Post B8fWiCKZEjWwZjdGvw by deutrino@mstdn.io
(DIR) More posts by deutrino@mstdn.io
(DIR) Post #B8ej7pcjHvKzyJ9oIq by ariadne@social.treehouse.systems
0 likes, 0 repeats
@juliank we do? not sure I follow
(DIR) Post #B8ejGmXG8hjX03rmKW by tris@chaos.social
0 likes, 0 repeats
@ariadne @juliank Context is Debian GR to ban LLM has been proposed: https://lists.debian.org/debian-vote/2026/07/msg00000.html
(DIR) Post #B8ejSib4bmTmGBLH0q by ariadne@social.treehouse.systems
0 likes, 0 repeats
@juliank @tris ok, but as a free software maintainer I have really yet to identify any case where I would want to use an LLM for a task that I didn't already have more suitable workflows for.
(DIR) Post #B8ejvkHpPCc5tkBJuy by zygoon@fosstodon.org
0 likes, 0 repeats
@ariadne @juliank @tris As a maintainer I can attest that LLMs drastically improve my workflow and work life balance. I have yet to find and area where they didn't help me greatly. I would love to exchange notes.Edit: this includes local models running on my laptop and my nvidia spark, as well as open and proprietary models operated by 3rd parties.
(DIR) Post #B8ejyL5fXZYCcejNom by juliank@mastodon.social
0 likes, 0 repeats
@ariadne @tris I can give you two examples:In a single reviewer project like APT I can either trust the test suite and myself, or I can add a vibe plausibility check.Also I spent a lot of my time either in or recovering from meetings, so having the AI produce code in the background and then have me review that is also helpful.Not all of the stuff is trivial enough for that though, but there's lots of low hanging fruits that the AI can reasonably work on.
(DIR) Post #B8ekQLatGI1eICgpV2 by ariadne@social.treehouse.systems
0 likes, 1 repeats
@juliank @tris like I have done a fairly exhaustive investigation of both frontier and open weight models, including having gpt-5.6-sol rebuild the HURD from scratch, which took ~5 billion tokens (good thing I paid for the subscription, huh?)and I have yet to find a use case where I can say with confidence that it significantly enhances my productivity.I have to experiment with these things at work, I frequently find myself arguing with Claude because it "hallucinates" shit and runs with it. and since I work at a cybersecurity lab, half the time it "hallucinates" a safety refusal.overall, it ain't there, and the ecological and economical cost to get it there is absurdly high.
(DIR) Post #B8ekcBKdF59NwKjD3w by ariadne@social.treehouse.systems
0 likes, 1 repeats
@juliank @tris @zygoon not interested. it isn't a lack of knowledge but a difference in philosophy. I solve my work-life balance by simply blocking out time on my calendar for my family.
(DIR) Post #B8eknGXUDodVH91fyi by ariadne@social.treehouse.systems
0 likes, 0 repeats
@tris @juliank using LLMs as a way of mitigating burnout will only lead to even harder burnout later.
(DIR) Post #B8em2wOSzR0JNnqTY0 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@juliank @tris @zygoon for me I like to feel connected to whatever I am doing. the deeper, the better.I don't like LLMs for coding for the same reason I don't like other tools that get between me and the code. I hold entire state transitions in my head, and am fully capable of exploring them in my head. and beyond this: I can't trust the accuracy of the LLM, so it is not useful to me for analysis tasks.sometimes I will use an LLM to analyze a log, but I rate it 50/50 on deriving useful insights.
(DIR) Post #B8em88JI70BtmzttBY by zygoon@fosstodon.org
0 likes, 0 repeats
@ariadne @juliank @tris thank you for sharing. I think I see your point only view.
(DIR) Post #B8enHfVMAG2dIieeDw by juliank@mastodon.social
0 likes, 0 repeats
@ariadne @tris I'd argue rewriting the Hurd from scratch is also not something you'd normally do yourself either and we should not hold LLMs to super human standards :D
(DIR) Post #B8fPgsHrZcNiRUn0ZE by ariadne@social.treehouse.systems
0 likes, 1 repeats
@tris @juliank I hold it to the capabilities the lab advertises. if you say it is good at systems reasoning, design and cyber, then it should not be hard to generate a hurd clone with it.in the US, this is a second or third year capstone project for a CS student specializing in systems engineering, not really a superhuman problem. and indeed, it took a stack of explorations and generated a hurd clone. for transparency, you can see the result on my github.but that hurd clone is also absolute slop. the design is so-so, but the code is unreviewable. slop that had I paid for the tokens would have cost around $200k to generate, thankfully I was smart and paid for the subscription instead.
(DIR) Post #B8fQ7TJIXqKunT8wF6 by ariadne@social.treehouse.systems
0 likes, 0 repeats
@tris @juliank @dysfun im just going to have to pop an edible and force myself to review it
(DIR) Post #B8fQu6dlcTAag2uMEa by ariadne@social.treehouse.systems
0 likes, 1 repeats
@tris @juliank @dysfun if a junior handed me this I would know who I was putting on a pivot this quarter 😂
(DIR) Post #B8fRE3HXodOi3L1Zpo by ariadne@social.treehouse.systems
0 likes, 1 repeats
@juliank @tris @evana ok, cool, so you used it as an ADHD forcing function.but I don't have ADHD.
(DIR) Post #B8fSGEeqajI1rrWs52 by deutrino@mstdn.io
0 likes, 0 repeats
@ariadne @juliank @tris @evana this is what I want out of an AI coding assistant as someone with AuDHD that's heavy on the "wait... what was I doing?" and who pays a huge penalty for context switches (which, ironically, so does genAI in many cases)my ideal coding assistant would be terse, knowledgeable, have a good view into whatever codebase I'm working with, and be able to help me stay on track.anyway just sayin'.
(DIR) Post #B8fSVLTDXrjhL8IJXc by ariadne@social.treehouse.systems
0 likes, 0 repeats
@juliank @tris @evana @deutrino yes, that is basically the only use case I've found for it, to be able to ask questions about patches, etc.
(DIR) Post #B8fTRIuPE7NkBveGFk by dngrs@chaos.social
0 likes, 0 repeats
@ariadne @tris @juliank alternatively it offloads the burnout onto others 🥳maybe the real use case all along, huh
(DIR) Post #B8fUBYAUpAm2q4E5s8 by juliank@mastodon.social
0 likes, 0 repeats
@ariadne @tris @evana @deutrino I like "here's a failing test case, make it work".Or "my test coverage is low. Hallucinate test cases to improve it"Or "add ACLS contracts to functionally verify this C function"But without a hard testable proof to work towards it gets led astray by nonsense
(DIR) Post #B8fUNuen2Q9M3qI6RE by ariadne@social.treehouse.systems
0 likes, 0 repeats
@tris @evana @deutrino @juliank I do not, because those are largely the problems I enjoy solving.
(DIR) Post #B8fUlpU5v9lcy7h09w by ariadne@social.treehouse.systems
0 likes, 0 repeats
@daemonspudguy @juliank @evana @deutrino @tris I don't think it is that. I do think a lot of folks are using LLMs to mitigate burnout and lying to themselves that the mitigation is a productivity gain though.
(DIR) Post #B8fVK1oGbNZP2bR2vo by juliank@mastodon.social
0 likes, 0 repeats
@ariadne @daemonspudguy @evana @deutrino @tris I mean that's certainly the case.But also wrt enjoyment I wonder what that even means? Would I rather be out cycling or spend the whole night coding after having spent the whole day working on the same stuff already?I don't think AI helps a lot because it has its own addictive properties where you watch it "think" and then interrupt it each time it goes astray.
(DIR) Post #B8fWiCKZEjWwZjdGvw by deutrino@mstdn.io
0 likes, 0 repeats
@ariadne @daemonspudguy @juliank @evana @tris interesting point on "mitigate burnout," because when I was burning out, heavily engaging with pair programming bought me another 18 months, and an ideal AI coding assistant (for me) would basically be a pair programming partner.
(DIR) Post #B8nBqW6upDmj50Czia by alois@functional.cafe
0 likes, 0 repeats
@ariadne sounds like you did not use them in a productive manner.I think the bigger question is not about if they should be used or not, but how they should be used.Just my 2 cents, I do appreciate your point of view... but wasting so many tokens without releasing early that the quality was bad is highly questionable.