Post ASOTK1upKu6EnHVQfo by mattmoehr@zirk.us
(DIR) More posts by mattmoehr@zirk.us
(DIR) Post #ASNLO0yVmWBOFAetXc by simon@fedi.simonwillison.net
2023-02-05T16:57:52Z
0 likes, 1 repeats
If you try to use ChatGPT and similar language models as search engines they're going to lie to you, a lot, and you're at risk of writing off the whole space as hypeThe trick is to learn what they're useful for and how to take advantage of them, which is actually quite a lot of work
(DIR) Post #ASNLmLvrVxgWBw1aPg by ARmuUmjGSNRUcd7K0u.alvaro@social.graves.cl
2023-02-05T17:00:11.556Z
0 likes, 0 repeats
@simon@fedi.simonwillison.net IMHO they might be great for scaffolding code
(DIR) Post #ASNLmMh0gfOmYA5Eu0 by simon@fedi.simonwillison.net
2023-02-05T17:01:50Z
0 likes, 0 repeats
@alvaro I use them for that a lot!Can you remember how to define a Web Component in JavaScript?I can't, but ChatGPT can get that up and running for me (with an initial rough outline of the feature I describe to it) in seconds
(DIR) Post #ASNMPoSjNFLZyCDNEe by erwinrossen@mas.to
2023-02-05T17:08:55Z
0 likes, 0 repeats
@simon @alvaro I asked it for some tricks for optimizing a specific SQL query. 1 advice was useless, 1 was untrue, but the third one is actually worthwhile investigating.
(DIR) Post #ASNMb78QI4cn4D6YPA by simon@fedi.simonwillison.net
2023-02-05T17:10:13Z
0 likes, 0 repeats
@erwinrossen @alvaro that definitely matches my experience with it - loads of things where the first few results from it are unusable, but asking a few follow up questions can open up an avenue that's actually worth exploring
(DIR) Post #ASNNgKelXpSEPDHiqG by mattbee@mastodon.org.uk
2023-02-05T17:23:17Z
0 likes, 0 repeats
@simon @erwinrossen @alvaro this isn't dissimilar to how I'm conditioned to use Google - skimming domain names, checking that it's not blogspam etc. Imagine uncritically acting on its first result!
(DIR) Post #ASNODHFDFfvfNa2f9U by college_physics@defcon.social
2023-02-05T17:29:06Z
0 likes, 0 repeats
@simon now that we know that an improved search experience is possible (was not too hard to guess tbh) a key question is what does it take to replicate this in an efficient, privacy respecting way. I suspect it doesn't take a gazillion parameter model to parse a select corpus (whether its a collection of pdf, online manuals etc) and improve on something like recoll or other desktop engine
(DIR) Post #ASNOb9Oc5Eq3zJp4HQ by simon@fedi.simonwillison.net
2023-02-05T17:33:36Z
0 likes, 0 repeats
@college_physics yeah, the thing I really want to be able to do is run a smaller GPT-3-like model on my own laptop which I can then use to run this trick to safely query my own local documents: https://simonwillison.net/2023/Jan/13/semantic-search-answers/
(DIR) Post #ASNQtP3IoQTvfLgaaO by ajs@toot.community
2023-02-05T17:59:07Z
0 likes, 0 repeats
@simon I don’t think anyone serious uses LLMs directly as a search engine. Rather the intriguing bit is using the language embeddings (a byproduct of building the LLM) to implement semantic search. That’s why the OpenAI embedding API is so interesting.
(DIR) Post #ASNRMvLJpsFODABojg by simon@fedi.simonwillison.net
2023-02-05T18:04:51Z
0 likes, 0 repeats
@ajs the problem is that this isn't at all obvious when you first start interacting with tools like ChatGPTI see so many people posting screenshots of ChatGPT conversations with howlingly obvious errors in them and then using those to hint that LLMs are an over-hyped waste of timeSimilarly: don't ask a LLM a math or logic puzzle and expect it to give you a useful answer!That's very non-obvious, because everyone asumes computers are good at maths
(DIR) Post #ASNSOruTfbi6677rsW by neia@are.waffles.fun
2023-02-05T18:16:19Z
0 likes, 0 repeats
@simon I feel like that means you need a lot of familiarity with the subject matter to use it, and that familiarity means it's not nearly as useful.
(DIR) Post #ASNSfB7Y0psqJQ1572 by dabeaz@mastodon.social
2023-02-05T18:19:13Z
0 likes, 0 repeats
@simon @ajs From my experience, the primary benefit of ChatGPT is in finding better questions rather than answers. Especially if I'm in unfamiliar territory.
(DIR) Post #ASNT2XPLJeUi5vTjsm by roland@devdilettante.com
2023-02-05T18:20:32Z
0 likes, 0 repeats
@simon ROFL. This totally reminds me of Stack Overflow and Unix man pages. >60% useless unless you have many years of domain knowledge which you do :-) and I do to some extent!
(DIR) Post #ASNUFVXBh0Wu0EtdPk by jerem@indieweb.social
2023-02-05T18:36:58Z
0 likes, 0 repeats
@simon I agree. Did you compile a list of what it is useful at? Right now I’m mostly navigating it from instinct, haven’t tried to write it down…
(DIR) Post #ASNWvijvBYlZ1t6oSG by ajs@toot.community
2023-02-05T19:07:06Z
0 likes, 0 repeats
@simon I agree; it takes a bit to realise that although everything it generates seems reasonable, it can’t reason. That said, what still amazes me is that if you point out flaws it can often self-correct. ML colleagues mumble about attention mechanisms being in use as if this is completely obvious.Maybe a parallel to draw about people trusting it unreasonably is that of social media as a good news source: a lot of people can’t distinguish between the two. (And confusingly, sometimes it is!)
(DIR) Post #ASNXaFWghmMSIZFEDg by kiview@ruhr.social
2023-02-05T19:14:19Z
0 likes, 0 repeats
@simon Where I found them useful: Things I roughly know and have an easy way to assess the correctness, but which I don't commonly use. Did a terrible job finding a certain scifi book from which I only remembered the cover vaguely :D
(DIR) Post #ASNY6eaN4Q1JHlQ2Nc by simon@fedi.simonwillison.net
2023-02-05T19:20:23Z
0 likes, 0 repeats
@jerem I keep meaning to try and write something up about that, based on scrappy notes and my ChatGPT history
(DIR) Post #ASNYIycaP5kOtsKau0 by dschwarz@toad.social
2023-02-05T19:21:57Z
0 likes, 0 repeats
@simon @erwinrossen @alvaro I tried using it to figure out how to implement a particular bit of functionality using a fairly obscure Python framework. It needs to indicate some kind of confidence level about each result; several answers were absolutely definitive sounding, plausible looking, and wrong.
(DIR) Post #ASNcm1iAOTU4hM35rk by cd_home@det.social
2023-02-05T20:12:07Z
0 likes, 0 repeats
@simon @college_physics Yes, i was wondering about that. Something that has the same eloquence, but the "knowledge" of only what you show to it/allow it to use.However, i doubt that it's possible. The sample size is the basis of its eloquence.And to say "use your complete language model, but only on the following facts" implies that it "knows" what a "fact" is.
(DIR) Post #ASNd5Qo7JO2s2B5Keu by PeterBronez@hachyderm.io
2023-02-05T20:15:18Z
0 likes, 0 repeats
@simon Kagi had an experimental AI-powered search up for a while. It showed both “AI Only” and “AI+Web” results. The “AI-Web” results were fantastic. Good information with links to sources.It was like giving a clever highschooler a search term and telling them to plug it into a search engine, read the first page of results, and write a paragraph summarizing what they found.The experiment is over now, but they plan to roll out a production version soon.
(DIR) Post #ASNdo2TWcn5BbIiwa0 by Sundruid@infosec.exchange
2023-02-05T20:23:26Z
0 likes, 0 repeats
@simon like dealing with -actual- humans?
(DIR) Post #ASNe1FkWaCJeXByaHY by jwz@mastodon.social
2023-02-05T20:26:28Z
0 likes, 0 repeats
@simon Every time someone says "you've just got to learn what these glorified Markov chains are good for" it sounds a lot like "but not ALL cryptocurrencies are Ponzi schemes, some are nice"
(DIR) Post #ASNeG3bO1qSxCglFUO by simon@fedi.simonwillison.net
2023-02-05T20:29:03Z
0 likes, 0 repeats
@jwz it's weird, I'm very firmly in the "cryptocurrencies are a waste of everyone's time" camp, but the more time I spend with large language models the more convinced I am that they are going to let me solve all kinds of problems that I couldn't solve beforeNailing down exactly what those problems are is a lot more involved than I think most people expect though
(DIR) Post #ASNecnP4QAPZcnqK36 by jwz@mastodon.social
2023-02-05T20:33:01Z
0 likes, 0 repeats
@simon What kind of problems?Even if there are upsides, the downsides seem pretty severe. These systems are optimized for bullshit and lies, like, that's their core competency to which all use degrades. See also "facial recognition is the plutonium of AI". https://jwz.org/b/yjMP
(DIR) Post #ASNfYpbCPpyxELBoTw by Javi@sfba.social
2023-02-05T20:43:35Z
0 likes, 0 repeats
@simon @Migueldeicaza yeah people say that a lot and pretend that everything you find in Google is A+ quality
(DIR) Post #ASNgNQre8e5ZHm0k8u by simon@fedi.simonwillison.net
2023-02-05T20:52:58Z
0 likes, 0 repeats
@jwz Extracting structured data from unstructured text is one promising angle - I'm interested in the potential for investigative data journalism, for problems like turning 20,000 ad-hoc poor quality scanned police complaint reports into actionable information, without spending six months on human-powered data entry first
(DIR) Post #ASNgaKgIREM7sEqyWG by simon@fedi.simonwillison.net
2023-02-05T20:54:21Z
0 likes, 0 repeats
@jwz There are plenty of NLP problems like that which I've previously not been able to take on myself because they were too difficult for me to justify the time and expense, which now seem feasible if I spend a much smaller amount of time messing around with some basic prompt engineering and a large language model
(DIR) Post #ASNhBgiVTOFavlhRVQ by webology@mastodon.social
2023-02-05T20:46:23Z
0 likes, 0 repeats
@jwz @simon I would argue that extreme power usage is closer to crypto than not, but I don't think we are far off from seeing another GPU shortage as more companies adopt it. Something about how the tech relies on tokens too feels very akin to crypto even though I know there are used differently.
(DIR) Post #ASNhBhHbMuK6gc7TCy by simon@fedi.simonwillison.net
2023-02-05T20:59:49Z
0 likes, 0 repeats
@webology @jwz For me one of the biggest differences in power usage is that with language models you do at least get something in exchange for your additional power - you can run more inferences, train bigger models etcWith proof-of-work crypto the power usage is purely a competition: if you burn more power all it does is encourage other members of the network to increase their power usage to match, in order to out-spend you and win more of the mining rewards
(DIR) Post #ASNhLVw2PjvoO5iLeC by amyhoy@mastodon.social
2023-02-05T21:00:08Z
0 likes, 0 repeats
@simon @jwz without spending six months on human-powered data entry first, you will never be able to know if the GPT result is a series of hallucinations, like so much of its output
(DIR) Post #ASNhWUvYV9BSRnj4wS by simon@fedi.simonwillison.net
2023-02-05T21:02:59Z
0 likes, 0 repeats
@amyhoy @jwz I'm talking about prompts like "Here is a copy and pasted police report. Return JSON with the names of the mentioned officers and the date of the incident"My hunch here is that spot checks on the results could help tell if it's working well enough, and that the end results would reach the same level of accuracy as asking human data entry people (who are also infallible) to do the same taskIf there's a better way to do this than using a language model I'm interested to hear it
(DIR) Post #ASNhnYqVBZmpWVUyx6 by amyhoy@mastodon.social
2023-02-05T21:09:03Z
0 likes, 0 repeats
@simon @jwz having done NLP work before, that doesn’t sound at all difficult to implement directly.
(DIR) Post #ASNiQsTQ8zwNgvRySG by jwz@mastodon.social
2023-02-05T21:15:59Z
0 likes, 0 repeats
@simon @amyhoy But how could you ever trust that data? You are asking for *facts* but the system is optimized to produce *believable answers* which are not at all the same thing.Suppose the system optimizes its march to the goal by just making up some numbers that subtly (or not so subtly) tilt the data one way or another. Now you've built a black box to confirm your biases.And the black box, by its nature, cannot "show its work" without lying.
(DIR) Post #ASNlzHSA4nRAsAxQAa by simon@fedi.simonwillison.net
2023-02-05T21:56:01Z
0 likes, 0 repeats
@jwz @amyhoy The black box thing is why I'm finding this whole space so utterly beguilingI hate that it's a black box. But I've spent my entire career working with computers that do exactly what you tell them... and now I'm faced with one that very much does not do thatIt's like someone's given me a spell that raises actual dragons from another dimension and challenged me to try and tame them!
(DIR) Post #ASNmBEsLUBfgLxuIMK by jeff@indieweb.social
2023-02-05T21:57:13Z
0 likes, 0 repeats
@simon @erwinrossen @alvaro A lot of people dismiss it as if the first iteration of the technology is as good as it’d ever going to get.Getting a lot of “this iPhone thing doesn’t even support copy and paste!” vibes from a lot of criticism.
(DIR) Post #ASNmN12Am3Hktq7S1g by simon@fedi.simonwillison.net
2023-02-05T21:57:42Z
0 likes, 0 repeats
@jwz @amyhoy I can't see this tech being un-invented, so the interesting question to me now is what I can build with it now that I couldn't build before - and what are the new, genuinely valuable problems I can solve for people
(DIR) Post #ASNmaD61EsP44eD4aW by amyhoy@mastodon.social
2023-02-05T21:17:02Z
0 likes, 0 repeats
@simon @jwz yep. my motto has been “bad data is worse than no data” ever since we started a biz that relies on data entry (time tracking) 15 years ago
(DIR) Post #ASNmaDePB1uPnIIXBY by lukasb@hachyderm.io
2023-02-05T21:34:30Z
0 likes, 0 repeats
@amyhoy @simon @jwz can do a two-pass model - have the AI look for stuff, if it turns up an interesting result, have a human verify.
(DIR) Post #ASNmaE3Dgll52FuLrc by simon@fedi.simonwillison.net
2023-02-05T22:00:26Z
0 likes, 0 repeats
@lukasb @amyhoy @jwz Right: for journalism, what matters is generating leads - giving skilled human journalists a hand in digging through mountains of data and trying to find the most likely threads that can lead to a storyYears ago I did that with crowdsourcing - random human volunteers are a highly unreliable source too, but they can still help kick off the process: https://simonwillison.net/2009/Dec/20/crowdsourcing/
(DIR) Post #ASNmaGBRl9WleLlwxs by lukasb@hachyderm.io
2023-02-05T21:35:58Z
0 likes, 0 repeats
@amyhoy @simon @jwz the advantage obviously is that the AI can look for stuff wayyyy faster than people. So even with a human verifying every interesting result, you still get a benefit.
(DIR) Post #ASNmaIpZuuojrOZ1nM by lukasb@hachyderm.io
2023-02-05T21:43:56Z
0 likes, 0 repeats
@amyhoy @simon @jwz let's say that gets your false positive rate near zero. You *do* potentially have a problem if your false negative rate matters. But there are ways to sample a % of negative results for human review to mitigate that as well.
(DIR) Post #ASNmm9am5mhSgKI5GS by simon@fedi.simonwillison.net
2023-02-05T22:02:24Z
0 likes, 0 repeats
@amyhoy @jwz Using what techniques? I've played with https://www.nltk.org/ a bit but I've never gotten to the point where I'd feel like I could point it at unstructured police reports and usefully pull out the data I need from them (I'm talking WILDLY unstructured - California has 20,000+ Police departments, all with very different standards of hand-written paper work)
(DIR) Post #ASNmxPBOT18YDjmfbs by jwz@mastodon.social
2023-02-05T22:06:38Z
0 likes, 0 repeats
@simon @amyhoy I've got no time for that attitude, whether applied to dangerous software or chemical weapons. We regulate things that cause harm.
(DIR) Post #ASNnAqahsJL4zT29DM by jwz@mastodon.social
2023-02-05T22:07:53Z
0 likes, 0 repeats
@simon @amyhoy I often say that we missed our demon haunted world so much that we've spent the last 50 years reinventing animism and making it real.
(DIR) Post #ASNnKNrTXsukmcQNu4 by ian@mckellar.social
2023-02-05T22:12:37Z
0 likes, 0 repeats
@jwz@simon @amyhoy Actually, we generally don't do a good job of regulating things that hurt working class people.
(DIR) Post #ASNnO5GCe7lIFckrjs by amyhoy@mastodon.social
2023-02-05T22:08:07Z
0 likes, 0 repeats
@jwz @simon you want people to get eaten by dragons? cuz this is how you get people eaten by dragons. i’m glad you can admit it’s beguiling — as in, deceiving, by definition — and you’re attracted to it because of its very wrongness. because the next step is admitting its promises are a lie, and it is wrong.
(DIR) Post #ASNndHQ1ivL8Isb0pk by amyhoy@mastodon.social
2023-02-05T22:10:33Z
0 likes, 0 repeats
@simon @jwz the technique of looking at a ton of them and figuring out the patterns of address.i’ve only done NLP on unstructured data, what other kind is there? otherwise it’s just database stuff.
(DIR) Post #ASNndIwg313F2QsbMe by amyhoy@mastodon.social
2023-02-05T22:10:55Z
0 likes, 0 repeats
@jwz @simon we didn’t use any toolkits 😂 we did it the hard way.
(DIR) Post #ASNo7jhh7pHyiPNmUa by sayrer@mastodon.social
2023-02-05T22:21:33Z
0 likes, 0 repeats
@ian @jwz @simon @amyhoy It is actually really good at syntactic problems. That seems fine to me. If you go looking for the truth, that might be misguided.
(DIR) Post #ASNp3ARsMRWGAYMvj6 by simon@fedi.simonwillison.net
2023-02-05T22:30:06Z
0 likes, 0 repeats
@amyhoy @jwz Oh you mean like iterating on a big list of regular expressions until you catch all of the edge-cases?That's the kind of solution which for me falls into the "it's going to take too long to figure all of those out, hence these 20,000 police reports will sadly go unprocessed and this story will remain untold"I'm always looking for technology which shifts projects from "too hard/expensive/time-consuming to do" to "that's now a feasible project"
(DIR) Post #ASNpIRcEofUzkvV4r2 by simon@fedi.simonwillison.net
2023-02-05T22:32:44Z
0 likes, 0 repeats
@jwz @amyhoy I'm not opposed to regulation for this stuff at all - but in order to have effective conversations about regulation, we need to have a lot more people with a strong understanding of how it works, what it's useful for and where it absolutely should not be applied
(DIR) Post #ASNpW31Z2leZXaphrs by nikclayton@mastodon.social
2023-02-05T22:34:42Z
0 likes, 0 repeats
@simon @jwz @amyhoy Narrator: "There were no effective conversations about regulation."
(DIR) Post #ASNpiJRNpjsvvkl3cu by amyhoy@mastodon.social
2023-02-05T22:36:38Z
0 likes, 0 repeats
@jwz @simon we built ours in 6 weeks. and we have a very different definition of “feasible” if imaginary results are an acceptable outcome.
(DIR) Post #ASNpuf1JQYAsKiS3Ky by sirjofri@mastodon.sdf.org
2023-02-05T22:37:28Z
0 likes, 0 repeats
@simon I can imagine it is quite helpful if you really just use it as a search engine kinda thing: to find resources on the internet. So not to five you answers to your questions, but to help you understand the question to deliver better search results. If it works out or will be used like that is something we'll see ...I guess most people want to see AI as a full solution for a problem, but it's really just a statistical tool.
(DIR) Post #ASNqibsXNiL34LAH8C by amyhoy@mastodon.social
2023-02-05T22:39:04Z
0 likes, 0 repeats
@ian @simon @sayrer @jwz there are so many cases of it doing wrong (very basic!!) math and “explaining” why its wrong stuff is right. not to mention anything more complex than 2+2=4, like word problems. so i would say no, it isn’t very good at syntax problems.
(DIR) Post #ASNqicVX2jWx1HPPua by simon@fedi.simonwillison.net
2023-02-05T22:48:52Z
0 likes, 0 repeats
@amyhoy @ian @sayrer @jwz It's a next-token-predicting language model, so using it for math is very much the wrong application of it - that's one of the many reasons I keep trying to convince people that these things are deceptively difficult to use effectivelyThe idea that a computer can be bad at math is very counter-intuitive!
(DIR) Post #ASNqvBuHFvC0Hbyg8O by amyhoy@mastodon.social
2023-02-05T22:50:42Z
0 likes, 0 repeats
@simon @ian @jwz @sayrer that argument would hold more weight if it didn’t do the same thing in all fields.
(DIR) Post #ASNr6EKUaWaCyzVGxU by josephholsten@mstdn.social
2023-02-05T22:50:47Z
0 likes, 0 repeats
@simon @jwz @amyhoy it’s a monkey’s paw. Sure, the first few wishes will seem like it’s only misunderstanding your request because you weren’t specific enough. So you’ll just keep doubling down until it’s eaten your soul and you can’t back out because of sunk costs.
(DIR) Post #ASNrGcn8Ug9HkJrF0y by simon@fedi.simonwillison.net
2023-02-05T22:51:02Z
0 likes, 0 repeats
@sirjofri It's actually specifically terrible at finding resources on the internet - the training process used for a language model causes it to entirely lose track of where any snippets of information originally came fromIf you ask it for URLs to things it will literally make up URLs that look legitimate but are actually to non-existent pages!Same if you ask it to cite sources - it will happily invent names of non-existent academic papers and credit them to non-existent people
(DIR) Post #ASNrT0DZDm1rsP7gCO by sayrer@mastodon.social
2023-02-05T22:51:17Z
0 likes, 0 repeats
@simon @amyhoy @ian @jwz I had it write a matrix math procedural macro in Rust, using the exact crates I instructed it to use. It did not write things as I would, and wasn't even exactly right, but it took care of a lot of trivia.
(DIR) Post #ASNrfRDCgQqx4XsmAq by simon@fedi.simonwillison.net
2023-02-05T22:52:59Z
0 likes, 0 repeats
@josephholsten @jwz @amyhoy Yeah, a monkey's paw is a pretty solid analogy for helping people understand the risks involved hereSee also the magic / spell-casting analogy, which is way more effective if your mental model of magic involves mispronounced spells that accidentally summon demons!https://simonwillison.net/2022/Oct/5/spell-casting/
(DIR) Post #ASNsygMEhDyXVBqc40 by anildash@me.dm
2023-02-05T23:14:08Z
0 likes, 0 repeats
@simon @amyhoy @ian @sayrer @jwz I think a key part here is making sure those of us who are critical are specific and fluent in the systems we’re criticizing, rather than blanket dismissal that sounds glib because we’re fast-forwarding to the conclusion instead of showing our work.
(DIR) Post #ASNvArdWmrURrls8hM by anildash@me.dm
2023-02-05T23:14:28Z
0 likes, 0 repeats
@simon @amyhoy @ian @sayrer @jwz What Molly White (and to some degree, Moxie) did in breaking down the faults and flaws of crypto assertions did far more to hasten good regulation than any amount of “it a bunch of dumb scams!” ranting did. Simon’s path here seems more likely to yield effective harm reduction.
(DIR) Post #ASNvAsEOZmyri77aAC by donaldball@triangletoot.party
2023-02-05T23:27:25Z
0 likes, 0 repeats
@anildash @simon @amyhoy @ian @sayrer @jwz I guess, but I also am very much stuck on the observation that these systems *rely* on obfuscated copyright theft at scale in order to function.
(DIR) Post #ASNvAsrOEoAlf3Miwa by simon@fedi.simonwillison.net
2023-02-05T23:38:44Z
0 likes, 0 repeats
@donaldball @anildash @amyhoy @ian @sayrer @jwz I'm very much looking forward to someone training a generative text or image AI entirely on out-of-copyright works - it would be very interesting to try that out and compare it to the existing modelsA ChatGPT-like system that has a cutoff date of 1927 (and a style of writing to match) would be quite a thing!
(DIR) Post #ASNvMN7I5LwwgYHTZw by amyhoy@mastodon.social
2023-02-05T23:40:24Z
0 likes, 0 repeats
@sayrer @simon @jwz @anildash @ian @donaldball that does sound fun and, like the current model, totally unfit for any professional or otherwise meaningful purpose other than fun
(DIR) Post #ASNvYFSgN8NaQ1wx6W by sayrer@mastodon.social
2023-02-05T23:43:06Z
0 likes, 0 repeats
@simon @amyhoy @ian @jwz Think in a more evil way. Every copyrighted work is in fact exactly this. 20th century artists like DuChamp, Warhol, and Richard Prince cover it well. Let alone DJ Shadow or J Dills.
(DIR) Post #ASO1NI4uu1914dzYNU by aaron@social.huslage.com
2023-02-06T00:48:20Z
0 likes, 0 repeats
@simon @jwz @amyhoy by the time the regulations catch up, this phase of AI will have ended and a new one taken its place. The goal is constantly moving and the map changes daily. How do you regulate that?
(DIR) Post #ASO1lAJHR81Xvm5vF2 by simon@fedi.simonwillison.net
2023-02-06T00:50:26Z
0 likes, 0 repeats
@aaron @jwz @amyhoy the EU are moving on that already - they seem to be mostly thinking about regulating the way the technology can be applied: https://artificialintelligenceact.eu
(DIR) Post #ASOGPAfmuA7rzxaoAS by TheSteve0@data-folks.masto.host
2023-02-06T03:29:54Z
0 likes, 0 repeats
@dabeaz @simon @ajs For me the primary benefit will be is that I now have a "calculator" for writing. I hope to use it to write first drafts of blog posts and doc. I know it will be wrong but it's much easier to edit versus write de novo.With ADHD, and some serious blockage around writing - I am excited to see if this can help
(DIR) Post #ASOGPB9Z7RwfUJWaa8 by simon@fedi.simonwillison.net
2023-02-06T03:34:10Z
0 likes, 0 repeats
@TheSteve0 @dabeaz @ajs I've been using GitHub Copilot for that kind of thing for a while - if you run it against a Markdown document it suggests sentences for you, and while they're rarely what I actually want to write they almost completely eliminate any writers block I feel when faced with that initial blank page
(DIR) Post #ASOJFN649rznUru2l6 by brennansv@sfba.social
2023-02-06T04:08:22Z
0 likes, 0 repeats
@simon Google has not reliably found what I need in many years. People should lose the notion that Google is all knowing.Google search has led me to many pages which were not helpful or were even misleading. It’s why I open results in tabs and quickly skim to determine which might be useful. Perhaps AI can do better to interpret requests and give me better leads. All I expect is a head start, not the answer.
(DIR) Post #ASOQXzrVT5tQcHvAga by mattmoehr@zirk.us
2023-02-06T05:30:12Z
0 likes, 0 repeats
@simon @jwz do you consider Github Copilot to be a LLM? Domain specific language models seem like the first thing that’s going to work. Medicine, programming, or maybe law. Industries that are textual but produce way too much text for any human to deal with.
(DIR) Post #ASOQprgqadikU5SIts by darrelplant@mstdn.social
2023-02-06T05:33:24Z
0 likes, 0 repeats
@amyhoy @simon @jwz Random number generators are largely unpredictable.
(DIR) Post #ASOSiRCcs4IWDSuEIC by simon@fedi.simonwillison.net
2023-02-06T05:54:43Z
0 likes, 0 repeats
@mattmoehr @jwz run Copilot on a Mqrkdown document and it will completely English language sentences just like GPT-3 does - it's running on an LLM called Codex that's a descendent of GPT-3
(DIR) Post #ASOTK1upKu6EnHVQfo by mattmoehr@zirk.us
2023-02-06T06:01:13Z
0 likes, 0 repeats
@simon @jwz right. I didn’t know those details but I knew copilot was a LLM at its core. So to respond to your Q“The trick is to learn what they're useful for and how to take advantage of them, which is actually quite a lot of work”Is you have to turn ChatGPT into a domain specific data extraction thing like copilot is a domain specific LLM. At least for now I don’t think LLMs can be flexible enough to write (bad) poetry AND parse scanned police reports
(DIR) Post #ASOTtJ5PUJ2gBQHGIi by simon@fedi.simonwillison.net
2023-02-06T06:07:45Z
0 likes, 0 repeats
@mattmoehr I'm confident that GPT-3 itself (in particular the text-davinci-003 model) is indeed flexible enough to do both of those things, based on my experiments with it using the Playground interface and the HTTP APIThat's what makes large language models so fascinating to me: once they grow beyond a certain size they grow an enormous range of capabilities
(DIR) Post #ASOU5KzWENadZBG5gm by mattmoehr@zirk.us
2023-02-06T06:09:46Z
0 likes, 0 repeats
@simonYou probably already know this company but… I worked for a survey many years ago and we used Abby software for data entry from paper forms. https://www.abbyy.com/solutions/forms-processing/
(DIR) Post #ASOYa6YWLpqPabYaEC by mattmoehr@zirk.us
2023-02-06T06:59:04Z
0 likes, 0 repeats
@simon Hmm. Interesting. So we’re on track for this?>>>2022: “WOW you can write a prompt and an AI will draw it!”2028: “You want to write a prompt? First you need to hire 10-15 promptOps Engineers to build out your PromptFlow pipelines which sends promptjobs to your PromptLake from the PromptQueue using the EventPrompt stream”~ Chris Albonhttps://twitter.com/chrisalbon/status/1567688342124503040?s=46&t=o1xEBVG5tq1AlY2px8AzpA
(DIR) Post #ASOZ0uaQxauEHpEZxA by simon@fedi.simonwillison.net
2023-02-06T07:01:33Z
0 likes, 0 repeats
@mattmoehr LangChain is starting to look a bit like that already! https://langchain.readthedocs.io/en/latest/modules/chains/combine_docs_examples/qa_with_sources.html
(DIR) Post #ASOfbjybyDzPosUIBU by sirjofri@mastodon.sdf.org
2023-02-06T08:18:52Z
0 likes, 0 repeats
@simon that actually makes total sense... We'll see what they do. Maybe they just formulate a better search term for their usual search engine.
(DIR) Post #ASOhKCSbWtaMr9HRFA by deivudesu@mastodon.social
2023-02-06T08:38:21Z
0 likes, 0 repeats
@simon @amyhoy @jwz Agreed that there *must* be a few ways in which LLMs will change the face of tech (whether we like it or not).Would contend that, for most classification/parsing tasks, they can offer, at best, POC-level solutions (before a more dedicated implementation solves the problem fully)…But for obvious reasons, LLMs shine on "creative" tasks, where accuracy is far less important than plausibility and coherence.
(DIR) Post #ASP69RDko9KEItIc1A by JMMaok@mastodon.online
2023-02-06T13:16:29Z
0 likes, 0 repeats
@simon I can see them being useful for ‘make this email sound more friendly.’ We should probably all start discounting the friendliness of emails by about 20% if that catches on, though.
(DIR) Post #ASP6gTbexXKjG6Ecc4 by JMMaok@mastodon.online
2023-02-06T13:22:23Z
0 likes, 0 repeats
@jwz @simon I think for government documents you will need to do the 6 months of data entry first, then compare with LLM models to see what it gets right and wrong. LLM may become useful for ongoing updates. People also tend to underestimate how much it will matter if the 20k reports are from the same police department or hundreds of them.
(DIR) Post #ASP75DNsPVmgBZaH6O by Obdurodon@hachyderm.io
2023-02-06T13:26:47Z
0 likes, 0 repeats
@simon @jwz @amyhoy A bit OT but..."I've spent my entire career working with computers that do exactly what you tell them... and now I'm faced with one that very much does not do that"That might explain a lot of programmers' extreme reactions, both positive and negative.
(DIR) Post #ASP8OkZLLzY6YypJHk by undergroundbeef@mastodon.social
2023-02-06T13:41:37Z
0 likes, 0 repeats
@simon I played with ChatGPT for about 20 minutes, then got bored with it. It seemed like a neat toy, but without being able to trust it, I'm not sure how it can enable me to do more than I can do on my own today. Maybe it can save me some typing. However, the time saved in typing will be lost to verifying what it typed is actually correct and in my voice... so I might as well do it myself. They've automated the monkey's writing Shakespeare; it's hard to see it as more than that.
(DIR) Post #ASPDCkaZxAeYv9VBK4 by exchgr@mastodon.world
2023-02-06T14:35:18Z
0 likes, 0 repeats
@simon @jwz @amyhoy 🚨🚨🚨sunk cost fallacy alert🚨🚨🚨
(DIR) Post #ASPEJroZaQv3JQ0MXg by simon@fedi.simonwillison.net
2023-02-06T14:47:37Z
0 likes, 0 repeats
@undergroundbeef that's exactly the challenge I'm talking about: figuring out what it can be reliably used for is a fascinating exercise in my opinionThere are plenty of useful applications, but they're not at all obvious(I'm going to have to write about this in detail soon I think)
(DIR) Post #ASPEzvd3nNL1Iwfzg8 by simon@fedi.simonwillison.net
2023-02-06T14:55:34Z
0 likes, 0 repeats
@exchgr i don't think that quite applies here: I didn't spend any effort building these things, so I'm not invested in making sure they're useful so that my existing effort isn't wasted
(DIR) Post #ASPLXw0pTZ7443OPTc by gretared@vis.social
2023-02-06T16:10:48Z
0 likes, 0 repeats
@simon @amyhoy @jwz every real use case for chat gpt I’ve heard is “write code snippets for me” so let’s put it in IDEs and move on. Do NOT let this shit tell teen girls how to get health care or anything else of import.
(DIR) Post #ASPLjMSdTu2LHJqc76 by gretared@vis.social
2023-02-06T16:12:52Z
0 likes, 0 repeats
@anildash you’ve become a disappointment
(DIR) Post #ASPNmarsZX8mjqK9aa by anildash@me.dm
2023-02-06T16:35:50Z
0 likes, 0 repeats
@gretared could you explain more of what you mean? Are you arguing that current efforts have been effective in regulating AI?
(DIR) Post #ASPPoqxXuovz8HBDDk by kidehen@mastodon.social
2023-02-06T16:56:53Z
0 likes, 0 repeats
@simon,Yes!Transforming unstructured data to machine-computable structured data has been the focal point of my experimentation with CHAT-GPT which has yielded positive results, more often than not. In my experience, you can't expect CHAT-GPT to function as a miracle all-answering oracle without any prior domain knowledge, since that's where its very weak -- and understandably so. /cc @jwz #chatGPT #UseCase
(DIR) Post #ASPQy7aHZf1BFnHWyG by exchgr@mastodon.world
2023-02-06T17:09:30Z
0 likes, 0 repeats
@simon yeah, i guess since you haven’t done the work yet, “sunk cost fallacy” isn’t exactly the right term, but there’s some element of “it exists, therefore it’s worth looking at” that doesn’t sit right given the plethora of examples to the contrary. almost like the is-ought problem
(DIR) Post #ASPSiEr2yWEX4rg4DA by brianwc@mastodon.social
2023-02-06T17:28:52Z
0 likes, 0 repeats
@simon @amyhoy @ian @sayrer @jwz I find it bizarre that everyone refers to it as "it" as if we know precisely what "it" is a few months after an initial alpha release and that "it" will remain constant and unchanging for the next 20 years or something. So "it" hallucinates right now. Good to know. What are the techniques for preventing that? Will those improve? What will "it" do a year or 3 years from now? This isn't the release of a new Ford truck where once out, that's it, it is what it is.
(DIR) Post #ASPY1vw7WOyZxyGFXc by micah@indieweb.social
2023-02-06T17:06:30Z
0 likes, 0 repeats
Yeah we had thus all under control, Google, Facebook etc all were well-regulated and well-behaved, no problems in years, and then you had to come along and suggest GPT might not be garbage
(DIR) Post #ASPY1wbx0sR83hpek4 by gretared@vis.social
2023-02-06T18:30:40Z
0 likes, 0 repeats
@micah better than whatever explanation I was trying to concoct
(DIR) Post #ASPYG6MjXhxMMlStv6 by jwz@mastodon.social
2023-02-06T16:22:48Z
0 likes, 0 repeats
@gretared @simon @amyhoy Except that the "write code for me" use case is the one where it is most obviously (even performatively) illegal! It is a straight-up license laundry to do copyright infringement at scale.The whole thing is like a *fractal* of bad ideas, all the way down.
(DIR) Post #ASPYG8AOqXHLx5i4yO by gretared@vis.social
2023-02-06T18:33:14Z
0 likes, 0 repeats
@jwz @simon @amyhoy had no idea! “Fractal of bad ideas” should be TM’d 😆
(DIR) Post #ASPYoD7yuLqtBTpU3s by jesse@metasocial.com
2023-02-06T18:37:12Z
0 likes, 0 repeats
@simon One thing I've hear at least one place selling is a language model explicitly trained on your documentation corpus, so the lies it's telling are at least consistent with the lies you were already telling.
(DIR) Post #ASPaDNcJ994YuRa0au by simon@fedi.simonwillison.net
2023-02-06T18:52:32Z
0 likes, 0 repeats
@jesse I've been playing with that idea a bit, it's pretty fun. You can reduce the chance of it hopelessly misleading people a little bit by having it include citation links to the documents it sourced the information from (my demo doesn't do that yet but it really should) https://simonwillison.net/2023/Jan/13/semantic-search-answers/
(DIR) Post #ASPc3Yi5FOuWKJj4uO by jesse@metasocial.com
2023-02-06T19:06:53Z
0 likes, 0 repeats
@simon Can you engineer the prompt to not volunteer information that’s not included in your prompt?
(DIR) Post #ASPgIy3hgQnb6YRRQ0 by simon@fedi.simonwillison.net
2023-02-06T19:59:58Z
0 likes, 0 repeats
@jesse You can certainly try - Perplexity.ai uses "You must only use information from the provided search results" as part of their prompt:https://simonwillison.net/2023/Jan/22/perplexityai/As with all things LLM it's an open question as to whether it will actually obey you across every possible situation though!
(DIR) Post #ASPonuA11fn3coemzQ by RAOF@toot.cat
2023-02-06T21:36:28Z
0 likes, 0 repeats
@simon @amyhoy @ian @sayrer @jwz how does this match up with your “take a police report and extract names and dates into a json format” example? It seems a predict-the-next-token engine is also not going to be a good way of accurately summarising data?Now, you could probably build an extract-names-and-dates system on top of GPT, but how much better than a regex + a database of names is it going to be?
(DIR) Post #ASPrJQHrpVB89KUCsC by undergroundbeef@mastodon.social
2023-02-06T22:04:35Z
0 likes, 0 repeats
@simon With reliability being an issue, I think it is best suited for those tasks where reliability isn’t too important. For example, treating it as an extra person in the room during a brain storming session, or potentially surfacing a basic level of understand on a topic when a person is so clueless they don’t even know what to Google. There are times when I need to do some searches to find enough info to do a useful search. Fast tracking that would be nice and an interesting use.
(DIR) Post #ASPs1Cejs45KcC3nJw by simon@fedi.simonwillison.net
2023-02-06T22:12:20Z
0 likes, 0 repeats
@RAOF it's extremely effective at answering questions (aka guessing the next token that comes after the question mark) against content that you've just provided it
(DIR) Post #ASQ4xy48xYSntpjMMi by leymoo@hachyderm.io
2023-02-07T00:37:32Z
0 likes, 0 repeats
@simon @Migueldeicaza I am job hunting and I’ve found it great for: - making my CV easier to scan- rewriting covering letters I draft. I have a rambly writing style, it helps me restructure, simplify and remove errors without having to involve my friends- writing covering letters from scratch if I don’t *really* want the job much
(DIR) Post #ASQ4y113zSmZ39JRTs by leymoo@hachyderm.io
2023-02-07T00:37:32Z
0 likes, 0 repeats
@simon @Migueldeicaza I was one of those students who got extra time, a word processor and stickers on my exam paper stating that they had to ignore my spelling & grammar. Chatgpt means I can churn out NT text effortlessly by just drafting and getting it to rewrite, and kills one of my perceived disadvantages. Same with awkward social cues - it can check emails for the right tone.
(DIR) Post #ASQ9TfFOrSBccPMhwe by simon@fedi.simonwillison.net
2023-02-07T01:28:11Z
0 likes, 0 repeats
@leymoo @Migueldeicaza this is a great example of what I'm calling the "calculator for words" use-case
(DIR) Post #ASRFnNgWf4Wmp9nXBg by maria@thelife.boats
2023-02-07T14:13:47Z
0 likes, 0 repeats
@simon Would love to see any success stories you can share? I've seen just a sea of red flags, so far.#chatgpt #language #search #research
(DIR) Post #ASRGMgU8kqHnGdq3I8 by Lucky_Th13teen@vermont.masto.host
2023-02-07T14:20:08Z
0 likes, 0 repeats
@simon Its the same thing every time some new technology comes out. Big explosion of the bleeding edge group using it for cutting edge ideas (this group tends to also be quite educated), youth utilization, news cycle catches on and sees the use by youth and early adopters and reports. Tech is a tool, and everything is a hammer if your looking to set a nail.
(DIR) Post #ASROpIqmVSmDzqaPK4 by simon@fedi.simonwillison.net
2023-02-07T15:53:15Z
0 likes, 0 repeats
@maria there are definitely so many red flags!I'm going to have to write about this in more detail, but I've had success as a brainstorming companion, a summarization and text rewriting tool, a learning companion (see https://simonwillison.net/2022/Dec/5/rust-chatgpt-copilot/), a code generator and a whole bunch of other things
(DIR) Post #ASRSLkNfz36SVh8PKq by maria@thelife.boats
2023-02-07T16:34:34Z
0 likes, 0 repeats
@simon I'd be most grateful to publish on your experiences at popula.compls lmk if this is possible!! thank you(maria@popula.com)
(DIR) Post #ASS5EHWMmXPbBjd9eq by maria@thelife.boats
2023-02-07T23:50:09Z
0 likes, 0 repeats
@simon thank you so much, gonna study these
(DIR) Post #ASXPkqG28ZO53uasj2 by bitbear@icosahedron.website
2023-02-10T13:33:19Z
0 likes, 0 repeats
@simon @donaldball @anildash @amyhoy @ian @sayrer @jwz I expect that that most AI models will become utterly useless or at least completely uninteresting if you remove every single unlicensed copyrighted work from their training corpus.https://icosahedron.website/@bitbear/109789749992233660