2026/04/03

Newest at the top

2026-04-03 03:01:19 +0000 <monochrom> Haven't you heard? Those LLMs that have learned from history are doomed to repeat it!
2026-04-03 03:00:44 +0000 <geekosaur> ^
2026-04-03 03:00:20 +0000 <Leary> mesaoptimizer: LLMs do that because they ate a huge amount of text written by humans who did that.
2026-04-03 03:00:17 +0000 <EvanR> butlerian jihad!
2026-04-03 02:59:42 +0000 <monochrom> "I do not trust it when a computer says 1+1=2 because it's a stupid computer!"
2026-04-03 02:59:37 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 272 seconds)
2026-04-03 02:59:37 +0000 <mesaoptimizer> okay, *at least* you aren't saying I'm wrong, I do appreciate that lol
2026-04-03 02:58:57 +0000 <monochrom> or perhaps ad machinem?
2026-04-03 02:58:54 +0000 <EvanR> that article isn't about LLMs or AI at all so
2026-04-03 02:58:44 +0000 <mesaoptimizer> erm, I have been paid to do research engineer work on neural networks, and have worked with researchers now at DM and Anthropic
2026-04-03 02:58:39 +0000 <monochrom> ad machina >:)
2026-04-03 02:57:31 +0000 <EvanR> we need a new latin fallacy like ad hominem but for attacking the LLM
2026-04-03 02:57:22 +0000 <mesaoptimizer> it is quite common for an LLM to add a concrete, supposedly humorous exageration as part of its triadic description for something, and you can see it with "a small municipal government's worth of bookkeeping".
2026-04-03 02:56:13 +0000 <mesaoptimizer> sidenote: "Your database layer uses connection pooling, retry logic, and mutable state internally. Your cache uses concurrent mutable maps. Your HTTP client probably has circuit breakers, pooled connections, and a small municipal government's worth of bookkeeping." this is a clear indicator that an LLM was used to write this essay
2026-04-03 02:55:45 +0000bitdex(~bitdex@gateway/tor-sasl/bitdex) bitdex
2026-04-03 02:54:56 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn
2026-04-03 02:54:22 +0000weary-traveler(~user@user/user363627) user363627
2026-04-03 02:45:39 +0000 <mesaoptimizer> EvanR: it started off quite interesting, yes, which is why I even put in the effort to ask about the thing that confused me, so yes, I am reading onwards.
2026-04-03 02:45:00 +0000arandombit(~arandombi@user/arandombit) arandombit
2026-04-03 02:45:00 +0000arandombit(~arandombi@2a02:2455:8656:7100:cd4b:38a2:fba4:622b) (Changing host)
2026-04-03 02:45:00 +0000arandombit(~arandombi@2a02:2455:8656:7100:cd4b:38a2:fba4:622b)
2026-04-03 02:44:56 +0000 <mesaoptimizer> ... sorry, that was for davean
2026-04-03 02:44:47 +0000 <mesaoptimizer> EvanR: I disagree about that claim when in a cognitively adverserial environment
2026-04-03 02:44:15 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 255 seconds)
2026-04-03 02:44:02 +0000 <davean> Yah the obligation is on you to actually think clearly about it.
2026-04-03 02:43:57 +0000 <EvanR> (there's a lot in there that is haskell specific, but as far as groundbreaking insights maybe not many)
2026-04-03 02:43:23 +0000 <EvanR> otherwise don't bother
2026-04-03 02:43:17 +0000 <EvanR> ok if you're going to analyze the article then yeah maybe read it
2026-04-03 02:42:58 +0000arandombit(~arandombi@user/arandombit) (Ping timeout: 248 seconds)
2026-04-03 02:42:56 +0000 <mesaoptimizer> so it still stands. It is not specific to Haskell.
2026-04-03 02:42:30 +0000 <mesaoptimizer> Leary: perhaps I misused the word, but my point was about semantic safety, in the sense of Benjamin Pierce's definition in TAPL
2026-04-03 02:42:19 +0000 <EvanR> as you wish
2026-04-03 02:41:16 +0000 <mesaoptimizer> EvanR: not fully, of course. I imagine maybe the author wrote a bunch of bullet points up and then used an LLM to expand upon it. But all right, I'll give the essay another shot and continue reading it
2026-04-03 02:40:40 +0000 <Leary> mesaoptimizer: You expect wrong; people use 'pure' just to mean 'pure'; there's no implicit /by construction/.
2026-04-03 02:40:38 +0000 <EvanR> none it seems especially controversial either
2026-04-03 02:40:25 +0000 <EvanR> tales from tech
2026-04-03 02:40:19 +0000 <EvanR> we've seen reports like this for years
2026-04-03 02:40:00 +0000 <mesaoptimizer> my conclusion is that the essay above is generated using a SOTA LLM, and therefore is adverserially generated to seem insightful, but actually be dangerously wasteful of human time
2026-04-03 02:39:57 +0000 <EvanR> it's classic "dev speak"
2026-04-03 02:39:47 +0000 <EvanR> that article isn't LLM style at all
2026-04-03 02:39:33 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn
2026-04-03 02:39:19 +0000 <mesaoptimizer> I accept the existence of parts of a language that are delineated such that they do not have semantic safety in terms of abstractions that protect you from the underlying computational substrate
2026-04-03 02:39:18 +0000 <davean> geekosaur: About unsafe stuff
2026-04-03 02:38:46 +0000 <geekosaur> since neither compiler can give you static guarantees, it's on you as programmer using unsafe features to prove that you are actually using them safely
2026-04-03 02:38:40 +0000 <mesaoptimizer> no I get that
2026-04-03 02:37:48 +0000 <geekosaur> this is no different from https://doc.rust-lang.org/book/ch20-01-unsafe-rust.html
2026-04-03 02:37:00 +0000 <mesaoptimizer> reminds me of LLM-generated text, which has triggered similar confusion and dicussion between I and a few of my acquaintances in the past, related to type theory and semantics. I guess its best I abandon reading this essay.
2026-04-03 02:36:39 +0000abbies(~abbies@tilde.guru) (Client Quit)
2026-04-03 02:35:44 +0000 <mesaoptimizer> geekosaur: yeah that is now my current interpretation of it. it does seem quite confusing though
2026-04-03 02:35:05 +0000 <mesaoptimizer> If you say something is pure, I expect that the language *guarantees* the semantics of it being pure, with the exception of the expression using, say, `unsafePerformIO`. Ergo, you cannot call those functions as those where 'the boundary cannot be violated' -- you already broke the semantic safety guarantees!