Newest at the top
| 2026-06-30 21:23:36 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 252 seconds) |
| 2026-06-30 21:18:59 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-06-30 21:08:55 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 276 seconds) |
| 2026-06-30 21:05:16 +0000 | Katarushisu6 | (~Katarushi@finc-20-b2-v4wan-169598-cust1799.vm7.cable.virginm.net) |
| 2026-06-30 21:03:38 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-06-30 21:03:26 +0000 | <int-e> | ('p' for 'd' that is) |
| 2026-06-30 21:02:44 +0000 | juri_ | (~juri@217-114-215-140.pool.ovpn.com) juri_ |
| 2026-06-30 21:02:41 +0000 | <int-e> | (unfortunately that's not a plausible typo. :) ) |
| 2026-06-30 21:02:12 +0000 | <int-e> | Well, it's not calle vibe coping for nothing |
| 2026-06-30 21:01:31 +0000 | <tomsmeding> | yes |
| 2026-06-30 21:01:15 +0000 | <int-e> | So... it feels better by an unspecified margin. |
| 2026-06-30 21:01:09 +0000 | polykernel_ | polykernel |
| 2026-06-30 21:01:09 +0000 | polykernel | (~polykerne@user/polykernel) (Ping timeout: 245 seconds) |
| 2026-06-30 21:00:13 +0000 | <tomsmeding> | subjectively improves output quality |
| 2026-06-30 20:59:54 +0000 | <int-e> | I dislike the word "works" in this context, it's too unspecific |
| 2026-06-30 20:59:37 +0000 | polykernel_ | (~polykerne@user/polykernel) polykernel |
| 2026-06-30 20:59:06 +0000 | Katarushisu6 | (~Katarushi@finc-20-b2-v4wan-169598-cust1799.vm7.cable.virginm.net) (Quit: The Lounge - https://thelounge.chat) |
| 2026-06-30 20:57:26 +0000 | <tomsmeding> | which, perhaps surprisingly, apparently works |
| 2026-06-30 20:57:14 +0000 | <tomsmeding> | int-e: supposedly, the fancy ones also do some kind of socratic method stuff: have another instance of the LLM (with a different prompt) judge the output, feeding that back, etc |
| 2026-06-30 20:56:51 +0000 | int-e | shrugs |
| 2026-06-30 20:56:29 +0000 | <int-e> | So the poor humans have to do less of that. |
| 2026-06-30 20:56:14 +0000 | <int-e> | Anyway. The point was that as I understand it, the harness is a kind of lector. It'll catch the most obvious bullshit (code that fails to compile, test suites don't pass, linter outputs etc.) and reject stuff, and I suppose add it to the prompt (call it "refining the prompt" if you want to be fancy) |
| 2026-06-30 20:55:50 +0000 | <tomsmeding> | lol |
| 2026-06-30 20:55:43 +0000 | <Clint> | depends how radioactive it is |
| 2026-06-30 20:55:16 +0000 | <EvanR> | how many infinities would it take a banana to write Shakespeare |
| 2026-06-30 20:54:58 +0000 | <tomsmeding> | don't make fun of bananas, bananas are nice |
| 2026-06-30 20:54:26 +0000 | <EvanR> | a human and a banana |
| 2026-06-30 20:53:20 +0000 | <tomsmeding> | fair |
| 2026-06-30 20:53:07 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 264 seconds) |
| 2026-06-30 20:53:04 +0000 | <int-e> | tomsmeding: Or maybe he just called a human a monkey. Which... eh we're not that far apart, biologically speaking. |
| 2026-06-30 20:52:33 +0000 | <monochrom> | heh |
| 2026-06-30 20:52:28 +0000 | juri_ | (~juri@217-114-215-140.pool.ovpn.com) (Ping timeout: 268 seconds) |
| 2026-06-30 20:52:10 +0000 | <tomsmeding> | That's certainly too harsh on Dan Brown :p |
| 2026-06-30 20:51:34 +0000 | <int-e> | monochrom: Though his punchline (as I recall it) was very different: "And a single monkey typing away over Columbus Day weekend will give you a Dan Brown." |
| 2026-06-30 20:51:06 +0000 | acidjnk | (~acidjnk@p200300d6e74def37a5af64098d23bd41.dip0.t-ipconnect.de) acidjnk |
| 2026-06-30 20:50:27 +0000 | acidjnk_new | (~acidjnk@p200300d6e74def67cefedba3bc5cab81.dip0.t-ipconnect.de) (Ping timeout: 272 seconds) |
| 2026-06-30 20:50:27 +0000 | acidjnk | (~acidjnk@p200300d6e74def67cefedba3bc5cab81.dip0.t-ipconnect.de) (Ping timeout: 272 seconds) |
| 2026-06-30 20:50:12 +0000 | <int-e> | monochrom: I'm stealing an old Colbert bit |
| 2026-06-30 20:49:45 +0000 | <tomsmeding> | what monochrom said |
| 2026-06-30 20:49:40 +0000 | <int-e> | (I'm not covering prompts part.) |
| 2026-06-30 20:49:39 +0000 | <monochrom> | That may be too harsh on Dan Brown. :) |
| 2026-06-30 20:49:07 +0000 | <int-e> | It's a million monkeys with a dictionary (LLM) on typewriters, but with a lector (harness) that rejects the most obvious bullshit. You may get a Dan Brown but not Shakespeare? (too harsh?) |
| 2026-06-30 20:49:04 +0000 | <tomsmeding> | (theoretically, in any case; not by personal experience) |
| 2026-06-30 20:48:16 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-06-30 20:48:00 +0000 | <tomsmeding> | monochrom: having an LLM generate proofs is actually the cleanest application of them that I've found yet |
| 2026-06-30 20:46:53 +0000 | acidjnk_new3 | (~acidjnk@p200300d6e74def37a5af64098d23bd41.dip0.t-ipconnect.de) acidjnk |
| 2026-06-30 20:46:42 +0000 | <monochrom> | or at least test with test cases. |
| 2026-06-30 20:46:30 +0000 | <tomsmeding> | well, the quotes are because the things are hundreds of gigabytes large these days; the word "concise" is because naively, a probability distribution of the next token conditioned on the previous 100k tokens would require super-galactic storage |
| 2026-06-30 20:46:21 +0000 | <monochrom> | It's also how I do math! I think up an attempt by experience of seeing tons of previous proofs by others or me (LLM), but I still have to check it with very rigid rules (proof checker). |
| 2026-06-30 20:45:40 +0000 | <tomsmeding> | oh |