2026/06/30

Newest at the top

2026-06-30 20:25:06 +0000ridcully(~ridcully@p57b52a2d.dip0.t-ipconnect.de) (Quit: WeeChat 4.9.2)
2026-06-30 20:25:05 +0000 <monochrom> I haven't needed to use LLMs.
2026-06-30 20:24:58 +0000 <monochrom> OK!
2026-06-30 20:24:50 +0000 <EvanR> see
2026-06-30 20:24:44 +0000 <EvanR> so they don't have to
2026-06-30 20:24:40 +0000 <EvanR> it's like when elite golfers hire robots to play golf for them
2026-06-30 20:24:32 +0000 <int-e> monochrom: I see it: Consider who the people who unironically call themselves "elite" are.
2026-06-30 20:24:19 +0000 <dcb> I'd try to minimize the number of tools needed to write code
2026-06-30 20:24:05 +0000Lord_of_Life(~Lord@user/lord-of-life/x-2819915) Lord_of_Life
2026-06-30 20:23:56 +0000 <monochrom> I don't see the correlation between elite and vibe coding.
2026-06-30 20:23:12 +0000 <probie> My hatred for LLMs doesn't come from their capacity, but the fact that they were trained on copyrighted works without permission
2026-06-30 20:22:22 +0000 <EvanR> gotta stir the pot!
2026-06-30 20:21:55 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 264 seconds)
2026-06-30 20:20:39 +0000 <tomsmeding> I grudgingly use LLMs for prototypes when on-the-clock
2026-06-30 20:20:16 +0000 <absentia> i can't really believe all the LLM hype if it is plausibly generated and hallucinated by LLMs themselves
2026-06-30 20:19:54 +0000 <absentia> or do you view this as a foolish and astroturfed bubble
2026-06-30 20:19:43 +0000 <absentia> are all of you vibecoding now
2026-06-30 20:19:39 +0000 <absentia> so this is a pretty elite channel
2026-06-30 20:19:26 +0000dtman34(~dtman34@c-73-242-68-179.hsd1.mn.comcast.net) (Ping timeout: 265 seconds)
2026-06-30 20:18:59 +0000dtman34_(~dtman34@c-73-242-68-179.hsd1.mn.comcast.net) dtman34
2026-06-30 20:16:50 +0000polykernel_polykernel
2026-06-30 20:16:50 +0000polykernel(~polykerne@user/polykernel) (Ping timeout: 245 seconds)
2026-06-30 20:14:49 +0000merijn(~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn
2026-06-30 20:13:53 +0000polykernel_(~polykerne@user/polykernel) polykernel
2026-06-30 20:13:18 +0000 <tomsmeding> (there's still plenty of engineering involved to make it good)
2026-06-30 20:13:02 +0000 <tomsmeding> it's the recognition of vectorisable structure from imperative loop code that's such a nightmare
2026-06-30 20:12:46 +0000 <tomsmeding> well once you have a high-level _array language_, at that level of abstraction it's much more doable, conceptually, to generate good code
2026-06-30 20:12:17 +0000 <monochrom> (Use rewrite rules for the code optimization? "It's not that hard"? :) )
2026-06-30 20:11:36 +0000 <monochrom> He didn't generate optimized code, but he spec'ed out the EDSL itself.
2026-06-30 20:11:22 +0000 <tomsmeding> (for GPU, naturally "vectorisation" is a thing already because if not then what are you doing)
2026-06-30 20:11:05 +0000 <monochrom> I didn't understand APL until that paper!
2026-06-30 20:10:45 +0000 <tomsmeding> currently, yes, but just this evening Ivo messaged me that a student would be looking at doing some vectorisation in accelerate already, because LLVM isn't doing a very good job with the generated code for reductions and especially scans and permutes
2026-06-30 20:10:45 +0000 <monochrom> Ooohhh Jeremy Gibbons did APL as EDSL into Haskell: https://www.cs.ox.ac.uk/publications/publication10857-abstract.html
2026-06-30 20:10:08 +0000 <jaror> Does accelerate rely on llvm for vectorisation?
2026-06-30 20:10:05 +0000 <tomsmeding> (surprisingly, to me)
2026-06-30 20:09:34 +0000 <tomsmeding> not massiv, by the way, because iirc that doesn't vectorise
2026-06-30 20:09:16 +0000 <monochrom> matlab
2026-06-30 20:09:14 +0000 <int-e> monochrom: fair, I was thinking EDSL
2026-06-30 20:09:11 +0000 <tomsmeding> accelerate, futhark
2026-06-30 20:09:11 +0000 <jaror> Futhark
2026-06-30 20:08:56 +0000 <tomsmeding> repa
2026-06-30 20:08:53 +0000 <monochrom> APL
2026-06-30 20:08:43 +0000 <int-e> jaror: welcome to.... numpy
2026-06-30 20:08:31 +0000 <monochrom> Oh, that.
2026-06-30 20:08:28 +0000 <tomsmeding> you're kidding, that would never work
2026-06-30 20:08:20 +0000 <jaror> :P
2026-06-30 20:08:18 +0000 <tomsmeding> O.o
2026-06-30 20:08:11 +0000 <jaror> or, here's a wild idea: a DSL which is specialized to array operations.
2026-06-30 20:08:04 +0000 <tomsmeding> that's not at all obvious any more :)
2026-06-30 20:07:50 +0000 <tomsmeding> monochrom: okay, the relatively easy cases (as compared to loop vectorisation as understood by gcc/llvm) are not too hard (as compared to other optimisations GHC does)