Newest at the top
| 2026-06-30 20:44:27 +0000 | <tomsmeding> | it's a "concisely" represented markov chain with a very long history dependence |
| 2026-06-30 20:44:11 +0000 | <newmind> | following and detecting patterns... just like humans :P |
| 2026-06-30 20:43:40 +0000 | <tomsmeding> | it's a language model, and if you understand what it does, it's not so surprising what it's good at and what it isn't good at |
| 2026-06-30 20:43:16 +0000 | <tomsmeding> | I hate the word "AI" for LLMs |
| 2026-06-30 20:43:02 +0000 | <tomsmeding> | (and I'm not even talking about the inevitable skill atrophy if you replace too much of your work) |
| 2026-06-30 20:42:50 +0000 | <monochrom> | It's why I don't call it AI when talking to people who know. I'm sure LLM is a great component of AI, but just one component, it's lacking f**king fact-checking. |
| 2026-06-30 20:41:53 +0000 | <tomsmeding> | the more falsifiable the output, the less risky it is. :p |
| 2026-06-30 20:40:44 +0000 | <tomsmeding> | mauke: they do offer that, and the success rate (of good models) is high enough that it's tempting to use it. But it's very dangerous, because it's often very hard to check that they're right (indeed, otherwise you would likely not have asked the LLM) |
| 2026-06-30 20:39:13 +0000 | Katarushisu6 | (~Katarushi@finc-20-b2-v4wan-169598-cust1799.vm7.cable.virginm.net) |
| 2026-06-30 20:39:04 +0000 | machinedgod | (~machinedg@d108-173-95-19.abhsia.telus.net) (Ping timeout: 245 seconds) |
| 2026-06-30 20:38:10 +0000 | polykernel_ | polykernel |
| 2026-06-30 20:38:10 +0000 | polykernel | (~polykerne@user/polykernel) (Ping timeout: 248 seconds) |
| 2026-06-30 20:37:47 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) (Ping timeout: 272 seconds) |
| 2026-06-30 20:36:34 +0000 | <tomsmeding> | (I'm still unsure about the latter, but don't tell them that) |
| 2026-06-30 20:36:18 +0000 | <tomsmeding> | unstated: why they were doing so at an academic institution |
| 2026-06-30 20:36:13 +0000 | polykernel_ | (~polykerne@user/polykernel) polykernel |
| 2026-06-30 20:35:50 +0000 | <tomsmeding> | monochrom: actually, the first time I heard someone describe what they were doing (which turned out to be harness engineering), I made the mistake of asking (because I genuinely didn't understand at first) how it was different from prompt engineering |
| 2026-06-30 20:35:44 +0000 | hc | (~hc@user/hc) hc |
| 2026-06-30 20:34:43 +0000 | hc | (~hc@user/hc) (Remote host closed the connection) |
| 2026-06-30 20:34:37 +0000 | <monochrom> | or rather s/meaningless/unguessable/ |
| 2026-06-30 20:34:22 +0000 | <jaror> | Yeah, I was talking to Ed Kmett at Zurihac and he is spending like 3-4 times my salary on tokens... |
| 2026-06-30 20:34:14 +0000 | <tomsmeding> | (that's why I offered a translation -- I didn't know its definition until a few weeks ago :p) |
| 2026-06-30 20:34:04 +0000 | <monochrom> | after bus factors and test pollution. |
| 2026-06-30 20:33:54 +0000 | <monochrom> | Oh harness engineering is going to be another of those meaningless "meaningful" words coined by programmers. |
| 2026-06-30 20:33:40 +0000 | lisbeths | (uid135845@id-135845.lymington.irccloud.com) (Quit: Connection closed for inactivity) |
| 2026-06-30 20:33:05 +0000 | <tomsmeding> | mind, compared to even the middle-of-the-road programmer these days, my experience using these tools amounts to a rounding error |
| 2026-06-30 20:33:03 +0000 | Lord_of_Life | (~Lord@user/lord-of-life/x-2819915) (Excess Flood) |
| 2026-06-30 20:33:01 +0000 | califax | (~califax@user/califx) califx |
| 2026-06-30 20:32:53 +0000 | merijn | (~merijn@host-cl.cgnat-g.v4.dfn.nl) merijn |
| 2026-06-30 20:32:50 +0000 | <monochrom> | I admit that it does something right, OK? For example, academically it shows that some linguists are right about semantics of a word containing a large factor of correlating with other words, much larger than most people want to believe. |
| 2026-06-30 20:32:24 +0000 | <tomsmeding> | (the precise interpretation of that latter statement is personal, so I'll leave that to the reader) |
| 2026-06-30 20:32:11 +0000 | <tomsmeding> | but if you do use them, you have to make sure you know _why_ you're using them, and if you actually think that's a good idea |
| 2026-06-30 20:31:31 +0000 | califax | (~califax@user/califx) (Remote host closed the connection) |
| 2026-06-30 20:31:07 +0000 | <tomsmeding> | and it's quite impressive how "harness engineering" (translation: writing code that practices the socratic method on the LLM, in addition to providing it a (textual) "API" for calling external tools) can increase their abilities |
| 2026-06-30 20:29:51 +0000 | <tomsmeding> | if it's not too conceptually difficult, I'm not surprised |
| 2026-06-30 20:29:38 +0000 | Katarushisu6 | (~Katarushi@finc-20-b2-v4wan-169598-cust1799.vm7.cable.virginm.net) (Ping timeout: 248 seconds) |
| 2026-06-30 20:29:25 +0000 | <jaror> | Some GHC developers seem to just be able to say "fix this bug" and have Claude churn away at it for them... |
| 2026-06-30 20:29:23 +0000 | <monochrom> | It's advertised as artificial intelligence. |
| 2026-06-30 20:28:49 +0000 | Lord_of_Life | (~Lord@user/lord-of-life/x-2819915) Lord_of_Life |
| 2026-06-30 20:28:48 +0000 | <monochrom> | LLMs will make great politicians and management types. Just needing to mince words. |
| 2026-06-30 20:28:35 +0000 | <tomsmeding> | if you don't write any boilerplate, I'm not surprised you don't find them useful :) |
| 2026-06-30 20:28:33 +0000 | <mauke> | I don't use LLMs, but how are they advertised nowadays? is it just code generation, or do they also offer to explain existing code to you? |
| 2026-06-30 20:27:52 +0000 | <monochrom> | This is what you get when you believe in "express in your own words to show you understand". |
| 2026-06-30 20:27:42 +0000 | Lord_of_Life | (~Lord@user/lord-of-life/x-2819915) (Excess Flood) |
| 2026-06-30 20:27:37 +0000 | <tomsmeding> | monochrom: they generate predictable text, that's all they do. If you ask them to generate boilerplate for you, they'll be mostly successful. The "higher-quality" the model, the more the boundaries of "boilerplate" shift, but fundamentally that's still what's happening |
| 2026-06-30 20:27:35 +0000 | <monochrom> | In one case it said "X is true" where the reference it cited said "X is false". |
| 2026-06-30 20:27:03 +0000 | <jaror> | it always reminds me of the "Great Galactic Grid" from the Zogg from Betelgeuse video on mathematics |
| 2026-06-30 20:26:34 +0000 | ridcully | (~ridcully@p57b52a2d.dip0.t-ipconnect.de) ridcully |
| 2026-06-30 20:26:12 +0000 | <probie> | [Google AI] did you know that the concept of "Supplemental Material" simply didn't exist until digital storage? |
| 2026-06-30 20:25:43 +0000 | <monochrom> | I wanted to give it a chance! Disillusioned. |