Newest at the top
| 2026-01-08 02:19:19 +0100 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 240 seconds) |
| 2026-01-08 02:18:36 +0100 | <EvanR> | in any language |
| 2026-01-08 02:17:55 +0100 | <EvanR> | haskell itself isn't making code inherently safe, you have whatever layers of stuff and planning for something like an eval bot |
| 2026-01-08 02:15:52 +0100 | <int-e> | EvanR: I'm kind of curious what evoked that picture. |
| 2026-01-08 02:14:36 +0100 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-01-08 02:13:43 +0100 | <newmind> | agents are doing |
| 2026-01-08 02:13:42 +0100 | <newmind> | obviously, yes. and i'd never advocate for blindly running untrusted code either. but it's a lot easier to reason about a function with a type signature than just executing a bash script blindly. it's not meant as a airtight sandbox that holds up against adverserial attacks, but it is another layer. and it is quite a bit more than what current |
| 2026-01-08 02:12:27 +0100 | <int-e> | huh |
| 2026-01-08 02:11:34 +0100 | <EvanR> | he's running in the equivalent of that underwater prison in avengers |
| 2026-01-08 02:11:16 +0100 | <EvanR> | the amount of infrastructure required for lambdabot xD |
| 2026-01-08 02:10:35 +0100 | <lambdabot> | why oh why indeed |
| 2026-01-08 02:10:34 +0100 | <int-e> | > text "why oh why indeed" |
| 2026-01-08 02:10:21 +0100 | <EvanR> | and not sure why that's even required |
| 2026-01-08 02:10:13 +0100 | <EvanR> | executing "untrusted" code is still a horrible idea in haskell |
| 2026-01-08 02:10:05 +0100 | <int-e> | You'd need {-# LANGUAGE AdditivePromises #-} for that. |
| 2026-01-08 02:09:40 +0100 | <EvanR> | some of those premises about haskell don't add up ... |
| 2026-01-08 02:08:55 +0100 | <int-e> | discover new ways in which well-typed programs go wrong |
| 2026-01-08 02:08:00 +0100 | prite | (~pritam@user/pritambaral) (Quit: Konversation terminated!) |
| 2026-01-08 02:06:28 +0100 | <newmind> | take your point, i was just generally looking for feedback, viewpoints and ideas |
| 2026-01-08 02:06:27 +0100 | <newmind> | the connection would be that haskell (or strongly typed languages in particular) provide guardrails and constraints to LLM generated code that does not exist in other languages, including that generated haskell code is safer (not absolutely safe) to generate and run than in most other languages (especially with the Safe extension enabled). but i |
| 2026-01-08 02:05:09 +0100 | omidmash5 | omidmash |
| 2026-01-08 02:05:09 +0100 | omidmash | (~omidmash@user/omidmash) (Ping timeout: 244 seconds) |
| 2026-01-08 02:05:09 +0100 | xff0x | (~xff0x@2405:6580:b080:900:4b0b:90a:cd82:2bd2) (Ping timeout: 244 seconds) |
| 2026-01-08 02:03:41 +0100 | <haskellbridge> | <sm> (I'm not excusing rudeness, but I understand it) |
| 2026-01-08 02:03:39 +0100 | <haskellbridge> | <sm> it means that if you come into a space and talk about the ai thing you're building, you need to be ready for all kinds of response |
| 2026-01-08 02:03:35 +0100 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 240 seconds) |
| 2026-01-08 02:03:10 +0100 | omidmash5 | (~omidmash@user/omidmash) omidmash |
| 2026-01-08 02:02:58 +0100 | <jreicher> | Fair |
| 2026-01-08 02:02:08 +0100 | <EvanR> | yeah I don't see the connection to haskell specifically |
| 2026-01-08 02:01:30 +0100 | <ncf> | take it to #haskell-offtopic please, this is not the place to discuss LLMs |
| 2026-01-08 02:01:26 +0100 | <geekosaur> | that was the theory behind UML, I think? |
| 2026-01-08 02:00:56 +0100 | <jreicher> | Never thought about this before, but maybe there's a market for AI /only/ because of informal specification. If the spec was formal non-AI program generation might be possible? |
| 2026-01-08 02:00:55 +0100 | <EvanR> | hopefully that bleeds onto programming at some point |
| 2026-01-08 02:00:03 +0100 | <EvanR> | people with mathematical or engineering maturity are better at specifying what they want or what they have |
| 2026-01-08 01:59:30 +0100 | <newmind> | in the meantime, for _a lot_ of real world software problems, it's a lot more grey: people can specify "what they want" only in very loose terms |
| 2026-01-08 01:59:15 +0100 | <EvanR> | that part is beyond my paygrade xD |
| 2026-01-08 01:58:49 +0100 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-01-08 01:58:48 +0100 | <jreicher> | Even if AI could do it, I think the bigger problem is motivating humans to write specs in the first place. That's been possible for half a century and still almost nobody doesit. |
| 2026-01-08 01:58:02 +0100 | <newmind> | least run code that's checked by a compiler and can't do anything completely nuts |
| 2026-01-08 01:58:01 +0100 | <newmind> | that would be the dream, yeah: provide a spec, and it spits out a program that fulfills that (and while we're at it, also does proofably terminate and run in limited space).. but current AI agents are doing the exact opposite, little more than running in yolo mode and just executing whatever comes to mind.. what i'm proposing is a middle ground: at |
| 2026-01-08 01:57:36 +0100 | <EvanR> | not "write me an MMO in haskell" |
| 2026-01-08 01:57:22 +0100 | <EvanR> | jreicher, correct in the sense that I explicitly specified the theorem |
| 2026-01-08 01:56:39 +0100 | ryanbooker | (uid4340@id-4340.hampstead.irccloud.com) ryanbooker |
| 2026-01-08 01:56:38 +0100 | <EvanR> | of code that is hardly a proof of anything |
| 2026-01-08 01:56:35 +0100 | <jreicher> | Even correct programs aren't proofs of theorems. "In retrospect it seems to be doing the right thing." |
| 2026-01-08 01:55:48 +0100 | <EvanR> | it's more like mass copywrite infringement search |
| 2026-01-08 01:55:04 +0100 | <EvanR> | in the sense that a program is a proof of some theorem, if I can get a computer to give me the correct program that's great. But that's not what we're doing right now with AI |
| 2026-01-08 01:52:16 +0100 | L29Ah | (~L29Ah@wikipedia/L29Ah) L29Ah |
| 2026-01-08 01:51:15 +0100 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 240 seconds) |
| 2026-01-08 01:49:16 +0100 | raincomplex | (~rain@user/raincomplex) raincomplex |