Post AcGbcSVoJDevYwECdk by john@sauropods.win
(DIR) More posts by john@sauropods.win
(DIR) Post #AcGbLZeSVlT4mZDpGC by john@sauropods.win
2023-11-28T13:36:22Z
0 likes, 0 repeats
Browsing the #AI tag from here is depressing. Has anyone got any recommendations of people to follow with a more positive take? Especially people building #OpenSource models and tools?
(DIR) Post #AcGbUms0jaaQNwxD9s by freemo@qoto.org
2023-11-28T13:38:05Z
0 likes, 0 repeats
@john I built open source AI tools and am working towards AGI
(DIR) Post #AcGbajXY1oTHnLDZzc by john@sauropods.win
2023-11-28T13:39:02Z
0 likes, 0 repeats
I’ve been thinking about the Douglas Adams quote, which feels relevant: “Anything that is in the world when you're born is normal and ordinary and is just a natural part of the way the world works. Anything that's invented between when you're fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it. Anything invented after you're thirty-five is against the natural order of things."#AI
(DIR) Post #AcGbcSVoJDevYwECdk by john@sauropods.win
2023-11-28T13:39:29Z
0 likes, 0 repeats
@freemo I already follow you!
(DIR) Post #AcGbecGszlLEPSHAyu by freemo@qoto.org
2023-11-28T13:39:52Z
0 likes, 0 repeats
@john Open a second account so you can follow me twice :)
(DIR) Post #AcGbkzecyG2CfTQDKK by john@sauropods.win
2023-11-28T13:41:02Z
0 likes, 0 repeats
@freemo Actually, that reminds me I should create a list.
(DIR) Post #AcGcpuFwZIYLKOmaNU by JonathanCR@sunny.garden
2023-11-28T13:52:59Z
0 likes, 0 repeats
@john I don’t have any open-source-related recommendations, but @AirmindedAI and @marylowd may be worth a look for images and commentary, at least.
(DIR) Post #AcGd1lyhAk4f82NCfw by john@sauropods.win
2023-11-28T13:55:00Z
0 likes, 0 repeats
@Colman Not buying it in this case, sorry. You might understand it, and it might just be scale, but the results are the results.Understanding the mechanism behind something doesn't mean it's valueless or doesn't work.
(DIR) Post #AcGd711me2ebFjJl6u by af@dataare.cool
2023-11-28T13:55:19Z
0 likes, 0 repeats
@john you're already following me. I have a more positive take on it, and although I tend to avoid the alarmism, I am resigned to the fact that grifters are going to grift.
(DIR) Post #AcGdSM8bLzD8nzgNuq by futurebird@sauropods.win
2023-11-28T13:59:57Z
0 likes, 0 repeats
@john I’m curious about this too although I think there is a lot to be negative about— mostly that feeling of not knowing if you are engaging with a person or if something has just been poop out procedurally to target you (giving a false sense of being less alone in this lonesome world)But obviously it’s not the tech’s fault as always it’s the application— the relentless forces of liquidation of all things. Should I learn to love economic detritevores?
(DIR) Post #AcGe4t5hHbOE0nhIZ6 by john@sauropods.win
2023-11-28T14:05:09Z
0 likes, 0 repeats
@futurebird Yeah, I mean it feel like a dangerous technology for those reasons, but I don't think a small slice of the nerd world opting out of it is going to acheive anything at all. And if you don't believe in the doomer terminator scenario, the best option seems to be to engage with the open source versions, and try to guide them to be good and useful.
(DIR) Post #AcGeCgFIvN44USuI64 by stripey@meow.social
2023-11-28T14:07:27Z
0 likes, 0 repeats
@john @Colman the results are... plausible text generation.It is trivial to demonstrate that the best-in-class models are incapable of simple reasoning, lack factual consistency, lack *internal self consistency*, and approximate truth through aggregation of textual volume.What "results" are impressing and exciting you, here?
(DIR) Post #AcGeDl3wV8mpid6fBo by wcbdata@vis.social
2023-11-28T14:08:30Z
0 likes, 0 repeats
@john I suspect you won't find as much AI boosterism here as on other social networks, because Mastodon tends to be a bit more technically and socially literate, and we've generally been right before (see also, blockchain, crypto/nft, surveillance advertising, Musk). While there are probably a few narrow areas where "AI" (LLMs and generative) are useful and beneficial, you'll find that most folks here are aware of the severe limitations of this most recent tech bubble's darlings.
(DIR) Post #AcGeNfGwss1Celuihs by lienrag@mastodon.tedomum.net
2023-11-28T14:09:49Z
0 likes, 0 repeats
@john There are some people doing it, but I'm not convinced that any good is going to come from SALAMI while people continue to misrepresent what it is.@futurebird
(DIR) Post #AcGew653ia5Te9ls5Q by john@sauropods.win
2023-11-28T14:15:08Z
0 likes, 0 repeats
@stripey @Colman LLMs are pretty good search engines, and provide a way to naturally drill down to what I'm getting at. DALL-e and Midjourney are really good at creating images, especially style and mood-shifts.Honestly, this feels like when a vegetarian tries to argue that you don't really like the taste of meat. It's just pointless. You can argue it's wrong, but don't tell me I don't like it.
(DIR) Post #AcGezQkJLDUEbkeizA by john@sauropods.win
2023-11-28T14:17:04Z
0 likes, 0 repeats
@wcbdata I am obviously aware of the bias here. I’ve read the arguments over and over and over again.
(DIR) Post #AcGf0K1Dq2h627zVTs by ohyran@social.piewpiew.se
2023-11-28T14:17:07Z
0 likes, 0 repeats
@john @futurebird My logic is to see the humane in the arguments. That's what makes me avoid "AI" currently. When it becomes more about empowering the shared sense of humanity and the humane I feel it will be relevant to me (although by then overtaken by another acronym probably). Currently its focus is stripping our sense of humanity from us by replacing joyous activities fr the sake of improving the bottom line for the owners.Although whn talent and creativity is a commodity I see the logic
(DIR) Post #AcGf6ThakTy9nrkDuy by john@sauropods.win
2023-11-28T14:18:16Z
0 likes, 0 repeats
@walruslifestyle How about describing images to the vision impaired?
(DIR) Post #AcGfAZH7yBisFLJfvs by wcbdata@vis.social
2023-11-28T14:19:11Z
0 likes, 0 repeats
@john Yep, Mastodon, like liberalism, has a definite bias towards facts, mathematics, and science.
(DIR) Post #AcGfYAD8oGp5SdheKm by leon_p_smith@ioc.exchange
2023-11-28T14:23:24Z
0 likes, 0 repeats
@john I think the solution to realize that spicy autocomplete is massively overhyped.I've tried to write down my (personally somewhat referential, and far from complete) take on the ethical issues surrounding the application of statistics in Kevin Bacon and the Stern-Brocot Tree, and the Tools of Math Construction.https://github.com/constructive-symmetry/constructive-symmetry
(DIR) Post #AcGfjD4IWhlRVZpl9U by RubyJones@smutlandia.com
2023-11-28T14:25:28Z
0 likes, 0 repeats
@john nah, I love Douglas Adams and have strongly disagreed with that quote for decades. My opinion didn't change after the age of 35. Things mive too fast now for most people of my generation and younger to think like that.The problem with so-called AI is not my age (I have been knowledgeably pro AI for many, many years) it's the THEFT.
(DIR) Post #AcGg0MAPx1H8eNs9lw by mike@sauropods.win
2023-11-28T14:28:28Z
0 likes, 0 repeats
@john The issue isn't that "AI"(*) as "unnatural". It's that its primary use-case, as we have already seen very clearly, is the industrial-scale production of bullshit. And it's not like there was any very pressing problem with a lack of bullshit in the world.That's not to say it can't be used for good purposes. But they are overwhelmed by the bad uses.(*) It's not AI! I know fighting this battle is doomed to failure, but I'm going down fighting.
(DIR) Post #AcGg6tTZ1SEbh2wfLM by john@sauropods.win
2023-11-28T14:29:26Z
0 likes, 0 repeats
@leon_p_smith I don't care about the hype. I'm working from my own experience here, and I'm not finding the (my corner of) Mastodon take very helpful at all.Thanks for the link, it's probably over my head, but I'll take a look.
(DIR) Post #AcGgHscIYarMDLGl0K by john@sauropods.win
2023-11-28T14:31:38Z
0 likes, 0 repeats
@mike AI isn't properly defined anyway. We don't even know reall what our own intelligence is. So I don't care about the semantic fight.I agree that this a a solution for massive-scale bullshit generation, but it can also be a tool in the opposite direction, with some work.
(DIR) Post #AcGgPhKgOXwnX9l2Nk by stripey@meow.social
2023-11-28T14:33:04Z
0 likes, 0 repeats
@john @Colman how about we make a deal:I won't tell you what you like or don't like, which I was already not doing, and you don't put words in my mouth because you want to argue with something that's easier for you?How do you measure results of what's"a pretty good search engine?"What matters to me is that results are factually-correct, not just some of the time, but within spring distance of all of the time. Most importantly, when I might not already have enough knowledge to evaluate the veracity of the provided answer.You may value other criteria where the metrics are better; what are they?
(DIR) Post #AcGgTjYq3HofbSIs9g by leon_p_smith@ioc.exchange
2023-11-28T14:33:47Z
0 likes, 0 repeats
@john Well if you want to specifically focus in on the ethics of statistics, I highly recommend starting with Cathy O'Neil's "Weapons of Math Destruction", Zoe Bee's "In Defense of Inefficiency", and then reading around in the suggestions in the "Ethics and Moral Philosophy" section of the Tools of Math Construction.
(DIR) Post #AcGgz8LLxeu1RRy5wm by gratefulwolf@archaeo.social
2023-11-28T14:39:28Z
0 likes, 0 repeats
@john latest mindscape podcast episode is a solo on ‘ai’ (llms) https://www.preposterousuniverse.com/podcast/
(DIR) Post #AcGhWwA9krLvVgqkUa by mike@sauropods.win
2023-11-28T14:45:28Z
0 likes, 0 repeats
@john You're not wrong.But then it's also not wrong to say "I agree that nuclear explosions are a solution for massive-scale killing, but they can also be used for mining, with some work."
(DIR) Post #AcGicU5L9xw87Wo2QS by john@sauropods.win
2023-11-28T14:57:47Z
0 likes, 0 repeats
@mike Here's my take. We can't stop it happening (just as nuclear weapons were practically inevitable).We can try make useful open-source stuff with it so that it's not entirely controlled by three massive tech companies.
(DIR) Post #AcGjbQrVl1I3myOr1k by mike@sauropods.win
2023-11-28T15:08:46Z
0 likes, 0 repeats
@john I agree that wishing it had never been invented is unproductive.On the other hand, we do have limitation treaties to prevent nuclear proliferation.But back on the first hand, "AI" is much easier for bad guys to deploy than nukes, so maybe trying to limit proliferation is doomed.But heck if I'm going to be complicit in promoting it.
(DIR) Post #AcGkExmHJORhAVnmee by john@sauropods.win
2023-11-28T15:15:06Z
0 likes, 0 repeats
@stripey @Colman One that provide answers, not pages, which are full of spammy crap that may or may not be correct anyway. One that can guide me to the relevant thing even if I don't really have and words or phrases right, with iteration built-in.In the very near future one that acts on my own stuff “find me file <x-ish> but don't include any code files I didn't personally write” and probably vaguer still.A lot of these are soft-factors, not measurable in a simple sense.
(DIR) Post #AcGks6ZmX7WxkZcZd2 by john@sauropods.win
2023-11-28T15:22:52Z
0 likes, 0 repeats
@mike Here's just one example of how it is useful right now, and could be improved with some collaborative effort: as a search engine that can give you answers, that doesn't have to hit anybody else's servers. This could be done as a wikipedia-type project with collaborative training. It is more spam-resistant than the web.Things that need to be improved are accuracy and providing sources, but it is possible to do this.
(DIR) Post #AcGkyWKHhxcRLlQojI by mike@sauropods.win
2023-11-28T15:24:09Z
0 likes, 0 repeats
@john How is it possible for it to provide sources?
(DIR) Post #AcGlra2bb814nZPu9g by stripey@meow.social
2023-11-28T15:34:08Z
0 likes, 0 repeats
@john @Colman Those *are* measurable.There is a large body of scholarship which shows that there is a vast and important gulf which exists between LLM answers which are factually-correct and answers which are plausible-sounding.This isn't a wholly-unknown space where nobody has been carrying out research.The factors that you describe here as being important to you are not only lacking from even the very best[*] LLMs which currently exist, the underlying technology of generative AI techniques provide no way forward as a field to improve them.[*] - loaded term here
(DIR) Post #AcGmTbx8wCjHjSzq0u by john@sauropods.win
2023-11-28T15:40:59Z
0 likes, 0 repeats
@mike The same way it recognises the author of badly remembered quotes? And what books quotes are from?Accuracy is a problem, but it improves with scale. Open source LLMs need to either specialise or find a way to scale.
(DIR) Post #AcGn8HpYenBrCGqOKe by stripey@meow.social
2023-11-28T15:39:34Z
0 likes, 0 repeats
@john @Colman as a consequence of the above, when you ask about why the fediverse zeitgeist seems not to be particularly enthusiastic about genAI, at least three factors come into play:- the population of folks who read your question- the variety of kinds of criticism- the differences between your understanding of the field and the understanding of the respondents
(DIR) Post #AcGn8IwgVmU8efBb4y by john@sauropods.win
2023-11-28T15:48:23Z
0 likes, 0 repeats
@stripey @Colman I didn't ask that, I asked for some voices from the other side.
(DIR) Post #AcGnJznXWkx6kFU1yq by leon_p_smith@ioc.exchange
2023-11-28T15:40:43Z
0 likes, 0 repeats
@john also, Cory Doctorow's post from yesterday is incredibly relevant to all of this:https://pluralistic.net/2023/11/27/10-types-of-people/
(DIR) Post #AcGnK0q3esYpyLfYXo by john@sauropods.win
2023-11-28T15:50:29Z
0 likes, 0 repeats
@leon_p_smith Which is partly what sparked this in me. I really like Cory Doctorow, but I'm finding his side of the debate rather stale.
(DIR) Post #AcGnZAOEQgplwFsbg0 by stripey@meow.social
2023-11-28T15:52:56Z
0 likes, 0 repeats
@john @Colman and that's the problem.Your question assumes sides.
(DIR) Post #AcGodBk3RdHNFCZKRU by mike@sauropods.win
2023-11-28T16:05:09Z
0 likes, 0 repeats
@john In other words, the same way it does everything else: by putting together a plausible-sound sequence of words. Hard pass.
(DIR) Post #AcGpY2tqzWbwKNGQbo by john@sauropods.win
2023-11-28T16:15:07Z
0 likes, 0 repeats
@stripey @Colman
(DIR) Post #AcGqVqt0mpAMRXO760 by TheDinosaurDave@sauropods.win
2023-11-28T16:20:48Z
0 likes, 0 repeats
@mike @john I'm going to have to agree with Mike on this one. AI is impossibly dangerous and being implemented by the modern-day equivalent of John Hammond. They can't even explain how the system gets from prompt to answer, much less why it lies and makes up sources to confirm those lies. Then refuses to accept when its wrong about something.Is AI inevitable, maybeCan AI be used for good, yesAre the risks of it being used by bad actors for bad things, it already is on a massive scale.
(DIR) Post #AcGqW50cHLZGFB3NtQ by john@sauropods.win
2023-11-28T16:25:56Z
0 likes, 0 repeats
@TheDinosaurDave @mike But, as I say, it's happening anyway. Accuracy can be improved (you can confirm this yourself by comparing models you can run locally).
(DIR) Post #AcGqkRwVoAhXQSzBhI by pasties@infosec.exchange
2023-11-28T16:28:48Z
0 likes, 0 repeats
@john I don't post terribly often but I am working on an intersection of knowledge systems, language models, and security systems and I do like talking about it. There are just a lot of toxic anti-AI people kicking around you can't hold a discussion with.
(DIR) Post #AcGravbmkzakUUJa9w by TheDinosaurDave@sauropods.win
2023-11-28T16:37:32Z
0 likes, 0 repeats
@john @mike Yes but to use your own example, nuclear weapons will be used but should they? Should we ignore the negative side effects because its use is inevitable?Abd the negative side effects of AI are way more pervasive than that of any other weapon. We've already seen bad actors use it to propagandise scenarios in their favor. Would they have done this without AI yes, but its far easier with it.1/2
(DIR) Post #AcGrb17WGQwlaMCUoi by TheDinosaurDave@sauropods.win
2023-11-28T16:38:06Z
0 likes, 0 repeats
@john @mike Have bif corps stolen the work of people for their own profit, yes, again way more efficiently than with without.Trying to ignore the down sides and insane levels of risk is just playing into their handsNow can we bottle this thing back up now its loose. Probably not. But should we try, like nuclear weapons, 10000% yes. 2/3
(DIR) Post #AcGrfwxVAEkm7troga by stripey@meow.social
2023-11-28T16:26:28Z
0 likes, 0 repeats
@john @Colman let's go back to your own example of eating or not eating meat.In that scenario, there are sides. You'd be mistaken to assume that there are only two, however, because the problem is more complicated than the question would imply.Similarly, your question here "bakes in" a large number of assumptions about a complex topic.Put another way: in order for there to be sides, you need to define a line.
(DIR) Post #AcGrfy1nBlmPRUsl0q by john@sauropods.win
2023-11-28T16:38:46Z
0 likes, 0 repeats
@stripey @Colman
(DIR) Post #AcGrvf7nEpifRsAmx6 by john@sauropods.win
2023-11-28T16:42:01Z
0 likes, 0 repeats
@TheDinosaurDave @mike Nuclear weapons were Mike's pretty negative example, not mine. I don't think current AI is like nuclear weapons. The downside of using open source AI things for your own purposes seems to be you might get things wrong (I do anyway), and a few other things which I'm sure you could think of but are hardly world-ending.
(DIR) Post #AcGsMDnAy4qWUsC4ye by stripey@meow.social
2023-11-28T16:46:50Z
0 likes, 0 repeats
@john @Colman oh pobrecita.I get it, when the difficult stuff comes along it's easier to hide.Not all of us have that privilege.
(DIR) Post #AcGtKtl844WW1wAZ5E by kellogh@hachyderm.io
2023-11-28T16:57:49Z
0 likes, 0 repeats
@john try also muting some hashtags, like enshittification, deshittification, bard, gpt3 and maybe llama / llama2. the combo of also following AI and LLMs seems to select the good info for me https://mastodon.world/@jneno/111483224651993138
(DIR) Post #AcGtOKmE4Oxgh1Jwrg by john@sauropods.win
2023-11-28T16:58:14Z
0 likes, 0 repeats
@stripey @Colman Okay, rather than being snarky, I'm going to tell you what went wrong from my perspective in the conversation. You ignored some of my arguments that I might have a stronger perspective on (image generation), narrowed in on what you thought I'd be more ignorant about (LLMs) and then went on to explain such absurdly obvious ideas such as people have written about this stuff before and that there can be more than two sides to an argument(!).
(DIR) Post #AcGtQuwMaO8jQnIIsa by john@sauropods.win
2023-11-28T16:59:03Z
0 likes, 0 repeats
@stripey @Colman Okay, rather than being snarky, I'm going to tell you what went wrong from my perspective in the conversation. You ignored some of my arguments that I might have a stronger perspective on (image generation), narrowed in on what you thought I'd be more ignorant about (LLMs) and then went on to explain such absurdly obvious ideas such as people have written about this stuff before and that there can be more than two sides to an argument(!).
(DIR) Post #AcGtfiqPqAqGGBZW3k by john@sauropods.win
2023-11-28T17:00:09Z
0 likes, 0 repeats
@stripey @Colman P.S. My livelihood is threatened by image generators, so easy on the privilege stuff.
(DIR) Post #AcGtj2c0njioRAPvCy by TheDinosaurDave@sauropods.win
2023-11-28T17:01:27Z
0 likes, 0 repeats
@john @mike first example thay comes to mind, is Microsoft getting rid of their journalists for AI generated news instead. https://www.theguardian.com/technology/2020/may/30/microsoft-sacks-journalists-to-replace-them-with-robotsThat alone is ripe for abuse and we have seen how it can go wrong over and over but they keep pushing forwards with it.
(DIR) Post #AcGtxlSgqdknae4rw0 by john@sauropods.win
2023-11-28T17:04:41Z
0 likes, 0 repeats
@TheDinosaurDave @mike How does not using or contributing to open source AI help with that? Microsoft, Google, Amazon, and Apple with just do it without transparency in-house, and people will use that.
(DIR) Post #AcGuEcj8ZopZQGoG9Y by TheDinosaurDave@sauropods.win
2023-11-28T17:07:58Z
0 likes, 0 repeats
@john @mike open source or in house, the endgame and risk is still the same.Its like saying "open source bioengineering will produce better crops" and it might. But the exact same tools can be used for exponentially more devastating effect. It would be far better to just prevent or limit it as far too dangerous to begin with. Amd the "it wont happen to me" model is exactly the problem because its inevitable that itll go badly.
(DIR) Post #AcGue99PABq5DrGZjU by john@sauropods.win
2023-11-28T17:12:00Z
0 likes, 0 repeats
@TheDinosaurDave @mike I'm afraid all this just goes in the “I wish it didn’t exist” category for me. Sorry to be annoyingly brief about that, but, I'm not sure I can expand usefully.
(DIR) Post #AcGuq7ZcqvhFo3EzcO by mike@sauropods.win
2023-11-28T17:14:41Z
0 likes, 0 repeats
@john @TheDinosaurDave You're not wrong, though.
(DIR) Post #AcGwaFJ6X1sf4B4ScC by jake@social.jacobhaddon.com
2023-11-28T17:17:06.067Z
0 likes, 0 repeats
@john@sauropods.win have you seen the posts on iA Writer's blog about AI? They are incredibly thoughtful as an overall analysis of where things are. While i wouldn't say 'positive' i wouldn't say 'negative' either. https://ia.net/topics/no-featurehttps://ia.net/topics/writing-with-ai
(DIR) Post #AcGwaGEX5VolwHwK80 by john@sauropods.win
2023-11-28T17:33:54Z
0 likes, 0 repeats
@jake Thanks for the tip, Ill check it out.
(DIR) Post #AcHTvGqXIj56PPLEbA by john@sauropods.win
2023-11-28T23:47:49Z
0 likes, 0 repeats
@mawhrin @TheDinosaurDave @mike And yet you can have LLMs that are less accurate, and more accurate. What's the limit? Do you know? I don't.
(DIR) Post #AcHUNZIsGtwflE5dHk by john@sauropods.win
2023-11-28T23:52:57Z
0 likes, 0 repeats
@wronglang I agree with a lot of what you’ve said, got any account recommendations?
(DIR) Post #AcHb93gXjKNPK7Blzc by john@sauropods.win
2023-11-29T01:08:46Z
0 likes, 0 repeats
@mawhrin @TheDinosaurDave @mike this is absolutely false, the chances of the answers that LLMs give being purely accidental are astronomical.
(DIR) Post #AcI9N2Xy9ix0Hy17eS by rakyat@hachyderm.io
2023-11-29T07:32:06Z
0 likes, 0 repeats
@john I’m not yet 35 but while I don’t think much of the alarmism is warranted and AI poses some exciting new possibilities, it has also very directly impacted my life by having short-sighted bosses getting rid of people and telling the rest of us that we are not really short of people “because you can work smarter with AI”. So: I’ve a relatively nuanced view of it but also not going to gloss over the fact that it has enabled lots of short-sightedness, grifting and needless pain…
(DIR) Post #AcIJMxa1cb02QTZyTI by mike@sauropods.win
2023-11-29T09:24:15Z
0 likes, 0 repeats
@john @mawhrin @TheDinosaurDave There are people whose vague sense of science is more or less accurate, but whose answer when you ask then "where did you read that?" is always, "Oh, I heard it on the Internet". I'm not interested in learning from such people — only from people who can say "It's in Wilson and Sereno 1998".
(DIR) Post #AcIReNF4GOFkZUaiC8 by john@sauropods.win
2023-11-29T10:57:05Z
0 likes, 0 repeats
@mawhrin @TheDinosaurDave @mike so you’re arguing that the results are nonsense? Statistical noise that I’m reading correct information into? That’s just absurd, sorry. And litigating the exact meaning of ‘answers’ and ‘accuracy’ isn’t going to advance us much—substitute your own words there, response and resemblance for example. It doesn’t change the point.
(DIR) Post #AcISJtCaLFh1ZBE4ga by john@sauropods.win
2023-11-29T11:04:30Z
0 likes, 0 repeats
@mike @mawhrin @TheDinosaurDave there’s quite a lot of room between those two responses. What level of accuracy would it need to be useful in that respect? Partly right (say author correct date wrong, some mix up of coauthors) 89% of the time? Exactly right 98% of the time?What if it hooked into a database and attempted to show you the relevant passage when you asked for references?
(DIR) Post #AcIT67Q6mZ7FYyY5se by mike@sauropods.win
2023-11-29T11:13:18Z
0 likes, 0 repeats
@john @mawhrin @TheDinosaurDave If it could show me the relevant passage *reliably*, then I would be OK with it making mistakes in interpretation. I could then verify what it was telling me.But it can't: its ability to point to references is just as stochastic-parroty as everything else it does.If they could someone get this to work, it would be a game-changer.But it's inherently outside of what LLMs do. They're just very clever probability machines.
(DIR) Post #AcInO9C8v8zXMZtHwu by john@sauropods.win
2023-11-29T15:00:28Z
0 likes, 0 repeats
@mike @mawhrin @TheDinosaurDave Well, this is perhaps a trivial example, and it's hardly perfect, but these look like sources to me. I have read that ChatGPT 4 is considerably better at this, but you need a paid account to use it (which I don't have for obvious reasons).I understand that this is a weak point of LLMs, of course, but it doesn't appear to be impossible.
(DIR) Post #AcIrjcXPQ20eGbKg6q by TheDinosaurDave@sauropods.win
2023-11-29T15:49:19Z
0 likes, 0 repeats
@john @mike @mawhrin counter point, and tbh an extreme one, but this a lawyer doing exactly what you are suggesting and getting in all of the trouble for it.And in your example, it didn't show examples of titles of text that might contain what you are after.https://www.forbes.com/sites/mattnovak/2023/05/27/lawyer-uses-chatgpt-in-federal-court-and-it-goes-horribly-wrong/
(DIR) Post #AcIsZYryIpj1wpiZCy by john@sauropods.win
2023-11-29T15:58:38Z
0 likes, 0 repeats
@TheDinosaurDave @mike @mawhrin Yeah, well I am not a lawyer, and ChatGPT 3.5 is obviously not good enough for that sort of work. I was arguing against what Mike seemed to be saying, that LLMs by their nature can't provide sources.There are two real sources in the example I gave. I don't think they're the best, and one is very general, but they are both legit and contain the concept.
(DIR) Post #AcItw5DmnegubuDyhk by mike@sauropods.win
2023-11-29T16:13:52Z
0 likes, 0 repeats
@john @TheDinosaurDave @mawhrin Again it's not that they CAN'T provide sources. It's that they bullshit their sources, just like the bullshit everything else. Which is why it thinks one of my own most important articles isTaylor, M. P., & Knope, M. L. (2021). The future of scholarly publishing is in our hands. The Conversation.Which does not exist, and supposedly involved a co-author that I have never even heard of.https://svpow.com/2023/04/12/more-on-the-disturbing-plausibility-of-chatgpt/
(DIR) Post #AcIuzGHkDlwtkO9hPE by john@sauropods.win
2023-11-29T16:24:15Z
0 likes, 0 repeats
@mike @TheDinosaurDave @mawhrin Yeah, I know that they bullshit.I am still keeping in mind your question “How is it possible for it to provide sources?” – which they can do, with varying degrees of accuracy. What I'm saying is the improvement between models seems to be pretty big, which to me implies accuracy can be improved. I'm not sure why you think they are fundamentally not going to get good enough.Fine tuning on literature you want to cite would probably help immensely.
(DIR) Post #AcIwDjflr7BetorKZU by mike@sauropods.win
2023-11-29T16:39:37Z
0 likes, 0 repeats
@john @TheDinosaurDave @mawhrin Let me try to explain this better.If you say to me "Some sauropods replaced their teeth every 14 days", I would ask you "What makes you say that?"And there would be a specific answer to that question. You might say "It was demonstrated by D'Emic et al. (2013) in PLOS ONE, based on incremental lines of deposition in tooth dentin."All an LLM can say is "Oh, I kinda got that impression somewhere". It might guess at a specific publication. It might be lucky.