this_@beige.party
2025-05-12T13:07:19Z
0 likes, 0 repeats
@futurebird @sabik @benh @chris_evelyn "Ants Eat Health Insurance Executives" would sell millions.
(DIR) Post #Au1DoU0QPiOjd6pik4 by xinniw@post.lurk.org
2025-05-12T13:09:02Z
0 likes, 0 repeats
@futurebird 8 is kind of the only one and personally I trust my own sense of communication way more than I trust the machine's output. 7 might be low risk but if I repeated the information without independently verifying I would feel like a fool.All of the others have failure modes that aren't acceptable. The only way it would be safe would be if I was already capable of doing the task myself and willing to basically do most of it anyway in order to check the work.
(DIR) Post #Au1EyMd8n52nbZDYjg by btuftin@social.coop
2025-05-12T13:22:47Z
0 likes, 0 repeats
@futurebird None of them as long as we don't know what the price is/will be.
(DIR) Post #Au1GzsTe4PNQZecrLc by sknob@mamot.fr
2025-05-12T13:45:27Z
0 likes, 0 repeats
@futurebird There are no appropriate, or to be more precise, ethical uses of LLMs, since at the very least, any use legitimizes LLMs, which are inherently unethical (trained on stolen data, environmental catastrophes, concentrate power in the hands of fasciste, devalue art and creative pursuits, etc., etc., etc.)
(DIR) Post #Au1HocFaA1QT14F3Ca by timsk@mastodon.social
2025-05-12T13:54:37Z
0 likes, 0 repeats
@futurebird I would trust GenAI to make a reasonable job of no.2. The rest of the list though, not so much. (and I say this as a GenAI-skeptic multi-lingual former student of computational linguistics, who has been arguing for decades that the roles of translator and interpreter are tragically undervalued and misunderstood)
(DIR) Post #Au1HuhgCt3jJaob8Xg by michael_w_busch@mastodon.online
2025-05-12T13:55:46Z
0 likes, 0 repeats
@futurebird None of the above.And especially not 6.
(DIR) Post #Au1KUWCjEIu8Ks10oC by NearerAndFarther@techhub.social
2025-05-12T14:24:36Z
0 likes, 0 repeats
@futurebird My take, focusing just on quality of output, and assuming a decent prompt:#1 - sure but wouldn't recommend (results are certain to be boring, although with a bit of back and forth you might get somewhere)#2 - sure, for personal use. results are minimally equivalent to google translate, often better -- just like with Translate, be alert to possible errors#3 - sure, but late in revisions, not from a first to second draft, and better for someone fairly confident in their own writing *style* (if not their grammar etc)#4 - with a web-capable LLM, sure but seems like overkill#5 - Sure, but confidence level is important here again. I'd also encourage a student to ask for an explanation of the output, engage with it past a "right or wrong" answer#6 - no way#7 - with a web-capable LLM, sure -- but it probably wouldn't be my one and only source#8 - maybe, but only for a native speaker or speaker with high proficiency -- mainly bc result is likely to sound artificial
(DIR) Post #Au1LjkGrJR9cGCluzI by Ascaso@mamot.fr
2025-05-12T14:38:36Z
0 likes, 0 repeats
@futurebird None of the above.
(DIR) Post #Au1TWNwHpV10rtI1Pk by lffontenelle@mastodon.social
2025-05-12T16:05:49Z
0 likes, 0 repeats
@futurebird IMO1. If you are ready to not picking any of the alternatives2. If readers are okay with unreliable descriptions or you're just saving keystrokes before revising3. If the fixes are not mission critical and/or you are going to check them later
(DIR) Post #Au1TqDuioSblZE18k4 by jplebreton@mastodon.social
2025-05-12T16:09:25Z
0 likes, 0 repeats
@futurebird re: 1, i've never seen one give anything that isn't trite garbage. it's the "nailed it" of people with subzero standards.
(DIR) Post #Au1VB6VDfAiEcIbpxo by KatS@chaosfem.tw
2025-05-12T16:24:23Z
0 likes, 0 repeats
@futurebird Honestly, I'm not sure I'd trust an LLM with any of those.ML might do a decent job of some, though.
(DIR) Post #Au1cFVeEoXOvRQsLBo by Oneironaut@infosec.exchange
2025-05-12T11:54:31Z
0 likes, 0 repeats
@benh @futurebird what’s the point of brainstorming if the ideas dont come from your brain
(DIR) Post #Au1cFWfh0c9ucEZ160 by benh@mastodon.scot
2025-05-12T11:56:55Z
0 likes, 0 repeats
@Oneironaut @futurebird They don't have to _come_ from your brainThey just have to be considered by your brain
(DIR) Post #Au1cFXZheMxhPwlkOm by david_chisnall@infosec.exchange
2025-05-12T11:59:42Z
0 likes, 1 repeats
@benh @Oneironaut @futurebird Rubber duck debugging works well in a lot of cases: explaining the problem often helps you come up with a solution. But, as I said in another thread, the difference between an LLM and a rubber duck is that the rubber duck knows to shut up when it has nothing of value to say.
(DIR) Post #Au1cFc3cz3H9Junhb6 by benh@mastodon.scot
2025-05-12T11:59:05Z
0 likes, 0 repeats
@Oneironaut @futurebird I can't draw. But I can scribble.And when I have scribbled enough (in pencil), I can look at the page to see it I can discern any images. And if I can then I go over them in pen, and say that is the beginning of a picture
(DIR) Post #Au1cFhSysDWi5zhWJE by benh@mastodon.scot
2025-05-12T11:59:38Z
0 likes, 0 repeats
@Oneironaut @futurebird The large language model is just automating the scribbles
(DIR) Post #Au1lxN7sl629LAjPW4 by arbeitstitel@nrw.social
2025-05-12T19:32:23Z
0 likes, 0 repeats
2, 3, 8 for 2nd language texts. ... maybe 1 in case of a "writer's block"@futurebird
(DIR) Post #Au1wpK1xU0D8v2KBzk by dpnash@c.im
2025-05-12T21:34:12Z
0 likes, 0 repeats
@futurebird Putting aside the known ethical issues with LLMs as they stand now, which are not trivial (plagiarism, pushing people out of specific kinds of work, etc):1, possibly 3, possibly 8. These are the only ones that (IMO) satisfy my main criterion: “factual accuracy of the output doesn’t matter for this specific case.” Case 1 often just requires getting some words or thoughts on paper (or screen). Case 3 involves language features that LLMs are capable of picking up on statistical grounds. Case 8 involves matters of opinion. The rest all require knowing what a correct answer is, which LLMs are not reliable for.
(DIR) Post #Au2VUp2wT66bQYkGP2 by datatrash@mastodon.social
2025-05-13T04:02:40Z
0 likes, 0 repeats
@futurebird 1 might work. results must be thoroughly scrutinised.2, 3 might save some time. but you have to be able to do it yourself to check the results for mistakes. could also cost time.4 it depends. might work, might as well be very prone to errors. there are better ways.5 not really. have to be able to check for mistakes. if you are, it makes more sense to do it yourself.6, 7 might help. results have to be checked, not trusted. might also cost you time.8 maybe. needs scrutiny