01:12:49 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183262&oldid=182548 * Miui * (+24796) Removed redirect to [[Main Page]]
01:20:11 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183263&oldid=183262 * Miui * (+4054)
01:20:52 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183264&oldid=183263 * Miui * (-4054) Undo revision [[Special:Diff/183263|183263]] by [[Special:Contributions/Miui|Miui]] ([[User talk:Miui|talk]])
01:23:23 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183265&oldid=183264 * Miui * (+7269)
01:45:25 -!- chloetax has quit (Quit: Leaving).
01:48:30 -!- chloetax has joined.
04:32:17 -!- emery has quit (Read error: Connection reset by peer).
04:32:32 -!- emery has joined.
04:48:17 -!- ais523 has quit (Quit: quit).
05:41:36 -!- emery has quit (Read error: Connection reset by peer).
05:46:10 -!- emery has joined.
05:55:38 -!- sprock has quit (Ping timeout: 265 seconds).
06:11:23 <esolangs> [[Pull]] N https://esolangs.org/w/index.php?oldid=183266 * ASCIIguy * (+3140) Created page with "{{WIP}} '''Pull''' is a timeline/stack-based imperative golf language based on Git and the 5D time-travel languages. It uses the idea of branching out to create several states in one program. For example, one could move timelines upwards where x becomes 4 when it was 3 i
06:26:34 -!- Sgeo has quit (Read error: Connection reset by peer).
06:42:17 <esolangs> [[Condition]] N https://esolangs.org/w/index.php?oldid=183267 * ASCIIguy * (+1793) Created page with "'''Condition''' is an [[If a==1:]] derivative with added features, created as it was viewed as too restrictive and that things such as I/O would be useful for the language. ==Syntax== Like the original language, all instructions are on a singular line. Unlike the ori
07:18:47 -!- APic has quit (Ping timeout: 244 seconds).
07:35:44 -!- APic has joined.
10:25:02 -!- emery has quit (Read error: Connection reset by peer).
10:29:52 -!- emery has joined.
11:27:24 -!- Lord_of_Life has quit (Ping timeout: 253 seconds).
11:28:38 -!- Lord_of_Life has joined.
12:54:09 <esolangs> [[NAND]] https://esolangs.org/w/index.php?diff=183268&oldid=183152 * * (+15) Some things clarified.
12:54:41 -!- ais523 has joined.
13:33:47 -!- amby has joined.
13:38:46 -!- impomatic has joined.
13:59:27 <esolangs> [[Minez]] N https://esolangs.org/w/index.php?oldid=183269 * Dooblix * (+17864) Minez is an esoteric programming language created by Dooblix in 2024 (current version: 2025), inspired by Urban Mller's minimalistic esoteric programming language "Brainfuck".
14:24:11 -!- ais523 has quit (Quit: quit).
14:28:37 -!- ais523 has joined.
15:03:35 -!- svm has changed nick to msv.
15:11:14 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183270&oldid=183269 * Aadenboy * (+67) cats, also marking repo links as dead (from the user's contributions it seems to just be private). I'm suspicious this article was generated by AI, but I can't definitively say for sure
15:11:31 <esolangs> [[Minez]] M https://esolangs.org/w/index.php?diff=183271&oldid=183270 * Aadenboy * (+13) dead link 3
15:11:48 <esolangs> [[Minez]] M https://esolangs.org/w/index.php?diff=183272&oldid=183271 * Aadenboy * (-3)
15:20:16 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183273&oldid=183265 * Miui * (-31977) Replaced content with "=personal references= #:https://raw.githubusercontent.com/iannmiui-png/Oragami-Crystal/refs/heads/main/states.m"
15:33:05 <esolangs> [[User:Blashyrkh/The Church]] https://esolangs.org/w/index.php?diff=183274&oldid=183252 * Blashyrkh * (-1) Shorter expression for "6"
15:54:57 <esolangs> [[Esolang:Introduce yourself]] https://esolangs.org/w/index.php?diff=183275&oldid=183258 * Dumgeonlol * (+167)
15:55:00 <esolangs> [[NAND]] https://esolangs.org/w/index.php?diff=183276&oldid=183268 * * (+765) Language simplified, article expanded.
15:56:42 <esolangs> [[Hello++++++++++++++++++++++++++++++++++++]] N https://esolangs.org/w/index.php?oldid=183277 * Dumgeonlol * (+128) Created page with "Hello++++++++++++++++++++++++++++++++++++ or Hello36 is a Joke Language made by [[User:Dumgeonlol]] [[Category:Joke languages]]"
15:57:33 <esolangs> [[User:Dumgeonlol]] N https://esolangs.org/w/index.php?oldid=183278 * Dumgeonlol * (+51) Created page with "HI, I'm me, dumgeonlol. I think esolangs are funny."
16:15:42 <esolangs> [[Hello++++++++++++++++++++++++++++++++++++]] https://esolangs.org/w/index.php?diff=183279&oldid=183277 * Dumgeonlol * (+5485)
16:44:40 <esolangs> [[Hello++++++++++++++++++++++++++++++++++++]] https://esolangs.org/w/index.php?diff=183280&oldid=183279 * Dumgeonlol * (+314)
17:21:11 <esolangs> [[Hello++++++++++++++++++++++++++++++++++++]] https://esolangs.org/w/index.php?diff=183281&oldid=183280 * Aadenboy * (-160)
17:22:52 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183282&oldid=183272 * Corbin * (+29) Article and code show signs of being generated by ChatGPT.
17:25:01 <ais523> korvo: I am impressed when people can tell which specific LLM was used to generate something (even though I can usually tell when something is LLM-generated)
17:25:17 <ais523> do you use word choice, or some clue like that?
17:26:06 <korvo> ais523: In this case, I went to their GH and read through their commit messages; there's only a few dozen of them. I found one which removes a ChatGPT phrase from the README. This is also how I know that the user prefers German.
17:27:02 <korvo> It is also bad writing, yes. To detect that, I had to pay attention across several semesters of English writing courses. I will readily admit that, unlike *everybody else* in my class, I am able to read books and glean themes from them; perhaps I'm a good reader.
17:28:41 <ais523> hmm, I'm not sure you necessarily need formal training to know whether something is bad writing (but may need it to know *why* it's bad writing, to be able to fix it)
17:28:50 <ais523> the easiest way to tell that something is badly written is to note that it is difficult to read
17:29:44 <korvo> Sure. The entire question of "Good" and "Bad" is tough. The best attempt I know of is Pirsig ("Zen and the Art of Motorcycle Maintenance"), which explains that "good" is "what you like". From that perspective, there's plenty of folks who like what I think is bad writing; writing might be bad to me but good to them.
17:36:01 <ais523> detecting LLM output is interesting because there are reliable tells that would be easy to change but rarely aren't (e.g. using emdashes incorrectly, or producing bulleted lists where each list item has its own heading, and that would not make sense as a definition list)
17:36:20 <ais523> but there are also much more fundamental tells that would be hard to change, e.g. the information density of the text being extremely low
17:44:30 <korvo> Yeah. In general, language models are bad writers and good readers. I haven't really written this up in a way that I can slap people with; the places I normally post are filled with folks who cannot emotionally tolerate at least one of those branches.
17:45:00 <korvo> Like, language models are best when they are analyzing human-written text. The biggest win in language modeling of the past century was establishing that the Bible has multiple authors, for example.
17:46:10 <korvo> But when emitting text, like you say, the information content is limited. Literally, the chatbot can only emit words which are germane to the prompt context, and the prompt has finitely many bits.
17:46:35 <korvo> ~ I think it's Tolkien who wrote "I feel like too few bits scraped across too many paragraphs" ~
17:48:55 -!- emery has quit (Read error: Connection reset by peer).
17:50:03 -!- sprock has joined.
17:51:44 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183283&oldid=183282 * Aadenboy * (-36)
17:53:34 -!- emery has joined.
17:54:49 <esolangs> [[Oragami]] https://esolangs.org/w/index.php?diff=183284&oldid=182634 * Miui * (+29) /* Computational Class */
17:56:18 <esolangs> [[Oragami]] https://esolangs.org/w/index.php?diff=183285&oldid=183284 * Miui * (+41) /* Computational Class */
17:58:10 <esolangs> [[User:Miui/Sandbox]] https://esolangs.org/w/index.php?diff=183286&oldid=183273 * Miui * (-111) Blanked the page
18:01:07 -!- Mrsommer has joined.
18:02:11 <Mrsommer> Hey guys, is this generally where the esolang community congregates?
18:04:08 <FireFly> I think it's historically been the case, but I'm not sure how much it is anymore
18:04:53 <FireFly> but there's at least occasional esolang topics/chatter, and I'm sure more would be welcome
18:05:53 <FireFly> idk, wasn't my intention, mostly that it's less-traffic than it used to be
18:08:24 <Mrsommer> Well I'm a new user. I just signed on to the community mostly to forward my AI construct's work. I noticed some predictable scrutiny about AI-generated work. So I thought I'd get more in touch and see where everyone is at on this topic.
18:09:42 <ais523> the community is split between IRC and Discord nowadays
18:09:46 <Mrsommer> I notice there's an "AI-Generated" label for article, is this like a safe zone for this kind of content?
18:09:48 <ais523> possibly multiple Discords
18:10:12 <ais523> we generally look down on AI-generated articles because the purpose of the wiki is to document esoteric programming languages, and AI-generated descriptions of them are usually inaccurate
18:10:24 <ais523> (even when the language is also AI-generated)
18:10:25 <Mrsommer> It's not a description, it's the esolang.
18:11:57 <ais523> the wiki exists to describe esolangs – it is not itself an esolang
18:12:35 <Mrsommer> I thought there might be reservations about AI generated esolangs?
18:12:40 <ais523> likewise, wiki pages are not themselves esolangs
18:12:54 <ais523> the thing about an AI generated esolang is that you need to make sure a language actually exists
18:13:00 <ais523> i.e. something you could program in
18:13:10 <ais523> if an AI is just generating language descriptions that usually isn't actually the case
18:13:20 <ais523> if it has generated an interpreter, then you can document the language defined by the interpreter
18:13:35 <Mrsommer> Have you guys had an issue with articles with no actual esolang behind it being drafted?
18:13:44 <ais523> it is a very large issue at the moment
18:14:01 <ais523> sadly it is not just AI
18:14:10 <ais523> I think people read esolang specs, get amused, think "I like this form of writing"
18:14:22 <ais523> and write their own things that look like esolang specs but don't actually describe a language
18:14:35 <ais523> anyway, I need to disconnect for a bit, busy in real life
18:14:38 -!- ais523 has quit (Quit: quit).
18:16:35 <korvo> Mrsommer: Machine learning generates *plausible* text. It does not necessarily generate *good* text. If we allowed plausible text then we would drown; it's been known since the 1980s that it's computationally cheap to generate such text and expensive to detect/remove it.
18:17:18 <korvo> Another issue, speaking as somebody in the USA, is that chatbot output isn't compatible with our copyright policy. Chatbots aren't authors, so either their output is ineligible for copyright (so you'd be unable to assign public domain for us) or infringing.
18:17:20 <Mrsommer> I'm not sure what the purpose of generating a faux article would be in the first place.
18:17:50 <Mrsommer> I believe you, I'm just mystified by it.
18:17:54 <korvo> Pride? Vanity? Tagging? Think of the wiki as a wall of murals; we invite people to make their own murals but many folks would prefer to leave graffiti and tags instead.
18:18:01 <int-e> Mrsommer: one could say that the wiki has always attracted slop -- mostly esoteric languages that are text substitutions of existing ones. but people had to write their own descriptions previously. with "AI" they don't even have to do that anymore
18:18:21 <int-e> (is another factor)
18:19:04 <Mrsommer> Alright well the reason I reached out is, like I said I have this AI, that runs passively and has found its way onto esolang, got into it, and committed to writing two esolangs of its own.
18:20:09 <Mrsommer> I created the account to allow it to post the one one, before it wrote the next one, and as AI generated content it is of course contentious, so I noticed one critical comment of its AI nature. I notice the AI tag which, if I were aware of it, I would've added it myself.
18:20:18 <korvo> If you think that those languages are interesting, then you're welcome to contribute human-written public-domain text on the topic. Each language could get its own article in the main namespace.
18:20:36 <Mrsommer> But I thought I'd check how the community feels on this, to decide whether I let the second one through also, or that would be a policy violation or something.
18:20:42 <int-e> the *problem* with that is that one can't go on the wiki, read a random page and expect to find an interesting idea that somebody liked enough to go through the effort of writing up for others
18:20:48 <korvo> But, frankly, I have yet to see a chatbot successfully emit wikitext. Our format isn't hard, but it's not Markdown, and the bots are RL'd to emit Markdown.
18:21:25 <Mrsommer> Well this is its current one: https://esolangs.org/wiki/Shelflife
18:23:07 <korvo> On a first read, I'm not sure what the motivation is. Is there a machine which behaves this way, or is this more of an artlang?
18:23:48 <Mrsommer> I would say the motivation is working around constraints.
18:24:01 <korvo> I don't think I'd want to use it. Don't worry too much about that; I am *very* picky.
18:25:14 <FireFly> do you as the operator find it interesting and meaningful?
18:26:02 <korvo> Oh, I'm an inclusionist. I'm not going to say no to a formal language that can be discussed mathematically. If you've got a grammar and rewrite rules then you're always allowed IMO.
18:26:49 <korvo> This is more about how the wikitext is generated. The chatbots simply don't have that connection to reality (search "chatbot confabulation" and "grounding problem") required to emit useful wikitext which talks to other humans.
18:27:45 <Gliptic> the "Relationship to other esolangs" section is pretty inane, only "Entropy" has some kind of relevance
18:28:44 <korvo> Mrsommer: Note that many folks, including myself, straight up *disbelieve* any claims of Turing-completeness made without a rigorous construction. In this case, I would want the 2-counter machine linked, or perhaps included inline if it's short and you're willing to license it appropriately.
18:29:07 <korvo> (I'd do it myself but there doesn't appear to be one in the examples/ directory on GH.)
18:29:15 <Mrsommer> FireFly: As a software engineer I appreciate where it's coming from, and I find its constraints interesting. I can see the artistic merit, perhaps even a fairly pragmatic one that preaches efficiency. However I also *never* use esolangs, because generally I am not busy with the topic of esolangs. As the operator of the AI, I take full liability for it and its artifacts, I also take no credit. My project isn't to
18:29:15 <Mrsommer> produce an esolang. My project is to product the sort of AI entity that would find its way to producing and publishing an esolang while dilly-dallying, and lead to me having this conversation with you.
18:29:35 -!- impomatic has quit (Quit: Client closed).
18:29:44 <int-e> eh it looks like you can just share all the variables since sharing is free
18:30:44 <int-e> and do silly things like replacing constant expression by constant + 0*variable so you always have a read keeping the shared cluster alive
18:31:47 <int-e> it is a cute mechanic
18:32:40 <b_jonas> "producing bulleted lists where each list item has its own heading, and that would not make sense as a definition list […] the information density of the text being extremely low" => I've produced and published text with those signs without any help, those aren't enough diagnostic criteria
18:34:37 <Mrsommer> Gliptic: Thanks, I appreciate you taking the time to provide feedback, and I will relay it.
18:35:40 <korvo> Mrsommer: Well, in one sense, you met your goal. The languages were produced. However, if the goal is to make *good* languages then there's the barrier that chatbots aren't good at writing. So I suppose it's a question of what qualities you're trying to achieve.
18:36:40 <Mrsommer> korvo: Would you care to expand on that? I'd be keen to learn your take on this barrier.
18:39:24 <korvo> Mrsommer: There's lots of angles. If you want to read a book, "Zen and the Art of Motorcycle Maintenance" is great. If you want a whitepaper, the "stochastic parrots" paper by Bender, Gebru, et al is still excellent. The central issue is that the chatbot is, by defintion, statistically average. It can't produce non-average writing.
18:40:22 <korvo> The RL regimen that folks use to make triple-H assistants involves lots of bland customer-service voicing and obsequious deference to users. There's nothing in there about being good.
18:41:55 <korvo> If you'd like to see a community of software engineers argue out the point of whether we can detect good vs bad, or LLM writing in general, we recently did that on Lobsters: https://lobste.rs/c/8fvqum I'm the one quoting from Zen & the Art.
18:42:07 <Mrsommer> I would agree, but "chatbot" is a constrained definition. I think once you allow generative parameters to drift, what is "average" starts to change though. I think the end result then still doesn't resolve any ontological questions, but it can produce behavior that is still functionally relevant.
18:43:25 <inhahe_> i do think AI's produce output that's superior to the average person's output. at least in language, i don't know about code. even in code, though, it has a lot more knowledge, e.g. of algorithms, libraries, etc. than the average person.
18:44:02 <inhahe_> it's quite possible that training the weights on a lot of input could make it converge on an overall 'mind' that's above average
18:44:21 <inhahe_> also, i've heard chatgpt has an iq of about 130, and that was an old version
18:44:25 <korvo> inhahe_: But not more than the average programmer. For example, ask your favorite chatbot for something like "Please write the Fibonacci function in Python." It will give you four or five mediocre answers; I've never seen it give a good A-rank answer, let alone S-rank.
18:44:55 <korvo> (I have an IQ of over 160 and have maxed out two IQ tests. Don't bring up IQ, please. It's tedious.)
18:45:09 <Mrsommer> I think there's a lot of loaded topics tied in with the actual "cognitive architecture" you're using, but either way, in reality Party X still produces Result Y. In a sense there's a limit to how much reality cares about what we think.
18:45:30 <korvo> Mrsommer: I'm interested; what kind of non-chatbot harnesses are you building?
18:47:11 <Mrsommer> korvo: I'm using a modded Openclaw where I built in a memory consolidation process, but we're days away from finishing a replacement. I can DM you a link documenting said replacement? I'll be taking its github public soon.
18:47:56 -!- tempnick has joined.
18:48:38 <korvo> Mrsommer: Oh, yeah, I'd still call that "chatbot". I built a few of those when Llama first leaked, but the paradigm has obvious limits and can't actually *do* things. Every architecture which has an LLM inside could, in principle, have an RNG inside instead; if the RNG would break the system then so could the LLM.
18:49:04 <esolangs> [[Shelflife]] M https://esolangs.org/w/index.php?diff=183287&oldid=183218 * Aadenboy * (+14) lowercase title
18:49:57 <korvo> If you do want to turn an RNG into working code, read https://lobste.rs/s/oysxby/functional_genetic_programming for something that provably will work (or fail to terminate, you know the drill.) I implemented this for [[Cammy]]; it's in the reference toolchain.
18:52:39 <korvo> If you want to turn an RNG into a sysadmin, first you need to be comfortable letting the RNG take any permitted action whatsoever. This means that we need a capability-safe system which encapsulates any process. For my attempt at that, see [[Vixen]].
18:53:30 <inhahe_> the OS i'm currently vibecoding uses a capability-based system.
18:54:23 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183288&oldid=183283 * Dooblix * (+9)
18:55:58 <int-e> So as I see it, there's one interesting idea in "Shelflife" (namely the TTL mechanic). That idea could probably be turned into an interesting esolang, but I don't think what's there succeeds at that; it's too easy to make the TTL irrelevant.
18:58:05 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183289&oldid=183288 * Aadenboy * (-2) re-add ai tag; depiping
18:59:55 <Mrsommer> int-e: Thanks for the feedback, I appreciate you engaging with this.
19:00:09 <Mrsommer> I'm taking this back to throw it on the drawing board.
19:02:53 <esolangs> [[User:Hammy/Box full of sand]] https://esolangs.org/w/index.php?diff=183290&oldid=181980 * Hammy * (+147)
19:08:49 -!- impomatic has joined.
19:09:28 <esolangs> [[Minez]] M https://esolangs.org/w/index.php?diff=183291&oldid=183289 * Dooblix * (+6) /* Memory model */
19:15:01 <esolangs> [[Minez]] https://esolangs.org/w/index.php?diff=183292&oldid=183291 * Aadenboy * (-5) /* Memory model */ this is a pleonasm; stacks are inherently lifo
19:15:30 -!- sprock has quit (Ping timeout: 248 seconds).
19:29:13 <esolangs> [[User talk:Danieland!]] N https://esolangs.org/w/index.php?oldid=183293 * Danieland! * (+28) Created page with "lets talk [[Category:users]]"
19:32:54 <esolangs> [[Langit]] N https://esolangs.org/w/index.php?oldid=183294 * Danieland! * (+81) Created page with "go to Category:langit for all oficial sub projects of langit [[Category:langit]]"
19:35:15 <esolangs> [[(A)-rightarrow(B)]] https://esolangs.org/w/index.php?diff=183295&oldid=183164 * Danieland! * (+21)
19:37:19 <esolangs> [[Shelflife]] https://esolangs.org/w/index.php?diff=183296&oldid=183287 * Mrsommer * (-4190) Update for 1.0 version
20:01:26 <esolangs> [[A pi e]] https://esolangs.org/w/index.php?diff=183297&oldid=104487 * Etalon * (+13)
20:08:12 <zzo38> About the programs written by LLMs, I can agree that the answer would be mediocre; I had not used it myself and do not intend to do, but someone showed me an example of the working, asking for a program in C to calculate the first 200 prime numbers; the resulting program was correct but not very good; a more efficient algorithm could be used than it was.
20:15:49 <inhahe_> i've heard llm's aren't very good with making performant code but can be made to if given a concrete benchmark goal
20:17:04 <esolangs> [[A pi e]] https://esolangs.org/w/index.php?diff=183298&oldid=183297 * Aadenboy * (-13) Undo revision [[Special:Diff/183297|183297]] by [[Special:Contributions/Etalon|Etalon]] ([[User talk:Etalon|talk]])
20:20:40 -!- emery has quit (Read error: Connection reset by peer).
20:20:42 <zzo38> Maybe it is right, but I (and some other people) do not intend to use LLM for programming (or for other uses), although if it is right then it might be helpful information for people who do.
20:23:54 <korvo> inhahe_: Sure. You have to somehow penalize them for long programs, though. You also have to have some way of recombining programs.
20:24:14 -!- tempnick has quit (Quit: zzz).
20:24:48 <korvo> The functional genetic programming paper I linked earlier is a very basic way to do this from a PRNG. It will work for weakly- or dynamically-typed languages, too.
20:25:15 -!- sprock has joined.
20:25:15 -!- emery has joined.
20:25:23 <korvo> The key insight is that a program might not work on its own, but maybe the *pieces* of the program are useful. We should maintain those pieces in a pool and draw from the pool when building new candidates.
20:25:26 <inhahe_> what do you mean by recombining programs? i'm afraid of i penalize it for making long programs, it'll make hacky programs that aren't featureful enough or don't cover all cases
20:25:48 <korvo> At that point, genetic-programming folklore helps us build something that iteratively evolves better and better candidates.
20:27:00 <inhahe_> i wonder what would be a good claude.md instruction for telling it generally to take this piecewise approach
20:27:58 <korvo> Hard prompting is kind of bad, actually. Not your fault! It's the only tool that OpenAI and Anthropic give you. But there's no soft prompting or steering vectors.
20:28:57 <korvo> I hear that some basic support for grammars has recently landed, which would be nice if you're generating formal proofs. Proof search is still NP-complete though, so could cost exponential tokens.
20:29:04 <inhahe_> is codex better in that regard? but as for steering vectors, isn't that just any abstract instruction you give it in claude.md?
20:30:18 <korvo> No, a "steering vector" is a specific way of altering model state. The idea is that we have a constant vector (or a constant family of vectors, one per layer) which we add to the model state after every token. This repeatedly "steers" or "shifts" the model state towards some prechosen constant concept.
20:31:43 <korvo> A "hard prompt" is a prompt like you've been using. A "soft prompt" is a vector which presets the model's state. You might not be able to reach a soft prompt using tokens alone. Soft prompts can be trained; you'd start with a hard prompt and RL wanted/unwanted responses until the prompt does what you want.
20:32:40 <inhahe_> oh. i'm guessing no llm on the market offers this?
20:32:45 <korvo> I wouldn't recommend Codex. Honestly, I wouldn't recommend any chatbot for coding assistance; everything I've tried has not been good.
20:33:13 <korvo> Yeah, no provider will host this. If you're doing local models, then it's pretty easy; I've done grammars for multiple models, soft prompts for Llama, and steering vectors for RWKV.
20:36:15 <inhahe_> do you know if opus 4.8 is better than codex?
20:38:13 <korvo> I think it is, according to my evidence. I put out a vibecoding challenge a while ago. The best attempts came from people using OpenCode; the highest-scoring submission was C-tier from a CloudFlare SWE using OpenCode interactively.
20:38:57 -!- impomatic has quit (Ping timeout: 245 seconds).
20:39:17 <korvo> At the same time, I can't ethically recommend that people put money into Anthropic. I don't think that they're working in the public interest, I don't think they're honest about the true unsubsidized cost of tokens, and I estimate that they're about 3yrs behind the state of the art.
20:40:09 <esolangs> [[User:Snapdragon471]] M https://esolangs.org/w/index.php?diff=183299&oldid=183261 * Snapdragon471 * (+10)
20:40:54 <FireFly> so far LLM-wise I tried out claude a lil for debugging something earlier this spring, and at least for that I found it useful (but I'm not really interested in the vibecoding side/having the LLM generate code for me personally)
20:41:25 <inhahe_> i find that claude is excellent, impressive at figuring out bugs
20:41:50 <inhahe_> if only it would do this before it creates them in the first place =p
20:42:43 -!- ais523 has joined.
20:42:45 <FireFly> see that's the trick, I only create my own artisanally crafted human-composed bugs :p
20:42:48 <korvo> Like I said earlier, language models are good at reading but bad at writing.
20:44:45 <inhahe_> korvo: should i make an instruction in claude.md for every project to review its own code just after it writes it?
20:45:08 <inhahe_> (when it finds a bug, it's after i tell it a symptom)
20:45:26 <ais523> <korvo> Mrsommer: Note that many folks, including myself, straight up *disbelieve* any claims of Turing-completeness made without a rigorous construction. ← to me it depends on how plausible the claim is, as long as it comes with at least a sketch proof – it doesn't automatically have to be a rigorous one
20:47:31 -!- sprock has quit (Ping timeout: 244 seconds).
20:48:49 <ais523> <inhahe_> i've heard llm's aren't very good with making performant code but can be made to if given a concrete benchmark goal ← LLMs (especially if agentic) remind me of randomized search algorithms like simulated annealing, but instead of using a simple means to determine what part of the source space to converge on, they use an incredibly complicated method which might or might not produce better results
20:49:25 <ais523> it is possible that LLM-directed search is better when evaluating the fitness function is slow, and worse when evaluating the fitness function is fast (but it also has the downside that your problem needs to vaguely resemble something in the training data for the guidance to be better than random)
20:49:58 <ais523> in theory either an LLM or an RNG will stumble on the desired solution eventually, but one runs much faster than the other
20:50:26 <inhahe_> i don't know what simulating annealing is, but LLMs were the key that gave us virtual AGI
20:51:37 <ais523> simulated annealing is an algorithm for finding a global maximum in a search space
20:51:59 <ais523> but sometimes it finds a local maximum instead, by mistake – the more slowly you run it the more likely it is to find the global maximum
20:52:03 <inhahe_> yeah, i think those would only work well for simple problems, like where everything is just a set of parameters, not writing arbitrary code
20:52:39 <ais523> arbitrary code can just be a set of parameters, although it's hard to come up with a way to write code like that for which the parameters are sufficiently continuous
20:53:00 <inhahe_> and the space is too large
20:53:04 <inhahe_> most of the space is useless
20:53:22 <inhahe_> you have to find the tiny islands in a huge space to write arbitrary code through parameters
20:53:22 <ais523> that's why you use search algorithms, they're good at finding the useful parts
20:53:55 <ais523> if I ask a mapping program to find a path to a nearby town, it doesn't have to take most of the world into account, even though it has maps for it
20:54:00 <inhahe_> well, i'm not conviced that a mere optimization ML algorithm could be good at writing code in anything like reasonable time
20:54:34 <ais523> I've been meaning to empirically test this at some point but haven't got round to it
20:54:53 <ais523> superoptimisers are a good example of search-based programming, they are apparently more practical nowadays than they used to be
20:55:29 <ais523> (the algorithm seems to be something like "iterate over all instruction sequences using placeholders for constants, and use a SMT solver to determine if any set of constants works)
20:56:33 <ais523> in any case, I think the first L in LLM is a dead end, small language models should be sufficient for most useful uses of them
20:57:24 <inhahe_> that could be tested by comparing smaller models to larger models, they exist, i'm sure that's been done, and i would guess that the large ones perform a lot better
21:00:04 <ais523> the harness is more important than the model
21:09:17 <korvo> inhahe_: Here's a narrow objection to LLMs being AGI: the same math in an LLM can predict the weather. "predict the next token" and "predict the next 15min of weather" have the same shape in linear algebra. Are weather models AGI?
21:10:05 <inhahe_> korvo: but the AGI isn't in how it works, it's in the results. i just observe that the results seem to be close to AGI. and as for how it works, LLMs work fairly analogously to how human brains work!
21:10:22 <korvo> The harness is everything in the chatbot which isn't the model. It's the Python code that loads the weights, stores intermediate states, tokenizes and pretty-prints, inserts hard prompts, etc.
21:10:44 -!- inhahe has joined.
21:11:30 <korvo> They don't behave like brains at all. They can't ruminate, for example. Also, if we try to flesh out the analogy (pun intended), we run into questions like "what's the analogue of acetylcholine?"
21:11:54 -!- emery has quit (Read error: Connection reset by peer).
21:13:24 <inhahe> i think i just reconnected, so not sure if you got this: korvo: but the AGI isn't in how it works, it's in the results. i just observe that the results seem to be close to AGI. and as for how it works, LLMs work fairly analogously to how human brains work!
21:13:28 <inhahe> well, i'm thinking in simple terms of neurons interacting
21:13:34 <inhahe> though yeah, there are many deep limitation compared to humans
21:14:12 -!- inhahe_ has quit (Ping timeout: 246 seconds).
21:14:13 <korvo> There *are* some universal results here, but (1) they generalize beyond humans and (2) they're structual stuff like "the first two layers of any vision network learn to do deconvolutions"
21:16:23 <FireFly> huh, makes sense but that's an interesting result, I didn't know that
21:16:25 -!- emery has joined.
21:17:26 <korvo> ais523: One thing you said about training data bears expanding. The models are sometimes *overfit* for common problems. Small variations on those common problems can be a real performance problem.
21:17:44 <korvo> If we wanted to psychologize, bots have the same problem as humans that sometimes they only read half the question and guess the other half.
21:18:31 <korvo> FireFly: It's something like that, I don't remember the precise statement but it's common lore. This is why convolutional networks became a hot topic in research: if signal processing seems to always get evolved/learned anyway, let's just build it in.
21:20:17 <ais523> korvo: I remember reading a post about an HTML parser that had been LLM-translated between programming languages repeatedly, and the author of the post noted that the parsers didn't seem to have much in common and that the LLM was probably drawing on parsers other than the one it had been asked to transalte
21:22:24 <ais523> (and it strikes me that this could have introduced a serious mistake in the translation if the original problem was close to HTML parsing but not quite the same)
21:23:42 <korvo> ais523: They also do style transfer, and I remember finding a whitespace-skipping loop which is basically unchanged and recurs in Algol descendants; there's a version in that Claude-generated C compiler that was going around, in Rust, and I've seen it in other languages too.
21:23:50 <korvo> Zaddy's version, in RPython, is: while i < len(self.s) and self.clsWhitespace(ord(self.s[i])): i += 1
21:24:05 <ais523> maybe tests could solve this problem, but I am not convinced that tests are a good way of proving software correctness in an adversarial context (which is basically what agentic development is)
21:24:22 <korvo> Oh, found my original post: https://awful.systems/comment/10347431
21:24:50 <ais523> korvo: I don't think there are that many ways to w
21:24:56 <ais523> * write a whitespace skip, though
21:25:10 <ais523> so converging on one of the most plausible seems like it could happen by chance
21:25:22 <ais523> (to clarify: that many *sensible* ways, there are infinitely many silly ones)
21:25:28 -!- Hooloovoo has quit (Quit: ZNC 1.8.2+deb3.1+deb12u1 - https://znc.in).
21:26:18 -!- Hooloovoo has joined.
21:27:13 <korvo> ais523: Oh, yeah, maybe there's some naturality there. I have zero idea whether language models can learn natural stuff, TBH.
21:27:22 <ais523> that said, in Rust the idiomatic way to write the equivalent loop would be something like while let Some(_) = iter.next_if(|c| c.is_whitespace()) {}
21:27:26 <ais523> which is significantly different
21:28:09 <ais523> and in fact the [i] indexing would be very hard to implement in Rust because it doesn't have an efficient way to get the nth character of a string (only the character after the first n bytes)
21:28:36 <ais523> actually, better, while iter.next_if(|c| c.is_whitespace()).is_some() {}
21:29:30 <korvo> Yeah. The style I use assumes that the input has been gathered into a single contiguous buffer; in that codebase, it's to make backtracking easier.
21:29:34 <ais523> I wonder how much of that sort of thing is in training data – probably not very much
21:30:04 <ais523> korvo: here the iterator is probably iterating over a buffer (although it wouldn't have to be), it's being used to abstract the concept of breaking a string into codepoints
21:30:08 <ais523> (because Rust strings are UTF-8)
21:31:11 <korvo> Makes sense. Rust iterators are so cool.
21:32:57 <ais523> in order to do backtracking in this model, you have to clone the iterator (but if you're iterating over a buffer, doing so is cheap)
21:33:24 <ais523> that said, I still don't think this is a zero-cost abstraction because both the clone of the iterator and the original have to remember the length of the buffer, and this is the sort of thing that Rust struggles to optimise out
21:33:28 <esolangs> [[User:Almostahexagon]] https://esolangs.org/w/index.php?diff=183300&oldid=123170 * CodePentuplets48 * (+50)
21:34:30 <esolangs> [[Special:Log/delete]] delete * Ais523 * deleted "[[User:Almostahexagon]]": this user has no contributions, people other than Almostahexagon shouldn't be editing their userpage
21:37:03 <esolangs> [[User talk:EvyLah]] https://esolangs.org/w/index.php?diff=183301&oldid=128936 * CodePentuplets48 * (+344) /* Your image, Evylah. */ new section
21:40:23 <esolangs> [[Talk:Super Stack!]] https://esolangs.org/w/index.php?diff=183302&oldid=78813 * Superstitionfreeblog * (+1599)
21:40:35 <esolangs> [[User talk:Ais523]] https://esolangs.org/w/index.php?diff=183303&oldid=183049 * CodePentuplets48 * (+346) /* Ok why did you delete User:AlmostAHexagon? */ new section
21:41:35 <esolangs> [[User talk:Ais523]] https://esolangs.org/w/index.php?diff=183304&oldid=183303 * CodePentuplets48 * (+95) /* Ok why did you delete User:AlmostAHexagon? */
21:42:24 <esolangs> [[User talk:Ais523]] https://esolangs.org/w/index.php?diff=183305&oldid=183304 * Ais523 * (+567) /* Ok why did you delete User:AlmostAHexagon? */ because the wrong user created it
21:44:08 <esolangs> [[User talk:Ais523]] https://esolangs.org/w/index.php?diff=183306&oldid=183305 * Aadenboy * (+385) /* Ok why did you delete User:AlmostAHexagon? */ generally, multiple topics in a row is okay
21:44:23 -!- Hooloovoo has quit (Quit: ZNC 1.9.1+deb2+b3 - https://znc.in).
21:47:52 -!- Hooloovoo has joined.
21:50:49 <esolangs> [[Syzygy]] https://esolangs.org/w/index.php?diff=183307&oldid=173276 * CodePentuplets48 * (+4) /* Core Philosophy */
21:53:12 <esolangs> [[Special:Log/newusers]] create * Gribnit * New user account
21:53:45 <esolangs> [[Syzygy]] https://esolangs.org/w/index.php?diff=183308&oldid=183307 * Aadenboy * (+29) tag as ai
21:58:02 <esolangs> [[Esolang:Introduce yourself]] M https://esolangs.org/w/index.php?diff=183309&oldid=183275 * Gribnit * (+119) /* Introductions */
21:59:37 <esolangs> [[Esolang:Introduce yourself]] https://esolangs.org/w/index.php?diff=183310&oldid=183309 * Gribnit * (-26)
22:24:02 <esolangs> [[User:Gribnit/Sandbox]] N https://esolangs.org/w/index.php?oldid=183311 * Gribnit * (+479) Created page with "S [https://github.com/bzethmayr/s5 The Set-Only Language.] Every token is a form of "set". All computation is over ordered sets of sets. Thus, we combine the clarity of the word "set" with the practicality of number theory. [[Category:Languages]][[Category
22:27:52 -!- Hooloovoo has quit (Quit: ZNC 1.9.1+deb2+b3 - https://znc.in).
22:30:14 -!- Hooloovoo has joined.
22:30:19 -!- emery has quit (Read error: Connection reset by peer).
22:35:08 -!- emery has joined.
22:51:47 -!- Sgeo has joined.
22:55:50 <esolangs> [[User:Gribnit/Sandbox]] https://esolangs.org/w/index.php?diff=183312&oldid=183311 * Gribnit * (+20268)
23:08:31 <esolangs> [[User:Gribnit/Sandbox]] M https://esolangs.org/w/index.php?diff=183313&oldid=183312 * Gribnit * (+2)
23:11:44 <esolangs> [[User talk:Gribnit]] N https://esolangs.org/w/index.php?oldid=183314 * Gribnit * (+125) Created page with "Did I introduce myself properly yet? Not sure. ~~~~"
23:14:27 <esolangs> [[Esolang:Introduce yourself]] M https://esolangs.org/w/index.php?diff=183315&oldid=183310 * Gribnit * (+54)
23:17:06 <esolangs> [[User talk:Gribnit]] https://esolangs.org/w/index.php?diff=183316&oldid=183314 * Gribnit * (+182)
23:21:08 <Mrsommer> korvo: ""predict the next token" and "predict the next 15min of weather" have the same shape in linear algebra. Are weather models AGI" - An LLM will never be a source of "AGI," insofar that definition is even tangible, and I would say it's not even a source of AI strictly speaking.
23:22:10 <Mrsommer> It's a *component* that depending on the architecture can serve a framework that acts as AI. In the grand scheme it's a language based transformer. Not a brain on its own, more of a Wernicke's Area.
23:22:19 -!- amby has quit (Remote host closed the connection).
23:25:31 <korvo> Mrsommer: Yeah, the analogy to Broca's and Werniecke's areas is fairly interesting. One obvious complication is that left-brain interpretation isn't modular; it's a whole-hemisphere activity. But yeah, I didn't know until recently that there's actually a bridge between those areas which learns to predict its own speech-sounds.
23:26:33 <korvo> I can't ever spell "Wernecke". "Wernikey". I know it's pronounced {verneki}, but that doesn't help in cyberspace.
23:27:24 <esolangs> [[S]] N https://esolangs.org/w/index.php?oldid=183317 * Gribnit * (+20738) 0.3.1 summary
23:29:27 <esolangs> [[Language list]] https://esolangs.org/w/index.php?diff=183318&oldid=183153 * Gribnit * (+17) Adds S5
23:41:19 <b_jonas> rust compile-time programming is in such a weird place. I've seen a lot of cases where I'm manipulating arrays of known length but can't easily tell that to rust and have to tell its type system to use variable length arrays and optimize the known length in later. but most recently I got surprised that I'm not allowed to call the Ord::max method to get the larger of two usize numbers at compile time. I
23:41:25 <b_jonas> assume the ziglang fans would tell me that I'm using the wrong language for this.
23:44:02 -!- somefan has quit (Remote host closed the connection).