Post B1JMjOP4FPiV0CcFYu by j@annihilation.social
(DIR) More posts by j@annihilation.social
(DIR) Post #B1I755ulsMKBbgoN60 by j@annihilation.social
2025-12-16T01:49:18.100105Z
0 likes, 1 repeats
@sun @p @dcc @lanodan @hjWhy doesn’t Claude do that well on HolyC？？ #templeOS #softwareenginee... [7584249227979590943].mp4
(DIR) Post #B1I80asDRbXc882AYi by p
2025-12-16T01:59:42.227210Z
5 likes, 0 repeats
@j @dcc @hj @lanodan @sun I don't know what the point of having a bot write recreational code, but LLMs do not form internal models of the structure or architecture of the code, they merely predict tokens.
(DIR) Post #B1I8X4FtjpZeWVb0Rk by j@annihilation.social
2025-12-16T02:05:33.928653Z
3 likes, 1 repeats
@p @dcc @lanodan @hj @sun yeah we aren't going to get consciousness out of an LLM
(DIR) Post #B1I9AqJiyntAsULsmG by lanodan@queer.hacktivis.me
2025-12-16T02:12:42.429201Z
3 likes, 0 repeats
@j @dcc @p @hj @sun Don't really need consciousness to generate code, there's a ton of codegen stuff out there.But they deal with a pretty different kind of structure than natural languages, which are also quite looser and generally allow to just append (because handwriting).
(DIR) Post #B1IA61DF8w57AxaGLQ by p
2025-12-16T02:23:05.655914Z
0 likes, 0 repeats
@j @dcc @hj @lanodan @sun This is why their code is dogshit, but Anthropic is positioning Claude as a thing that can replace an entire dev team, so people wonder why Claude can't do HolyC. The main answer is "HolyC isn't on the list of languages included in the LLM competitive benchmarks."
(DIR) Post #B1IA7OsedEiCs0c2C0 by lanodan@queer.hacktivis.me
2025-12-16T02:23:19.194614Z
2 likes, 0 repeats
@j @dcc @hj @p @sun Also I don't think consciousness is the right terminology, old chatbots like eliza/cleverbot/… aren't comatose or KO.Reasoning would be more like it but I'd argue just pissing code out where the problems have already been solved is what most people wish LLMs could do.
(DIR) Post #B1IAdWMyK6yrMb6rMu by sun@shitposter.world
2025-12-16T02:29:08.412663Z
4 likes, 1 repeats
@lanodan @dcc @j @p @hj I think every big llm service now has tons of secret sauce over top of the llm that works like cleverbot did, in addition to some actual but rudimentary reasoning functionality.
(DIR) Post #B1IBWJNoJB5b0ucznU by lanodan@queer.hacktivis.me
2025-12-16T02:39:01.628103Z
2 likes, 0 repeats
@sun @dcc @j @p @hj I don't think there's any reasoning at all but just having more data than a single human could even see and most of us being lacking in imagination (seen for example when people have to make up numbers, specially somewhat large ones).And in a way that's also why the eliza effect happens so easily, humans are capable of reason but that's not majority of what they do, specially for most jobs.
(DIR) Post #B1IBdlD9F74ZLdC0Js by sun@shitposter.world
2025-12-16T02:40:22.789015Z
1 likes, 1 repeats
@lanodan @dcc @j @p @hj there are arxiv papers on partially-reasoning llms that describe what they're doing. there is some disagreement but it's more sophisticated than just "stochastic parrot" now.
(DIR) Post #B1ICClAhNiUPMQjgm0 by p
2025-12-16T02:46:43.361286Z
3 likes, 0 repeats
@sun @lanodan @dcc @hj @j No, they do; they put classifiers in front and then more classifiers on the output. You wouldn't burn GPU time on kids spamming gamer words if you could avoid it.
(DIR) Post #B1IDbZxMJfIrlstXpQ by p
2025-12-16T03:02:24.783842Z
2 likes, 0 repeats
@sun @lanodan @dcc @hj @j > partially-reasoning llms that describe what they're doingThey use a "here's how you think about this" template to seed the token-predictor.
(DIR) Post #B1IDtsijI4zxBBmO6S by vii@dsmc.space
2025-12-16T03:05:41.244786Z
1 likes, 1 repeats
@p @dcc @j @lanodan @hj @sun I'm doing this right now. There's a whole lot of "wait" "let me think" "no that's not right" "wait no" "but what about" "is that right?" added between reasoning and re-reasoning descriptions for answers, to encourage these tokens highly and trick the model into "reconsidering" mistakes.
(DIR) Post #B1IEYLpYTmGJ2vGlZw by sun@shitposter.world
2025-12-16T03:13:01.536273Z
1 likes, 1 repeats
@p @dcc @j @lanodan @hj there are ones with cause-effect reasoning now. again not saying they're even as complex as an animal reasoning but
(DIR) Post #B1IF2NKlSE04OdBT60 by ageha@tomo.airen-no-jikken.icu
2025-12-16T03:18:25.883211Z
1 likes, 0 repeats
@sun @dcc @j @p @lanodan @hj limitation is feedforward design. brains have feedback from "higher layers" (really a spoke-hub-y web with many inputs), so something like part of your visual system that sees abstract objects talking back to the edge detectors and telling them what to see kinda architecturehttps://pmc.ncbi.nlm.nih.gov/articles/PMC4345765/without these many-layered loops you have to try lut-ing, meaning extreme bloat and little perturbations on the low level being able to disrupt the abstract high level too easily
(DIR) Post #B1IFA07fdXZgizhBr6 by sun@shitposter.world
2025-12-16T03:19:48.502984Z
1 likes, 1 repeats
@ageha @dcc @j @p @lanodan @hj even the reasoning LLMs don't really have the concept of objects yet
(DIR) Post #B1IGi8VV4PxoWogNSC by p
2025-12-16T03:37:12.793104Z
1 likes, 0 repeats
@vii @dcc @hj @j @lanodan @sun Yeah, I mean, I wonder how far they can push this. Like, reportedly the "tools"-style functionality was discovered almost by accident.
(DIR) Post #B1IGkjQml6r7coiNxQ by p
2025-12-16T03:37:40.940340Z
1 likes, 0 repeats
@sun @dcc @hj @j @lanodan Please do elaborate.
(DIR) Post #B1JMdDq6bwlyV9Jw24 by j@annihilation.social
2025-12-16T16:18:15.653209Z
3 likes, 1 repeats
@lanodan @dcc @p @hj @sun LLM "reasoning" is just using AI to expand the question to generate more text to ask the AI to hopefully get a more accurate answer out of the AI. That's literally all it is.
(DIR) Post #B1JMjOP4FPiV0CcFYu by j@annihilation.social
2025-12-16T16:19:22.524223Z
1 likes, 1 repeats
@p @dcc @lanodan @hj @sun I would but I'm also not a Multi-Billion dollar company/scam.
(DIR) Post #B1JMmvcRD1XOt5YWJ6 by j@annihilation.social
2025-12-16T16:20:00.844431Z
1 likes, 1 repeats
@p @dcc @lanodan @hj @sun no elaboration before marriage :jahy_baka:
(DIR) Post #B1JS3x4pXcIc6RlywC by Suzu@detroitriotcity.com
2025-12-16T17:19:04.058693Z
2 likes, 0 repeats
@j @dcc @p @lanodan @hj @sun I have been using ChatGPT and Gemini in their Pro versions for my personal bullshit, and I find the reasoning it does "behind the scenes" pretty interesting. It basically does that, yes, it expands the question to add more info and also to direct how it will go on about solving the problem and answering the question.In my use, it does a lot of "the user wants to do X, and by our conversation so far, he's using Y and likes to do things Z way".I find it even more interesting because there are a lot of insights and extra info, and sometimes even the correct solution inside the thought process that then the LLM won't put in the final answer. Reading the thoughts is an invaluable tool, at least in my experience.
(DIR) Post #B1JS7xWNpZhEdiZUJc by p
2025-12-16T17:19:50.799351Z
3 likes, 0 repeats
@j @dcc @hj @lanodan @sun > scamA thing's worth whatever people will pay for it.
(DIR) Post #B1JSHmJgLSOSmgJRZI by lanodan@queer.hacktivis.me
2025-12-16T17:21:35.703783Z
1 likes, 0 repeats
@p @dcc @j @hj @sun Would need to at least take stuff like buyer's remorse into account, otherwise well… nothing's a scam.
(DIR) Post #B1JSmfj4tjZFJzvGWO by j@annihilation.social
2025-12-16T17:27:11.506084Z
3 likes, 1 repeats
@p @dcc @lanodan @hj @sun hey @dcc we need an :enron: emojienron.png
(DIR) Post #B1JSwqrOMA9AkVfJaK by j@annihilation.social
2025-12-16T17:29:01.770026Z
1 likes, 1 repeats
@Suzu @dcc @p @lanodan @hj @sun using AI just feels like arguing with a well educated retard :akko_shrug_one_arm:
(DIR) Post #B1JT4NGKHgytGmuo9A by dcc@annihilation.social
2025-12-16T17:30:23.166147Z
2 likes, 1 repeats
@j @p @hj @lanodan @sun :enron:
(DIR) Post #B1JUS2uhgxyz3zeXw0 by j@annihilation.social
2025-12-16T17:45:52.846584Z
2 likes, 1 repeats
@dcc @hj @lanodan @p @sun let's go! :enron:
(DIR) Post #B1JUUFG0llprvSfU1o by dcc@annihilation.social
2025-12-16T17:46:16.569749Z
2 likes, 1 repeats
@j @hj @lanodan @p @sun I will take all emojies :alexjoneswant:
(DIR) Post #B1JUaLetskLTW7DhIm by Suzu@detroitriotcity.com
2025-12-16T17:47:19.887734Z
1 likes, 0 repeats
@j @dcc @p @lanodan @hj @sun honestly, I think it depends on how you use it.I don't claim to be an expert in tardwrangling AI, but I feel like the biggest mistake people do is "humanizing" it and treating AI like an expert they are asking for advice, or as a colleague they are brainstorming with.AI is actually a tool. You can use it very effectively to quickly do some repetitive task or to quickly scan over something (like a block of code or a log) to check for errors.As I mentioned another day when talking about AI, it can be used now (especially Gemini or newer versions of chatGPT) to condense and speed up internet search. There are a lot of times when I have some weird error or want to set up something that I know that has some quirks due to my specific hardware or the system I use, and then I search the internet for something with these restrictions. What I most often than not find is a bunch of reddit posts (because, as much as we like to dunk on reddit and call every redditor a retard, it's a fact that still there are a lot of people who use it for tech stuff and it's readily indexed, so almost every niche situation was already discussed there). Usually these reddit posts have some insight, and then someone posts a link to some forum post or blog with the actual solution, but this link never shows in the search engines, you have to find it through a reddit post (honestly, if I had a coin for every time it happened to me, I could buy more RAM).AI can speed this process because it will do the search, scour the threads and find the relevant links for you. I have done some tests here and it has been pretty reliable in doing it, and providing the actual sources for the info.
(DIR) Post #B1JUabO7OuQcIhUWMi by j@annihilation.social
2025-12-16T17:47:25.694553Z
2 likes, 1 repeats
@dcc @hj @lanodan @p @sun add all the state flags under the 2 letter abbreviation please. If you want I can collect all the state flags and name the files for you but it will have to wait till this afternoon or maybe tomorrow
(DIR) Post #B1JUdRJFKT6huYZ7Cq by dcc@annihilation.social
2025-12-16T17:47:56.284633Z
2 likes, 1 repeats
@j @hj @lanodan @p @sun I'll take em
(DIR) Post #B1JUkFsyCWQaVYWbTc by j@annihilation.social
2025-12-16T17:49:10.289491Z
5 likes, 3 repeats
@Suzu @dcc @p @lanodan @hj @sun lmao you greatly overestimate how I feel about AIa026ff33c49456aad0e82c26811a5523edb36c4b1966eb8b4507573fb84af992.gif
(DIR) Post #B1JUmXRvrzArYoy132 by j@annihilation.social
2025-12-16T17:49:35.104256Z
2 likes, 1 repeats
@dcc @hj @lanodan @p @sun we're going to use the old Mississippi flag because the new one sucks
(DIR) Post #B1JVF6aKpXC0djWEWO by dcc@annihilation.social
2025-12-16T17:54:44.655122Z
2 likes, 1 repeats
@j @Suzu @hj @lanodan @p @sun Que mgs2 ai video:
(DIR) Post #B1JXjL4liZ6sB7IYUq by p
2025-12-16T18:22:37.633672Z
1 likes, 0 repeats
@lanodan @dcc @hj @j @sun Well, a thing is worth what people will pay for it: misrepresenting the thing is another matter, but if idiots will pay for it, someone will sell it.
(DIR) Post #B1JXr8LSvySL0lpD4y by lanodan@queer.hacktivis.me
2025-12-16T18:24:00.701375Z
2 likes, 0 repeats
@p @dcc @j @hj @sun True, that's a better way to put it.
(DIR) Post #B1JXtLRDE7SKCGZEYa by p
2025-12-16T18:24:26.134732Z
2 likes, 0 repeats
@j @Suzu @dcc @hj @lanodan @sun > arguing with a well educated retardPoorly educated retard. It's like every single normie.
(DIR) Post #B1JY6rdwnbZFqQyEka by vii@dsmc.space
2025-12-16T18:26:51.423321Z
3 likes, 3 repeats
@p @dcc @j @Suzu @lanodan @hj @sun "It was trained on the entire internet (derogatory)"
(DIR) Post #B1JYGCZ0O4e7Km6xvs by lanodan@queer.hacktivis.me
2025-12-16T18:28:32.591303Z
2 likes, 0 repeats
@vii @dcc @j @p @hj @sun Aka it was trained on linkedin, imageboards, twitter, reddit, and "news"papers.
(DIR) Post #B1JYjgpAxPkLaHjgvY by mint@ryona.agency
2025-12-16T18:33:52.218435Z
5 likes, 2 repeats
@p @dcc @j @Suzu @lanodan @hj @sun Big Tech drained the oceans and raised RAM prices tenfold just to create an artificial redditor.
(DIR) Post #B1JYmhF9r731kjK4FU by p
2025-12-16T18:34:26.341395Z
3 likes, 1 repeats
@Suzu @j @dcc @hj @lanodan @sun > or as a colleague they are brainstorming withWell, you know, why the fuck *not*? Something that helps you reason shit out would be more useful. It's useful like a fuzzy reference; make it also act as a debugging duck and you have a real thing. People have finite patience for listening to my bullshit; I think it's probably more useful to me to talk to people than to talk to a machine, but at some point, basically anyone starts to think I'm shitting them or they lose patience with my rabbit holes, and it's that way with anyone, right? Mechanical research assistant would be pretty cool.That's a thing that is useful to people that want to do their own thinking but it's not good for engagement numbers. They want engagement numbers: https://archive.is/v4dPa . They've hired the Product Guy. They're A/B-testing ways to make normies talk to the chatbot. They're not making these things to as a force-multiplier for people's intelligence, they're trying to hustle idiots, because there are more idiots and it's easier to give an idiot an experience than it is to give him a tool, and the idiot wants an experience instead of a tool.
(DIR) Post #B1JZaaliCftW9NLEuG by mint@ryona.agency
2025-12-16T18:43:25.930981Z
1 likes, 1 repeats
@p @dcc @j @Suzu @lanodan @hj @sun >but at some point, basically anyone starts to think I'm shitting them or they lose patience with my rabbit holes, and it's that way with anyone, right?Probably has to do something with the fact in current year and day, you're talking like you're still on usenet. Everyone except other uncs sees you as an unc. It's about age and generational gap.
(DIR) Post #B1JZh6z9Kqh1o2jfwO by mint@ryona.agency
2025-12-16T18:44:36.626503Z
1 likes, 1 repeats
@p @Suzu @dcc @hj @j @lanodan @sun Not saying it's a bad thing, just a general observation. Even older zoomers have trouble understanding what their younger peers are yapping about and vice versa.
(DIR) Post #B1Ja3smCKVJoOlXGmu by p
2025-12-16T18:48:45.003626Z
4 likes, 0 repeats
@vii @Suzu @dcc @hj @j @lanodan @sun "Idiots plus whoever has consistently been able to afford good PR over the last hundred years of media."
(DIR) Post #B1JaFP3EfnC3gMzxTM by p
2025-12-16T18:50:49.930184Z
5 likes, 0 repeats
@mint @Suzu @dcc @hj @j @lanodan @sun Now they are spending even more money to achieve the impossible: make the redditor *likeable*.
(DIR) Post #B1JaYNGhWeitegLJjM by Suzu@detroitriotcity.com
2025-12-16T18:54:12.114918Z
1 likes, 0 repeats
@p @dcc @j @lanodan @hj @sun idk. >They're not making these things to as a force-multiplier for people's intelligence, they're trying to hustle idiotsExactly, and that's the main problem.But honestly, I think part of the problem is the way people use it. Like I said, if you know what you are doing, and you use AI as a tool to help you do what you need to, it can act as a force multiplier.But yeah, the allure of it being some "magical entity" that will fulfill all your wishes like some sort of djinn is too attractive for most people, and that's why they are pushing it, and that's why some people are huffing paint and falling for it.For instance, the other day someone reposted in fedi a Xwitter post by some retard saying that "He doesn't need to know about the resistance limits of concrete to build, because he has access to an expert that can take it into account when he asks for help" or something like that, I can't recall the exact words.
(DIR) Post #B1JbviNNJWshjZ9rCy by p
2025-12-16T19:09:41.028684Z
7 likes, 0 repeats
@mint @Suzu @dcc @hj @j @lanodan @sun > Probably has to do something with the fact in current year and day, you're talking like you're still on usenet.Cap, on god. Lil cuzzo got no drip: my rizz gives you aura.Also I was never on Usenet. I think @sun is the only person in this thread that has met me in person, but we've been on A Mumble. The easiest thing to do is type the way I speak, so that's what I do. I type like a nerd because I talk like a nerd: I am a nerd.Anyway, people have finite patience. This is not because I am old: people had finite patience when I was tiny babies. It is impossible to get someone to humor me indefinitely down 80 rabbit holes and still pay enough attention to supplement with facts or point out contradictions or cite research. I suspect that many people have this problem.The problem OpenAI has is that they want to boost the KPIs. Being useful does not boost the KPIs. Engagement metrics boost the KPIs.ytcracker--i_invented_the_computer--17_simping_on_her_chaturbate_lol_gooner_secret_vault_rare_loud_obnoxious_track_spam_is_life.mp3
(DIR) Post #B1JcNEKsb2Yofr6xAe by irie@merovingian.club
2025-12-16T19:14:36Z
3 likes, 0 repeats
@j @dcc @p @lanodan @hj @sun obviously not, it's like trying to create life out of stones. Too many tech bros live in lala land.
(DIR) Post #B1JdpJAiJFwDjeJCDY by p
2025-12-16T19:30:56.166850Z
6 likes, 1 repeats
@mint @Suzu @dcc @hj @j @lanodan @sun I'm not the first person to become An Old; I'm aware of how it works. This is a separate issue from me getting fascinated by something weird and then chasing it down the rabbit hole while everyone else says "Holy fuck, what the hell are you talking about?" and then I have no way to account for myself because it seemed obvious to me that this thing should be chased down the rabbit hole and I look at them the same way they're looking at me and I say "How the hell do you not care about this?" This is the 'Tism, not the Old; the Old is what makes me call stuff fuckin' awesome. The 'Tism is also what makes me care about attributing all of my horrible personality features to the correct cause.birdideas.jpgterry-s_commodore.jpgworking.jpg
(DIR) Post #B1Je3BoUSskiy3X6cC by p
2025-12-16T19:33:26.692779Z
3 likes, 0 repeats
@Suzu @dcc @hj @j @lanodan @sun > But yeah, the allure of it being some "magical entity" that will fulfill all your wishes like some sort of djinn is too attractive for most people, and that's why they are pushing it, and that's why some people are huffing paint and falling for it.Boomers loved Eliza. People still think the stupid Markov bot is a person and they argue with it. I have argued with Markov bots.> For instance, the other day someone reposted in fedi a Xwitter post by some retard saying that "He doesn't need to know about the resistance limits of concrete to build, because he has access to an expert that can take it into account when he asks for help" or something like that, I can't recall the exact words. This was the belief around "expert systems" in the 80s. It's the same magical thinking.
(DIR) Post #B1JgtRkQfKjLM1gATI by irie@merovingian.club
2025-12-16T20:05:14Z
3 likes, 0 repeats
@j @dcc @p @lanodan @hj @sun I wanted to add. AI has no intelligence. Intelligence means understanding, something that is intelligible. To understand something implies a Mind that can understand. AI means Artificial Intelligence, it simulates understanding, it simulates a mind, but it has no mind. No understanding, no intelligence.
(DIR) Post #B1JoOfgrZqThFhqbSq by sun@shitposter.world
2025-12-16T21:29:21.376389Z
2 likes, 1 repeats
@irie @j @dcc @p @lanodan @hj what's the problem if its not really intelligence if it can do stuff you want? I told it the other day to write me a complete embedded raspberry pi pico project with wifi to control an LCD panel and it did it all and it worked perfectly.
(DIR) Post #B1Jobk8WMFtCo3181w by sun@shitposter.world
2025-12-16T21:31:43.769377Z
0 likes, 1 repeats
@p @dcc @j @Suzu @lanodan @hj > They're not making these things to as a force-multiplier for people's intelligenceit works great for it if you use it like that, because yes they are
(DIR) Post #B1JprFQ2Qrws7pASw4 by lanodan@queer.hacktivis.me
2025-12-16T21:45:43.219536Z
1 likes, 0 repeats
@sun @dcc @j @p @hj LinkedIn kind of advertisement material there.A good half of this thread is hackers, not tech enthusiasts, canned stuff like that isn't going to work.
(DIR) Post #B1Jq4DAw5LagvoHl1E by All_bonesJones@shitposter.world
2025-12-16T21:48:04.209176Z
2 likes, 1 repeats
@sun @irie @dcc @hj @j @lanodan @p this tbhif it can turn my psuedocode into properly formatted javascript that does what i was describing who gives a fuck where it lands on the spectrum of consciousness and agency?
(DIR) Post #B1JrYKabBWUkZDDOPA by lanodan@queer.hacktivis.me
2025-12-16T22:04:42.848374Z
1 likes, 0 repeats
@All_bonesJones @dcc @j @p @irie @hj @sun Yeah, although that bit is like Google Translate / DeepL, and of course it can. Of course not reliably, but we all know that for Google Translate / DeepL, or at least quickly learn it.But that kind of stuff is different than the AI side of things, which is purely just marketing, and marketing that tends to go away when things become tools.Like how Siri is barely considered an AI anymore and more taken as a speech-to-text interface.
(DIR) Post #B1JsELnR5TboTOFvZw by sun@shitposter.world
2025-12-16T22:12:19.105670Z
1 likes, 1 repeats
@lanodan @dcc @j @p @hj what you said reads like an implication that a hacker wouldn't find it useful therefore you're not really a hacker. unless I missed some point of the thread, which went on for a while while I was asleep. what do you mean by canned stuff, conversational AI like off the shelf Claude is perfectly fine for exploring things. if you want it to do all your work for you it will, and it might be okay or it might not. but you can use it to learn a lot and it works for that. there are tons of stuff it doesn't know and if you try to use it to solve hard problems you still need discipline and curiosity. some people won't go any deeper than what the AI does for them, that is a them problem not a tool problem.
(DIR) Post #B1JsuxQ3b2jpJxdYwK by lanodan@queer.hacktivis.me
2025-12-16T22:19:58.709758Z
2 likes, 0 repeats
@sun @dcc @j @p @hj I mean that this reply wasn't something actually technical, felt terribly out of place, and has been told again and again. Like people repeating marketing material (aka lies, for example making something seem better than it is, is a lie).And hackers as in people how go beyond surface level to actually learn how something works, to learn the limits and often make their own as well.
(DIR) Post #B1JtwXopqgpGSLY0qe by sun@shitposter.world
2025-12-16T22:31:30.103238Z
3 likes, 2 repeats
@lanodan @dcc @j @p @hj squeezing use out of bad, limited tools is a hacker virtue
(DIR) Post #B1Juk7p8lvaSJTKf6O by sun@shitposter.world
2025-12-16T22:40:28.497280Z
1 likes, 1 repeats
@lanodan @dcc @j @p @hj I don't know what to say except I started out as skeptical but I kept using the broken bad tool until I figured out how to get useful work out of it. I feel like this counts.
(DIR) Post #B1Jw1bFkhxPYSmfz7Y by lanodan@queer.hacktivis.me
2025-12-16T22:54:49.427366Z
2 likes, 0 repeats
@sun @dcc @j @p @hj I think there's more than that but roughly yeah.Big part of the issue is you didn't put it as limited until being effectively called out on it.Also I really think you should have read the thread, specially on such a controversial topic.
(DIR) Post #B1Jx4cRUqGxGyWAmZM by sun@shitposter.world
2025-12-16T23:06:35.137856Z
2 likes, 1 repeats
@lanodan @dcc @j @p @hj I am giving you more credit than how you're acting right now, I made a statement that amounted to it works for me and iunstead of extending good faith because we've known each other for so long that I actually am trying to use it for things and having success you're treating it equivalently to marketing copy.
(DIR) Post #B1JxzcVFI1nLfX6g3U by lanodan@queer.hacktivis.me
2025-12-16T23:16:52.549518Z
1 likes, 0 repeats
@sun @dcc @j @p @hj Well I did because it reads to me (it still does) as "They work great as force-multiplier for people's intelligence, because yes they are" and no further statement/arguments, directly on a post about how they aren't.Not sure how one could take that in good faith, although sadly I remember that you previously have trolled on the AI topic.By the way, I wrote "A good half of this thread is hackers" to avoid something like gatekeeping on such a loose term, I guess "at least" would be better wording.
(DIR) Post #B1JzucusWcB3PZqKKe by i@declin.eu
2025-12-16T23:38:21.689846Z
3 likes, 0 repeats
@lanodan @dcc @j @p @hj @sun it insists upon itself as the kids say, it works because it works, and it wasn't a holy cow worth a trillion dollars we wouldn't be needing this thread and accept it for what it is
(DIR) Post #B1K0eGSiLFBX5kQnKq by MagicMoshe@annihilation.social
2025-12-16T23:46:38.580037Z
4 likes, 3 repeats
@p @dcc @j @Suzu @lanodan @mint @hj @sun :terrylaugh:Terry Davis.mp4
(DIR) Post #B1K5mLy8FdBoG3nqlc by j@annihilation.social
2025-12-17T00:44:08.220830Z
1 likes, 1 repeats
@irie @dcc @p @lanodan @hj @sun nah they're just trying to hype up a big box of nothing
(DIR) Post #B1K5oZk40vvvYOHYzw by sun@shitposter.world
2025-12-17T00:44:31.987866Z
1 likes, 1 repeats
@lanodan @dcc @j @p @hj I don't troll people at random, in particular I don't troll people that don't reflexively make the shittiest arguments, and so I've never trolled you. but this thread is weak and I'm out.
(DIR) Post #B1K5z5wL9pA0JH2Zea by j@annihilation.social
2025-12-17T00:46:26.278753Z
0 likes, 1 repeats
@p @dcc @Suzu @lanodan @hj @sun nah AI is slightly above normies in some aspects
(DIR) Post #B1K7NWX82GDXgxSbs8 by deutrino@mstdn.io
2025-12-17T01:02:02Z
1 likes, 0 repeats
@p @dcc @j @Suzu@detroitriotcity.com @lanodan @hj @sun I'm missing half the thread because of my fedi neighborhood, but I've been saying that what I want in a coding assistant LLM is a rubber duck which can puzzle thru things with me, unlocking my own abilities by acting as a domain-specific interactive Rorschach blot ... some hallucinations and incorrectness would be fine; when I still had a straight dev job it wasn't uncommon to figure things out in pair programming by barking up the wrong tree a while.
(DIR) Post #B1K7lz0znCtAyb5rto by deutrino@mstdn.io
2025-12-17T01:06:27Z
1 likes, 0 repeats
@sun @dcc @j @Suzu@detroitriotcity.com @p @lanodan @hj I have no experience with megacorp LLMs for coding, can you use any of them as an interactive rubber duck that *isn't* overly verbose and *doesn't* have a main interaction style of vomiting a mishmash of example code? I want something that acts like a bright junior dev pairing with me. one who remembers stuff about code patterns and can roll the dice & win on suggestions often enough to help my actual human brain figure things out / stay semi on track.
(DIR) Post #B1KRJ3LOcHzj5W7OFM by p
2025-12-17T04:45:21.463628Z
2 likes, 0 repeats
@irie @j @dcc @lanodan @hj @sun> No understanding, no intelligence.Well, that's the thing, right? Define "intelligence" in objective terms.Back when AI was a respectable field (it only manages to be respectable long enough for a breakthrough before hype overwhelms it; this was what happened 50 years ago, so if that was any indicator, we'll have an "AI winter" where the research resumes when the money dries up--the money dicks care about the industry and not the research--and then it'll go back to being respectable again), there was an observation (I forget who originated it but I got it from Hofstadter and I recommend reading "Goedel, Escher, Bach" and then ignoring everything else he has done since then, because GEB is entertaining and interesting and gets at a few of the problems in AI while also doing, like, a tour of philosophy and math and enough to stay interesting but not so much that it ruins non-programmers' enjoyment of the book, and this is really rare) that however you define "intelligence", a machine will be able to do it. That is, an objective definition for intelligence means an objective benchmark for it.Hamming's lectures at the Naval Academy are all available to the public; I got them from Youtube. In particular, he did two about AI and basically everything he said holds: people are more interested in whether or not a machine can replace a human than what a human can do when given an intelligent machine as a tool. (Hamming was a rung up from Doug McIlroy, who was the boss of dmr, ken, bwk, and everyone else at Bell Labs. In addition to the things the lab did while Hamming was in charge of it, Hamming personally developed automatic error correction and fiberoptics, among other achievements. Brilliant guy. *All* the lectures are interesting, but the AI bits are relevant here.) That's something Jobs observed, also: he trotted this anecdote out in at least two interviews, someone had left a copy of "Popular Science" on the table at the office and it had a chart on efficiency of movement: how many calories per mile? By a long shot, humans performed the worst, least effecient, and birds of prey performed the best, because they can lock their wings and just coast on air currents and float up there with almost no effort. But, someone had added one other entry to the chart: a human on a bicycle. If you give the human a bicycle, we pass the birds of prey again. So he said he thought of computers like that, a bicycle for your mind. And, long before he decided to fuck everything up and lock it all down, that was more or less what Apple was making.So what Hamming said was that you've got to make up your mind whether or not you believe in a soul, and you've got to then be aware of how that will fuck you up: a person with an ardent belief in a metaphysical component to humanity will consistently underestimate machines and miss opportunities, and a person that believes very strongly that man is just a biological machine is going to consistently *over*estimate machines, inevitably over-investing and falling over.I'm in either Hamming's or Hofstadter's camp (it's the same camp, they arrived from opposite ends; I think Gosper or Greenblatt or somebody--I'm tired and I will continue fucking up attributions--said something like "I don't want to simulate a human: I want to make something different"): whether the machine is "intelligent" or not doesn't really matter. There the machine is, and this is what it can do, this is what you can use it for. I think I've mentioned C. elegans and OpenWorm about a million times, but it's got just a handful of neurons and we can simulate it with only about 80% accuracy. We are not going to get near to simulating a human without hollowing out the moon and replacing it with vector machines. (And even then, we'll have to worry about heat.) But the machine to date has been very good at doing things that are tedious or impossible for a human to do. Now there are LLMs and they predict tokens and they're pretty good at it, and anyone that is smart will be looking at what new niche we can put the machines into so humans don't have to do it any more, but as I said about as soon as I saw a paragraph written by a machine, so far the things are mostly used for spam. (And see my other remarks about the machines obsoleting propaganda by giving anyone that can buy enough GPUs an entire fleet of words-of-mouth to replace the traditional broadcast outlets: the CIA doesn't need to buy newspapers any more, it just has to roll out a few hundred thousand bots and then everyone can have their own tailor-made propaganda machine addressing their specific concerns. We all hop to fedi and it's decentralized and they can't coopt or take down at will...and this place is even easier to flood with bots. I can't wait to say "Huh, that's weird" about eighty times and then someone leaks that they government was using a bot army for surveillance and astroturfing and that all the weird shit was neatly explained by a revolution in influence operations.)
(DIR) Post #B1KRhYQYhXggRPDSgS by tard@clew.lol
2025-12-17T04:49:45.942423Z
1 likes, 0 repeats
@p @dcc @j @irie @lanodan @hj @sun My GPU is smarter than me. I realized this two years year when I was ERPing with it. I plan on selling it on Ebay before it sucks me back in.
(DIR) Post #B1KV9Z6q3j7YWj2Ze4 by lanodan@queer.hacktivis.me
2025-12-17T05:28:26.111126Z
1 likes, 0 repeats
@p @dcc @j @irie @hj @sun I kind of feel like you could have a definition of intelligence that a machine pretty much can't do except mimic it.For example, darwin awards are like "stupidity awards", so could put identifying danger as one aspect, as seen with lab rats too.And I'd put communicating with others and teaching as a form of intelligence too because helping your peers avoid danger as well effectively makes you stronger.Then I'd guess things like records and history because you got good enough at immediate preservation that you can try to plan more long-term than a lifetime.Meanwhile computers/software is ridiculously good at destroying itself all the time.And so far it seems like what's there is low-level and immediate reactive stuff like error-correction and basic forms of pattern recognition on filters, which feels more like an immune system kind of thing, like your body learning stuff rather than yourself.
(DIR) Post #B1KZuOF0gTieFMRhdw by p
2025-12-17T06:21:44.739169Z
3 likes, 0 repeats
@sun @Suzu @dcc @hj @j @lanodan > it works great for it if you use it like thatShow your notes: which model, what settings, what runtime? Because I am going to have to dispute this unless we're going to be specific. Maybe you're using different models than I have seen; you've been up to your elbows in this stuff for a minute, I have just prodded it on occasion. But there's no monolithic "it" that "works great" because there are a million of these fuckin' things. Please be specific, because I want what you are saying to be true. I think we mean different things by "research assistant" or "force-multiplier for human intelligence".I would very much like to use it like that and I tried to use it like that and then I tried interrogating it to ask it how to use it like that. So far, no success.It's useful as a fuzzy reference (unless it hallucinates): "How do you do this in $language? What's the equivalent of $x from $other_language?". If I write a program, it can often understand the program: it surprised me a few times. It cannot make sense of a lot of stuff, though, and it doesn't understand the operating environment: it can do anything that 50 other people have already done on GitHub. I fed it some of my weird shit and asked it what it did and what it was trying to solve: it could reliably "guess" the problem if the problem was spelled out in the comments. So I came up with a differentiator: about half the robots would fuck it up and about half would get it right, and of those, most would come up with a really tragically suboptimal solution and of those, some would revise after being nudged and some would make it worse. "Write an awk program that prints only the prime numbers from stdin." Small problem, trivial, first-year shit, I expected to mostly look at algorithms, not "pass/fail". Nobody must have checked that into Github, though: the goddamn thing first gets hung up on stdin, I tell it, "Look you're just going to be dealing with one natural number per line." Half of them shit gibberish out. The other half does trial division counting from 1 to n-1. I figure, you know, okay, I didn't say it had to be fast or even "not an embarrassment": so I say "speed it up", 0% success; fine. I say "If you're going to do trial-division, special-case 2 and start from 3, counting up by 2s until sqrt(n)". They shit their pants. They start going back and trying to split $0. For fun, I tell them "Try a Hamming sieve" or "Try memoizing the results" and the thing starts crying. For fun, I will attach my attempt at dealing with dolphin-mistral because that was the odd case out: it actually *did* eventually get it right, but it took a while and a lot of "No, no, that's not what I told you". This is, I'd like to stress, a really trivial task. `seq 1 9999999999 | awk 'p($1); function p(n){if(n%2==0)return 0; s=sqrt(n);for(i=3; i<s; i+=2)if(!(n%i))return 0; return 1}'`. Trial division is slow and you're better of with a sieve even in awk, but most of them did `for(i=1;i<n;i++)`. I would have taken 'p($1){print $2} function p(n){c="factor " n; c|getline; return NF==2}'. It'll make a login page if I ask it to make a login page: Github has plenty of implementations login pages. When one or another goddamn idiot bot starts saturating my entire goddamn pipe because you can't call it a "Large" Language Model unless you DDoS every cgit deployment that dares to make itself publicly visible, they'll have even more implementations of login pages. Fine, whatever.Fine: they're shit at generating code. Maybe they can do the "research assistant" thing where I say "Look, I'm gonna think aloud, you just supplement, like a research assistant." They wanna push back if I say anything that a redditor wouldn't expect me to be able to say. The thing will sit there and confidently spew lies at me and then lose its shit if I say that Bob McNamara was responsible for the one-child policy in China. It'll argue try to stop me from spreading misinformation if I say "Well, the World Bank an IMF are semi-autonomous agents of the UN." That's in the goddamn charter for both organizations. I swear and type in all-caps to manipulate the salience estimate: the fuckers drift, so you write "I GODDAMN TOLD YOU TO FUCKING $x NOW YOU GIVE ME A THREE-PARAGRAPH ESSAY ABOUT WHY YOU WILL DO $x FOREVER" and the salience plus the reinforcement will sometimes trick it into surviving another few exchanges. If you can hold a train of thought while someone keeps asking you if you want "(a) a detailed refutation of everything you have said, (b) the phone number for the suicide hotline, or (c) a list of conquistadors whose names rhymed with those of English monarchs", then by all means. I did not feel "assisted".So I can get it to, if I color inside the lines, act as a reference work for anything that is so uncontroversial I can do it it in my sleep. Now, I get that bugmanning is not a fundamental limit of the technology itself but the drift is; some of these issues are fundamental and some are just because you either build your own from scratch or you deal with the one they've built, and they didn't build it for researchers, they built it for normies or code-monkeys. Whatever.I read an interview with Bill Joy, around 1980-ish; early 80s. vi was still new and exciting, :sun: was a podunk operation. The interviewer had asked him what he thought about make(1). His verdict was that he liked it: "the things that used to be scribbled on a note next to the terminal are in a Makefile now". In philosophical terms, there's still the objection I had initially; I mentioned it to you in January. Maybe you don't need to worry about that: maybe instrumental code is no longer worth keeping because you can just run another one off. The level at which you interact with the machine has ceased being serialized in a repeatable manner. You told it to write your Pi Zero program, fine: the program is checked in and not your instructions to the LLM, and the LLM itself is stochastic. Maybe next time, it spits out a shitty one that breaks for reasons that you don't understand. This means you've got to keep the working version: the code is now a rune handed to you, an artifact handed to you. So you check in this *thing*, it's code you generated instead of writing, and that's fine, but you don't check in the prompt, the conversation, the lengthy cajoling. We end up not being able to continue insisting that the code is a throwaway artifact.One thing it can do is NLP: I guess it shouldn't come as much of a surprise that a "Large Language Model" can do reasonable "Natural Language Processing". So I can use it for tasks like that; I've had some success with that. I tell it four or five paragraphs about who I am and then tell it to summarize my notifications, and then because it is a goddamn monkey-paw but without the ironic twist being *clever* in any way, I have to spend a paragraph explaining that a summary should be *shorter* than the thing it is summarizing. And then I tell it, "Tell me if anything urgent is going on, and otherwise list the participants and summarize the topics." Then I've gotta run it through cmark; them's the breaks, it was easier than trying to convince the thing to stop using markdown. And because that's low-stakes, I don't have to care that it fucks up about 20% of the time. I can laugh, I can enjoy the chatbot. I can ask it "Elixir must have a way to do this, how do you do this in Elixir?" or Python or JavaScript because it *loves* Python and JavaScript; I don't. It takes a lot of effort and yelling to get it to not try to generate the entire goddamn whatever from scratch. I say "No, this won't work because of that." and it shits an entirely new program instead of just "No, it wouldn't: we should change $x to $y." The goddamn things actively steer you into pics related. "Stop trying to write the goddamn thing, robot, I will never run your code: I want you to discuss the architecture because I am working on a toy problem so that I can apply the approach to a bigger problem. Do you see?" The robot replies: "'See'...'see'...He must mean 'option (c)'. Hey, here's a list of all those goddamn conquistadors!"I swear it is *exactly* like dealing with the goddamn Genuine People Personality doors from the Hitchhiker's Guide. I am willing to look past that specific horror as long as I can make them do research assistant shit. No luck so far. Please explain how I can do that so that I can do that.dolphin-mistral.jsoncaptain.jpg554295-walle4-918613477.jpg
(DIR) Post #B1KdaegeEGhJHGDbVo by p
2025-12-17T07:02:59.963479Z
0 likes, 0 repeats
@lanodan @sun @dcc @hj @j :zerocool:
(DIR) Post #B1KdgQX7lTnT7OW3N2 by p
2025-12-17T07:04:02.547844Z
0 likes, 0 repeats
@All_bonesJones @sun @dcc @hj @irie @j @lanodan Well, that was more or less what Hamming and the others thought, and I'm inclined to agree: it's not a meaningful question. The thing is there, however you want to classify it.
(DIR) Post #B1KeCzT1PvAajL7mme by p
2025-12-17T07:09:55.650610Z
2 likes, 0 repeats
@sun @lanodan @dcc @hj @j > what you said reads like an implication that a hacker wouldn't find it useful therefore you're not really a hackerIt seemed more like "'I told it to write a program and it did' is not enough information to matter to the people in this thread".> ClaudeFUCK CLAUDE, FUCK ANTHROPIC, I WILL GO INTO THEIR OFFICES AT NIGHT AND SET THE BUILDING ON FIRE AND THEN TELL THEM THAT THEY MAY HAVE APPLIED COMMON SENSE AND DETERMINED THAT NO REASONABLE PERSON WOULD DO THIS AND BESIDES SETTING THEIR OFFICE ON FIRE WAS NOT IN /robots.txt SO I SHOULDN'T DO IT, BUT I WILL POINT OUT THAT THEY DIDN'T SPECIFICALLY FORBID ME IN /arsonists.txt AND THAT THIS GIVES ME LICENSE TO HALT THEIR OPERATIONS BECAUSE I NEEDED THE DATA FROM THEIR ASHES AND IT IS FINE FOR ME TO FUCK UP THEIR DAY IN ORDER TO MARGINALLY IMPROVE MY "LARGE ASH MODEL"
(DIR) Post #B1KeQ754Q9fw6mB9Si by p
2025-12-17T07:12:17.949204Z
1 likes, 0 repeats
@sun @lanodan @dcc @hj @j > squeezing use out of bad, limited tools is a hacker virtueSkepticism about hype is another hacker virtue; getting novel use out of tools by ignoring the tools the normies use is not just a virtue but practically a hacker trademark.
(DIR) Post #B1KefCN91sh4rapHwe by p
2025-12-17T07:15:01.526187Z
1 likes, 0 repeats
@sun @lanodan @dcc @hj @j Well, I eagerly await your reply to my other question asking for specifics, because the things are basically an over-hyped search engine with some fun NLP features as far as I can tell.
(DIR) Post #B1KerxjUlnZRDyMo6K by p
2025-12-17T07:17:19.932670Z
0 likes, 0 repeats
@lanodan @sun @dcc @hj @j > Also I really think you should have read the thread, specially on such a controversial topic. Well, like, :moon: and :mgsgb_7: have mentioned doing this stuff for day-jobs, so they definitely know some things we don't.
(DIR) Post #B1Kf1nc9FG92jpcYL2 by p
2025-12-17T07:19:06.600158Z
1 likes, 0 repeats
@sun @lanodan @dcc @hj @j I didn't read it that way; I think it was a remark about the level of detail, not a matter of good or bad faith.
(DIR) Post #B1KhECAZyU5sHKlRzc by snacks@netzsphaere.xyz
2025-12-17T07:43:38.422487Z
1 likes, 0 repeats
@p @dcc @j @Suzu @lanodan @hj @sun normies can't tell.me about how i've been using some compiler internals wrong at runtime
(DIR) Post #B1KhJzPYK6NmRjWyfo by p
2025-12-17T07:44:48.450370Z
5 likes, 2 repeats
@MagicMoshe @Suzu @dcc @hj @j @lanodan @mint @sun I just wake up and turn on my computer and do whatever I want.I dine on nectar of the gods.terry_dines_on_nectar.jpg
(DIR) Post #B1KhUyowrIKpnheNIu by p
2025-12-17T07:46:47.626805Z
0 likes, 0 repeats
@sun @lanodan @dcc @hj @j > but this thread is weak and I'm out. :terrymad:
(DIR) Post #B1KhXB1OT36CDgWEaG by p
2025-12-17T07:47:11.448085Z
1 likes, 0 repeats
@j @Suzu @dcc @hj @lanodan @sun Normies are retarded.
(DIR) Post #B1KiEUhlA3MribzTtI by p
2025-12-17T07:55:01.165257Z
2 likes, 0 repeats
@deutrino @dcc @j @lanodan @hj @sun I mean, that's what I'd really like. If I'm sitting back in a chair with a notepad then I can go down infinite rabbit holes if I want. It already manages to do many things that meet or pass human standards::mgsgb_right: It can be snotty about using a search engine for you, so you're like "What was that guy's fuckin' name?" and it will say "I can search the web for that." all passive-aggressively.:mgsgb_right: It get annoyed by my idiot tangents.:mgsgb_right: It doesn't get my jokes.:msgsb_right: It lies like a motherfucker.If it didn't constantly bust up my train of thought, it would be the *perfect* research assistant.
(DIR) Post #B1KiUcOWL4A2CWgEjo by p
2025-12-17T07:57:56.005655Z
1 likes, 0 repeats
@deutrino @sun @dcc @j @lanodan @hj > an interactive rubber duck that *isn't* overly verbose and *doesn't* have a main interaction style of vomiting a mishmash of example code?Yes. Some of them. GPT will reliably fuck up this style of interaction and it will default to cloying and shitting cookie recipes onto your screen, *exactly* like a monkey-paw. I haven't found one that reliably does this. I suspect fine-tuning would do it.
(DIR) Post #B1KilWTNbmkcl50uoK by p
2025-12-17T08:00:59.310229Z
2 likes, 0 repeats
@tard @dcc @hj @irie @j @lanodan @sun did you have mental chat sex with a floating point number2007-12-10--ebuggery.gif
(DIR) Post #B1KiuyAA1jZJFAWqDA by meeper@udongein.xyz
2025-12-17T08:02:33.347994Z
1 likes, 0 repeats
@p @dcc @j @lanodan @hj @sun Not really an NLP search engine. I kinda like how people equate it to as lossy semantic decompression (this holds more true when you use it to generate stuff like videos, or audio, where you can see what's equivalent to the artificial smoothening of an approximated curve or whatever and that It tries to reporduce a distribution based on the training data, but that reallly doesn't capture the autoregressive nature of these things, I read the abstract of one paper which proved the kinda obvious fact that these can be modeled as markov chains, tho ofc they have to an extent a continous space of states and no defined start or end state.So it's kinda a mix of all these things, the autoregressive nature really works on language though as it's kinda forced to act more convergent (as they do capture the semantic relationships in language it has to kinda at least make some sense)
(DIR) Post #B1KjJ50yChWHRaj6wa by p
2025-12-17T08:07:03.189761Z
1 likes, 0 repeats
@lanodan @dcc @hj @irie @j @sun Well, I mean, that's still informal, right; I think the essence there is that "consciousness" and "intelligence" have thus far evaded a *formal* definition and they are not going to get any more solid now that robots are continuing to inch towards the latter.
(DIR) Post #B1KjYs7XuXU65UKgc4 by meeper@udongein.xyz
2025-12-17T08:09:46.269127Z
1 likes, 0 repeats
@p @dcc @j @deutrino @lanodan @hj @sun They really should be treated as 'Generative' AKA a plausible bullshit generator -- Sometimes what you want is plausible bullshit slop, or a bullshitted set of guidelines that you can somewhat verify and it's up to you to make it generate useful bullshit.
(DIR) Post #B1Klij0kpoAkdfywiW by lanodan@queer.hacktivis.me
2025-12-17T08:34:04.214560Z
1 likes, 0 repeats
@p @dcc @j @irie @hj @sun Yeah, that's the most annoying bit about it, and part of why I tend to avoid using "intelligent", specially in technical discussions, there's usually more precise words without getting wordier.
(DIR) Post #B1KzkVm8PlxRWgtENE by i@declin.eu
2025-12-17T11:11:15.835736Z
1 likes, 0 repeats
@p @dcc @j @Suzu @lanodan @hj @sun the 4gb cpu models and the 1tb gpu models are a world apart too
(DIR) Post #B1LLOMiTc52q3E1XX6 by irie@merovingian.club
2025-12-17T15:13:43Z
2 likes, 0 repeats
@sun @dcc @j @p @lanodan @hj there's no problem with using mindless AI for whatever work you want it to do, that's the ideal usage for it imo. The problem arises when people expect it to understand things, or expect it to "develop" consciousness. This error stems from how modern neuroscience theorises about human consciousness (network) so they cannot see why a complex-enough artificial network wouldn't achieve consequences too, which is all nonsense of course (dead ≠ living)
(DIR) Post #B1LM1L79VGKONO9MQ4 by Suzu@detroitriotcity.com
2025-12-17T12:45:43.484566Z
1 likes, 0 repeats
@i @dcc @j @p @lanodan @hj @sun just to complement it. I'm not running any cool models locally yet, but as I mentioned, due to some circumstances of my work, I have both chatGPT Pro and Gemini Pro.Here is me asking the awk prime numbers thing to ChatGPT.What I have noticed is that it will present the simplest code it can generate and try to explain it, probably because it assumes the user is trying to learn. It then mentions the code can be optimized, and if you ask it, it will explain there are different optimizations for different uses and ask what you want to do. I tried throwing something ridiculous at it, and the response seems pretty sound (although it was wrong when it said my upper limit was 10^100, it was merely 10^89).I think it handled that pretty well. But I comprehend p might be annoyed with it starting with the simplest and most non optimized solution and then having to be nudged into doing something better. But yeah, I can understand it going with the simplest first because I just asked it to do it in a simpler way, and there are different kinds of optimization needed for different scenarios. I assume that if I had prompted it in a more detailed way, it would go with a more optimized solution from the get go.
(DIR) Post #B1LQ4IC5DY9q0J7Ujg by p
2025-12-17T16:06:11.797213Z
1 likes, 0 repeats
@snacks @Suzu @dcc @hj @j @lanodan @sun If you got all of the normies together and poured their brains into one normie, then that normie might be able to do so.
(DIR) Post #B1LQdm2EGFTcgpQgoS by p
2025-12-17T16:12:36.541659Z
2 likes, 0 repeats
@RedTechEngineer @dcc @deutrino @hj @j @lanodan @sun > No nasi kangkang in Islam.I saw some dude, he was like, an imam or something, and he was like "I've solved the problem of allowing women to do office work in Islamic countries. We just have to drink milk out of their tits. This makes it a familial relationship, so it's not unclean to work in the office with her."
(DIR) Post #B1LQr4ILgLyf5RJCL2 by p
2025-12-17T16:15:00.698697Z
1 likes, 0 repeats
@meeper @dcc @hj @j @lanodan @sun > I read the abstract of one paper which proved the kinda obvious fact that these can be modeled as markov chainsWell, they predict tokens. I'd like to read the paper if you remember who wrote it or you still have it laying around.You're talking about what the things *are*, though, and was carefully avoiding that question and describing their capabilities.
(DIR) Post #B1LQyVdCRdZftbbrLE by leyonhjelm
2025-12-17T16:16:21.386728Z
2 likes, 1 repeats
@p Yep. An imam decided that makes you her son socially. you can drive her alone in a car, work in an office or a home, because of course you can be around your mom. I could probably dig up the fatwa @RedTechEngineer @dcc @deutrino @hj @j @lanodan @sun
(DIR) Post #B1LR9l9Z4gbnDp733Y by p
2025-12-17T16:18:23.379599Z
3 likes, 0 repeats
@meeper @dcc @deutrino @hj @j @lanodan @sun So I can have the robot write some shit and then on the other end they have the robot read some shit and summarize it and I say "Reply to Johnson: we're accepting the proposal." and then the robot generates some shit and emails it to Johnson and then Johnson's robot reads it and says "Hey, Johnson...you're married now."
(DIR) Post #B1LRID1tJbp1zlYJqS by thatbrickster@shitposter.world
2025-12-17T16:19:54.576455Z
2 likes, 1 repeats
@meeper Bullshit, you say? https://blog.kagi.com/llms@dcc @j @p @deutrino @lanodan @hj @sun
(DIR) Post #B1LSAURmlyxXeWY0sy by p
2025-12-17T16:29:43.603811Z
0 likes, 0 repeats
@i @Suzu @dcc @hj @j @lanodan @sun That's still not the name of a model.I used the dolphin bit as an example because it was really terrible; as noted, some of them actually managed it. The one that got it right on the first try was "Mistral Nemo". It seems to lose its shit really quickly, though, so "No, that's not quite what I'm looking for. Isn't there something in the stdlib for this? I'm sure I remember something in the stdlib docs." never works.
(DIR) Post #B1LSR8fczXED2Ixdjs by p
2025-12-17T16:32:44.136010Z
2 likes, 0 repeats
@irie @sun @dcc @j @lanodan @hj > The problem arises when people expect it to understand things, or expect it to "develop" consciousness.You can't stop that. They treat computers as magical and now the computer talks.
(DIR) Post #B1LSclmduUFFWqmvGy by j@annihilation.social
2025-12-17T16:34:49.334857Z
2 likes, 1 repeats
@p @dcc @irie @lanodan @hj @sun10011011111001100100111.jpg
(DIR) Post #B1LSi2eScezvIATVQ0 by NonPlayableClown@postnstuffds.lol
2025-12-17T16:35:44.523433Z
2 likes, 0 repeats
>They treat computers as magical and now the computer talks.Yeah it's been pretty much doomed as such the moment computers were mass marketed.
(DIR) Post #B1LSyeq6PQJmMXqLZ2 by i@declin.eu
2025-12-17T16:38:45.572941Z
1 likes, 0 repeats
@p @dcc @j @Suzu @lanodan @hj @sun it's at the bottom of the pic (https://huggingface.co/moonshotai/Kimi-K2-Instruct) and why it's popular to let the smaller models call a search/text dump rpc for it to crawl through as needed
(DIR) Post #B1LTCWbznO1ytprDyS by lanodan@queer.hacktivis.me
2025-12-17T16:41:16.405671Z
2 likes, 0 repeats
@thatbrickster @dcc @j @p @deutrino @hj @sun @meeper This article is kind of fun because it reads like "Nah, see this software isn't too bad, you just have to be careful about [proceeds to list a bunch of faults and user traps]" which we see a lot in FOSS circles but this time it's LLM version.
(DIR) Post #B1LTwqsclrpB55j3cu by p
2025-12-17T16:49:40.322277Z
2 likes, 0 repeats
@Suzu @i @dcc @hj @j @lanodan @sun > it will present the simplest code it can generate and try to explain it, probably because it assumes the user is trying to learnI'm willing to explain myself to a human, but a *machine* that second-guesses me is infuriating.> I think it handled that pretty well.This one, sure. "Not embarrassing." I don't like the code but that's notation, right, the algorithm is fine. The thing is I expected all of them to be able to do this, really trivially. I have a really low evaluation of these things and I still overestimated them.works.pngit-does-not-print-hello-world.png
(DIR) Post #B1LUIU4MGH0choWraq by p
2025-12-17T16:53:34.888662Z
1 likes, 0 repeats
@leyonhjelm @RedTechEngineer @dcc @deutrino @hj @j @lanodan @sun I almost CC'd you because I thought you'd appreciate it, but you were, as I should have expected, way ahead of me on this one.:saddam: New girl in the office.:arafatno: DIBS! I AM NOT GOING LAST THIS TIME
(DIR) Post #B1LUMdYMwSihHJXxYG by leyonhjelm
2025-12-17T16:54:19.921652Z
1 likes, 0 repeats
@p It is the most based thing Islam could ever do@RedTechEngineer @dcc @deutrino @hj @j @lanodan @sun
(DIR) Post #B1LUSrWxt4x2MjrHxA by lanodan@queer.hacktivis.me
2025-12-17T16:55:26.020368Z
2 likes, 0 repeats
@p @dcc @j @i @hj @sun brainfuck-awk one reminds me that I often wish you could see the tokenisation it does.Like ever used Wolfram|Alpha? Sometimes you get pure bullshit, but it displaying what it assumed and how the tokenisation got done helps a lot on not only getting the limits but also debugging.
(DIR) Post #B1LUZEFwJ92CaJvlku by p
2025-12-17T16:56:36.463513Z
4 likes, 0 repeats
@thatbrickster @meeper @dcc @deutrino @hj @j @lanodan @sunfuck_this_guy-s_website.png
(DIR) Post #B1LUbM7K3GZdXoukeO by Suzu@detroitriotcity.com
2025-12-17T16:56:52.771826Z
1 likes, 0 repeats
@p @dcc @j @i @lanodan @hj @sun can you send me that print one? I want to check what does ChatGPT Pro say about it.
(DIR) Post #B1LUng7lUWsp4QztPU by lanodan@queer.hacktivis.me
2025-12-17T16:59:11.669231Z
2 likes, 0 repeats
@p @dcc @j @deutrino @hj @sun @thatbrickster @meeper Heh that's a fun one, Kagi's blog being Google hosted… one job moment.
(DIR) Post #B1LUsiEz3A2h5Am3yi by i@declin.eu
2025-12-17T17:00:05.777669Z
1 likes, 0 repeats
@lanodan @dcc @j @p @hj @sun a paying customer would have gotten a vm that simply runs the script for an answer, same for people who ask for help with chess positions and get a solver instead of a one token answer
(DIR) Post #B1LVJP7QFkMlayxtSq by lanodan@queer.hacktivis.me
2025-12-17T17:04:55.720571Z
2 likes, 0 repeats
@i @dcc @j @p @hj @sun Yeah, although it should be able to evaluate the answer without running it, specially as a lot of programs loop indefinitely.While if you analyze the program, you can either know the answer, ballpark one, or describe the kind of output.
(DIR) Post #B1LYf8icCzaqxsPwzQ by meeper@udongein.xyz
2025-12-17T17:42:22.711579Z
1 likes, 1 repeats
@p @dcc @j @lanodan @hj @sun Here was the paper I was talking about, honestly I never made more than a cursory glance at the abstract https://arxiv.org/abs/2410.02724
(DIR) Post #B1LZ0Gbq4AYHWDK2Ii by meeper@udongein.xyz
2025-12-17T17:46:11.655100Z
1 likes, 0 repeats
@p @dcc @j @deutrino @lanodan @hj @sun More or less though key point bullshit generation, not summarization so it really can just be used when you want semanticly plausible noise based on the entire internet as a dataset
(DIR) Post #B1LZsuYRlavf2nnAO0 by i@declin.eu
2025-12-17T17:56:09.888662Z
1 likes, 0 repeats
@lanodan @dcc @j @p @hj @sun especially in the case of interpreting, it doesn't see single character operators, it sees clusters of them, and putting single spaces between all of them might actually improve it's output
(DIR) Post #B1LaWXIRev2GUgCa3s by p
2025-12-17T18:03:21.070426Z
1 likes, 0 repeats
@i @Suzu @dcc @hj @j @lanodan @sun Ha, you weren't joking about 1TB.
(DIR) Post #B1Ld2Ey1XgwEoytaXg by Suzu@detroitriotcity.com
2025-12-17T18:31:24.046777Z
1 likes, 0 repeats
@p @dcc @j @i @lanodan @hj @sun wait, is this model split in 61 parts of 17 GB each?Holy shit. How do you even run this thing in consumer grade hardware?
(DIR) Post #B1LdV9mftMEH90NLHc by p
2025-12-17T18:36:43.019379Z
2 likes, 0 repeats
@lanodan @dcc @hj @i @j @sun "It's a Brainfuck program. It must just print 'Hello, World!'. That's what people do with Brainfuck."I tried to use WolframAlpha a few times. Never went past playing it; I don't know if you could check its assumptions at the time.
(DIR) Post #B1LdVJIYWC4weUMQj2 by i@declin.eu
2025-12-17T18:36:42.093652Z
0 likes, 0 repeats
@Suzu @dcc @j @p @lanodan @hj @sun it helps that it's 384*32b models in disguise as a 1T model, so you're really only activating like ~48gb of it at any time and it still fits on a single A/H 100
(DIR) Post #B1LdddXd8ejdkfNkrQ by p
2025-12-17T18:38:14.968610Z
2 likes, 0 repeats
@Suzu @dcc @hj @i @j @lanodan @sun Yeah, sure. I wrote it to have a thing to put on the DevTerm when I uploaded a screenshot: echo '++++[->++++<]>[-<+++++>>+++++++>++<<]<------.>>+++++.--.+.>.<[-<+>>>+<<]>[--<++>>-<]>---.+++++++++++++.+.<<<.>>>-------.---.<<<--.>.>>+++.-------.++.++[->+<<+>]>++++++.<<.<<.>[-<<->>]<<++++.[>>>--<<<-]>>>+.' | \ sed -E 's/(.)/\1\n/g' | \ awk 'BEGIN{print "BEGIN{p=0;"}END{print "}"}/\./{print "printf \"%c\",a[p]"}/\+/{print "a[p]++"}/-/{print "a[p]--"}/</{print "p--"}/>/{print "p++"}/\[/{print "while(a[p]){"}/\]/{print "}"}' | \ awk -f /dev/fd/0
(DIR) Post #B1LdfXxarPcJ1MR9iC by Suzu@detroitriotcity.com
2025-12-17T18:38:30.165375Z
1 likes, 0 repeats
@i @dcc @j @p @lanodan @hj @sun kek, I should hold no delusions of running anything decent in my 9070XT :Kekw:
(DIR) Post #B1LdsyXrKUzBElnrkW by i@declin.eu
2025-12-17T18:40:58.965680Z
2 likes, 0 repeats
@p @dcc @j @Suzu @lanodan @hj @sun
(DIR) Post #B1LdvJIH9pg2AQcy5g by p
2025-12-17T18:41:26.615822Z
1 likes, 0 repeats
@i @lanodan @dcc @hj @j @sun Must be a pretty fun place to work.
(DIR) Post #B1LeHgjWgsbdEb2od6 by Suzu@detroitriotcity.com
2025-12-17T18:45:22.704405Z
1 likes, 0 repeats
@p @dcc @j @i @lanodan @hj @sun I'm also showing the "thinking" part, where it used some novel approach to it:
(DIR) Post #B1LeHymfZvodDYea1o by p
2025-12-17T18:45:32.427799Z
0 likes, 0 repeats
@lanodan @i @dcc @hj @j @sun Yeah, but it's much easier to just run it with a timeout.
(DIR) Post #B1LeMY0GZlqTbgBw0m by caekislove@caekis.love
2025-12-17T18:46:19.794740Z
1 likes, 0 repeats
I hope that website unemploys every Jeet in India
(DIR) Post #B1LeX5HtgEwvjSSjeS by lanodan@queer.hacktivis.me
2025-12-17T18:48:14.634663Z
1 likes, 0 repeats
@p @dcc @j @i @hj @sun Yeah, but I'd say that reduces it's usefulness.Like there's a reason things like linters, static analyzers, … are made the way they are.
(DIR) Post #B1M2J9YtOG4TplImX2 by tyil@fedi.tyil.nl
2025-12-17T23:14:33.026Z
0 likes, 0 repeats
@leyonhjelm@fsebugoutzone.orgI could probably dig up the fatwa Please do, I need this for personal reasons
(DIR) Post #B1M2a0rSzWqjlz1s48 by leyonhjelm
2025-12-17T23:17:43.206761Z
1 likes, 0 repeats
@tyil https://wikiislam.net/wiki/Adult_Suckling#2007_Azhar_fatwasHe got his ass handed to him by other Islamic authorities but I maintain he was correct and his backpedals were under duress.
(DIR) Post #B1MBegIblfcJ2H1d7g by tard@clew.lol
2025-12-18T00:59:23.497506Z
1 likes, 0 repeats
@p @dcc @j @irie @lanodan @hj @sun No children were involved, imaginary or otherwise.
(DIR) Post #B1NvfTGsTsOvyAUTGy by p
2025-12-18T21:09:41.639951Z
0 likes, 0 repeats
@i @lanodan @dcc @hj @j @sun > it doesn't see single character operatorsYeah. Very expensive way to make a computer bad at counting.
(DIR) Post #B1NviK88vHaIUKlUjQ by p
2025-12-18T21:10:12.572366Z
0 likes, 0 repeats
@Suzu @dcc @hj @i @j @lanodan @sun > How do you even run this thing in consumer grade hardware? Doesn't seem like you can.
(DIR) Post #B1NwGfw4MFXFxhUvdA by i@declin.eu
2025-12-18T21:16:23.264418Z
4 likes, 1 repeats
@p @dcc @j @Suzu @lanodan @hj @sun funnily enough, apple has the 512gb unified memory mac studio, so you can drop 40k, rdma the ram over thunderbolt, and have a cluster in less than 300w
(DIR) Post #B1Nwz4gJiLifS4Iin2 by p
2025-12-18T21:24:26.517075Z
1 likes, 0 repeats
@i @Suzu @dcc @hj @j @lanodan @sun Did it have to execute the Brainfuck to do it or did it manage that much with static analysis?> ```nginx:penceshiggy:> Rewrite it as clean Brainfuck.Oh, fuck that stupid fucking robot. This is about as readable as Brainfuck *gets*. You can see, like, first I make 16 so that I can make 0x50 and 0x70 and 0x20, so I'm near most of the alphabet plus I've got a space. After that it's mostly ad hoc but it's not hard to follow to the extent that Brainfuck can be made readable.
(DIR) Post #B1NxXShQvFPRmvDiWe by p
2025-12-18T21:30:39.394837Z
0 likes, 0 repeats
@Suzu @dcc @hj @i @j @lanodan @sun That's kinda hilarious.
(DIR) Post #B1Nxb5E88Qf29CEzaa by p
2025-12-18T21:31:18.664867Z
0 likes, 0 repeats
@caekislove @Suzu @dcc @j @i @lanodan @hj @sun That's not who Microsoft is coming for.
(DIR) Post #B1NyY9qianQLZKSW9I by i@declin.eu
2025-12-18T21:41:56.861558Z
1 likes, 0 repeats
@p @dcc @j @Suzu @lanodan @hj @sun for this one it was purely static, and all it means to make it clean is add newlines
(DIR) Post #B1NyecaGfdX4QQqXgG by p
2025-12-18T21:43:09.327772Z
1 likes, 0 repeats
@lanodan @dcc @hj @i @j @sun Well, yeah.Incidentally, last time I touched Node, there wasn't a "Just do a syntax check" flag. At some point someone wrote a parser. So you could *install* a syntax-checker to get around the stupid chodejs people not having a way to tell their interpreter "Just do the parse step and then stop". Almost any other interpreted language has that, like, "I want to find out if this is syntactically invalid without executing it." So you can save time by shoving it in the build step, you can call it from your editor.I kept a terrible shitty "mozjs" binary around for *years* because you could pass it `-c` and find syntax errors. Eventually it stopped running or building and it didn't support the new syntax extensions anyway, so it was nice when Fabrice did quickjs because quickjs compiles and then you can use it for this shit. `find -name '*.js' -print0 | xargs -0 -n1 -P32 $whatever`, and it takes less than a second even in these ridiculous projects and it's also much easier than waiting for the build system or a runtime error to chug nuts for five minutes and then tell you "Well, shit, the file with the most recent timestamp is missing a goddamn close-paren."nodemodules.jpg
(DIR) Post #B1Nzon6uUK11vGWmhs by p
2025-12-18T21:56:11.909166Z
3 likes, 0 repeats
@tard @dcc @hj @irie @j @lanodan @sun Well, yeah, sure, I didn't think you were buggering Phillipe but you did just dodge the question about if you fucked a GPU.banging_a_website.gif
(DIR) Post #B1O0X60Vak79Gtexbk by lanodan@queer.hacktivis.me
2025-12-18T22:04:10.732763Z
1 likes, 1 repeats
@p @dcc @j @i @hj @sun Somehow never occurred to me that you could use qjsc to effectively syntax-check JS code, but well I also barely ever do JS.And reminds me of why I avoid writing python code, typo on a variable name being a runtime error is tolerable in shell due to things like envvars not being separated, but python has no excuse, specially as perl does it.$ perl choke_me.plGlobal symbol "$typo" requires explicit package name (did you forget to declare "my $typo"?) at choke_me.pl line 7.Execution of choke_me.pl aborted due to compilation errors.$ perl choke_me.pl fooGlobal symbol "$typo" requires explicit package name (did you forget to declare "my $typo"?) at choke_me.pl line 7.Execution of choke_me.pl aborted due to compilation errors.$ python choke_me.pystartdone$ python choke_me.py foobarstartTraceback (most recent call last): File "/home/haelwenn/tmp/2025-12/choke_me.py", line 7, in <module> print(typo) ^^^^NameError: name 'typo' is not defined. Did you mean: 'type'?$ head choke_me.pl choke_me.py==> choke_me.pl <==#!/usr/bin/env perluse strict;print("start");if ($#ARGV == 1) { print($typo);}print("done");==> choke_me.py <==#!/usr/bin/env pythonimport sysprint("start")if len(sys.argv) == 2: print(typo)print("done")
(DIR) Post #B1O80jgZqWXcPRPwUC by p
2025-12-18T23:27:59.843818Z
0 likes, 0 repeats
@i @Suzu @dcc @hj @j @lanodan @sun That's hilarious.
(DIR) Post #B1O8K08sswdcwzqPNQ by p
2025-12-18T23:31:28.749900Z
2 likes, 0 repeats
@i @Suzu @dcc @hj @j @lanodan @sun > for this one it was purely staticThat's impressive.> all it means to make it clean is add newlines awib-0.3.b is the correct use of whitespace in a Brainfuck program.awib-0.3.b
(DIR) Post #B1O8acsYdBBFiv8QZE by p
2025-12-18T23:34:29.017461Z
0 likes, 0 repeats
@lanodan @dcc @hj @i @j @sun > Somehow never occurred to meEver since standalone SpiderMonkey stopped compiling, I had been looking for something.> typo on a variable name being a runtime error is tolerable in shell due to things like envvars not being separated, but python has no excuseHa, I actually don't mind it. Some languages you declare things; some languages you just do whatever.
(DIR) Post #B1OAdrxzBy1pIHLenQ by lanodan@queer.hacktivis.me
2025-12-18T23:57:27.330474Z
1 likes, 0 repeats
@p @dcc @j @i @hj @sun Well, I can be fine with shrugging it off for some, specially for like scripts that aren't really important (roughly half of shell scripts).Python on the other hand is often used for applications, so can easily have entire functions barely exercised and even hard to test.Add the fact it loves to introduce breakages from one release branch to another and it's a recipe for "Dammit, it worked 6 months ago, what now?".It's like they saw the npm ecosystem and wanted it builtin so you can't ignore/reduce the churn.
(DIR) Post #B1OKGJdjeOhNiGHeLI by r000t@ligma.pro
2025-12-19T01:45:15Z
2 likes, 1 repeats
@j @dcc @Suzu @p @lanodan @hj @sun
(DIR) Post #B1PRGoHIdGlHZYvrA8 by p
2025-12-19T14:38:30.547564Z
3 likes, 0 repeats
@lanodan @dcc @hj @i @j @sun > It's like they saw the npm ecosystem and wanted it builtin so you can't ignore/reduce the churn. Or like the only reason the language remained stable was someone was in charge of it, someone felt responsible, and then he was replaced with a committee.That's the thing that still makes no sense to me, like, Python was supposed to be the serious language and looked down its nose at Ruby, but code I wrote for Ruby 1.6 still works in 3.3.
(DIR) Post #B1PYMazhdLGBn3PJqK by p
2025-12-19T15:57:59.390498Z
1 likes, 0 repeats
@r000t @j @dcc @Suzu @lanodan @hj @sun :shodan: *That* is better than I was hoping for.System_Shock_Digital_Horrors_OC_ReMix.mp3