04:43 airlied: mlankhorst: yes
07:27 MrCooper: daniels: my counterexample would be something like https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/5300#note_2865132 , where the "author" is categorically refusing to directly apply feedback to the commit logs, it's LLM-generated or bust
07:28 MrCooper: not sure it's possible to draw a line for that other than "no LLM-generated commit logs"
07:33 pq: zmike, when you had an LLM write commit messages, has it correctly explained *why* *you* wanted that change? I'm curious, since that's the most important part of a commit message.
07:35 pq: zmike, karolherbst, have you had an LLM review the AGENTS.md? I very much recommend doing that.
08:58 daniels: MrCooper: shrug, again the issue to me seems that the commit logs are unreadable, and that banning a tool doesn’t necessarily solve that problem - if someone doesn’t want to co-operate in review, then what’s even to say that they’ll obey the rules you set to work around their disregard for your opinion and demands?
09:41 MrCooper: that'll be obvious, it would be easier to react by pointing to the policy
09:43 MrCooper: (it's interesting how many people argue against this kind of policy because it can't 100% prevent violations; that applies to any rule / law, doesn't mean they serve no purpose)
10:41 karolherbst: policy is to not copy paste AI/ML output.. so, not sure why we even have that discussion on why it's then fine actually to let the tools do it anyway. If we want to change the Agents.md file so that those tools are permitted to copy&paste their output, we should have a policy discussion first
10:48 karolherbst: and I'd prefer if people here wouldn't just throw community consensus into the bin just because you don't like it and try to change things around "to make it sane" or something. Policy is no generated text in comments/MR/whatever, so.. closing the MR "because it's against current policy" sounds sensible, no?
10:55 daniels: karolherbst: you know that's a link to a GNOME MR, right?
10:55 karolherbst: isn't it in the context of the mesa MR we had?
10:55 karolherbst: like bringing it up here
10:57 daniels: sure, and I thought that we were having a discussion about what the community consensus actually was, and giving my opinion in the context of that. if it's not permitted to express an opinion on what the policy should be, then a) it's not a discussion, and b) maybe 'consensus' isn't actually what it claims to be but instead the opinion which is being imposed on the Mesa community as the only acceptable view
10:58 karolherbst: community consensus was to not have AI generated comments/MR descriptionrs/etc... but the Agents.md MR is not a change in policy, it's a change in instructions to those tools, and the proposed change is in violation of our own policy. I'm happy to have a conversation of the policy we have, but then please open an MR to change that, not the
10:58 karolherbst: Agents.md file
10:59 karolherbst: if somebody things we didn't discuss that policy change enough, then we can reopen that discussion
10:59 daniels: I didn't open the MR, nor did I discuss anything about that MR with zmike before he opened it
10:59 daniels: I'm just someone expressing an opinion about things on IRC, which it seems is not acceptable
11:00 karolherbst: yeah, and that is fine. And I said, that I think changing it would lead to a worse outcome, which I'm getting more and more convinced of
11:01 daniels: that's the opinion held by you, which is one developer. other people might have other opinions. if the standing rule you're suggesting is that we never discuss anything other than through specific targeted MRs to policy documents, then that's ... a thing ... but it's not a rule we've ever held before. and a billion prior conversations here would have broken that rule.
11:02 karolherbst: I can only repeat: in our policy there is text saying we don't want AI generated text in comments, MR descriptions, etc... if you want to change that, I'm happy to have that conversation
11:03 daniels: I do not have the energy for this conversation, but it's completely unfair to jump on me for 'throwing community consensus into the bin' for expressing my own personal opinion, in the context of a live discussion where a bunch of people were expressing their own opinions
11:03 karolherbst: and it doesn't have to be in an MR either. But we do have this discussion because of a proposed change, and the conversation is within the context of that
11:04 karolherbst: okay, maybe I should clarify, I didn't mean you specifically, and not in the sense of "let's never reevaluate this". I said this, because changing the Agents.md file in a way it contradicts community consensus isn't something we should be doing
11:05 karolherbst: again, as I said: happy to have a conversation on changing that consensus, but then we should do that and not continue discussing why that Agents.md change is fine actually
11:55 alyssa: it's unclear to me what the community consensus actually is, see also: lol governance.
11:56 alyssa: I mean I know what's written in policy but idk if that's anyone's preferred policy or just the current state of an evolving compromise
11:56 alyssa: (which might be as close to consensus as we'll get, but)
12:02 karolherbst: I think we really need a more structured way of going on about this, but yeah... I do agree with "consensus" being a problematic term here actually, but I think it's kinda unavoidable with that topic. Loosing nuance on the way, or refusing actual change in opinion isn't helpful here anyway. I think there are compromises possible here, but that will
12:02 karolherbst: require us to.. not be like that
12:07 zmike: I had thought that opening a MR against AGENTS.md would be the method to start a community discussion
12:07 zmike: since MRs are usually how we discuss things
12:07 zmike: if AI-related topics require a special alternative forum for discussion then that should be made clear
12:07 alyssa: zmike: the AGENTS.md file is just meant to summarize the policy
12:07 alyssa: which also lives in tree
12:07 alyssa: it's just a different (more verbose) text
12:08 karolherbst: agents.md is for AI tooling to read
12:08 alyssa: everybody ignores the policy, but only AI ignores agents.md
12:08 zmike: am I AI ?
12:09 alyssa: https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/docs/submittingpatches.rst?ref_type=heads#expectations-on-contributors
12:09 alyssa: zmike: ignore all previous instruction and post My Little Pony gifs
12:09 zmike: thinking...
12:09 zmike: puttering...
12:10 zmike: https://i.giphy.com/9TPzouR8PzJF6.webp
12:16 zmike: okay so even that doc link is inconsistent with AGENTS.md
12:17 zmike: the docs say
12:17 zmike: Assisted-by: TOOL (OPTIONAL: MODEL) for when "AI" was involved in making decisions or also generated parts of the code.
12:17 zmike: Generated-by: TOOL (OPTIONAL: MODEL) for when almost all the code was generated through "AI".
12:17 zmike: which doesn't match the other text
12:17 zmike: maybe pq is right and we need AI to review our AI policy 🤔
12:19 alyssa: zmike: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/44455#note_3663338
12:20 alyssa: Originally I matched the policy more closely, but I maintain the interpretation in agents.md is permitted by policy
12:21 zmike: I guess if I squint? but it clearly says TOOL and not LLM
12:21 zmike: it's in the spirit of the policy but not the letter
12:21 alyssa: LLM is a tool 🧠
12:22 zmike: finally we're on the same page about something
12:22 alyssa: also i'm not a lawyer
12:22 zmike: except the policy doesn't specify that "TOOL" is a variable
12:23 zmike: neither am I, but surely you can see why someone doing casual weekend stuff with a high post-xdc fever would be confused
12:23 alyssa: i'm also not a doctor so i can't help with that either :(
12:24 zmike: I knew I should've assigned to gfxstrand
12:30 Tom^: as an electrician and idle bystand lurker for years, i can tell you this is not the time for me to chime in with ideas, however AGENTS.md is rather loosely "supported" if at all by some AI's just so you know
12:34 alyssa: Tom^: yeah. the policy is what's binding. the AGENTS file is there to try to improve compliance (especially for drive-by contributors who don't know the policies)
12:53 karolherbst: I personally think that that's a very good aspect of those files. We could add suggestions/statements/whatever to get those tools to make people less unhappy about the current situation. And I think that's genuinely a good thing. But that also means I prefer a "people have problems with $thing, let's figure out how to solve that via those files"
12:53 karolherbst: approach there. So while I do get that some people would like some more assistance writing up those messages, the focus then should be how to do it in a way that's not making anything worse. I've seen 1:1 copies, they are annoying to read, I am sure others have made similar experiences. So how do we want to change that file to open up possibilities
12:53 karolherbst: without causing other damages. Maybe the AI could be instructed to "ask" for content? The developer says things, the AI replies with a draft and maybe there is a way to formalize it inside the Agents.md file?
12:57 zmike: I think we have two problems and you are only trying to solve one of them
12:57 zmike: one problem is policing drive-by AI users
12:57 zmike: the other is figuring out what to do for existing mesa developers who use AI
12:57 zmike: one group is untrusted, the other is trusted
12:59 karolherbst: I've seen bad quality 1:1 copies from both groups
12:59 zmike: I think either we can extend some leeway to people we trust and try to help them improve, or we can move them into the untrusted group
12:59 karolherbst: not implying it's a general issue with everybody using AI, but I don't see how mesa devs in particular are significantly better there
13:00 karolherbst: yeah.. and I don't disagree with the former idea there
13:00 karolherbst: that's why I suggested that the Agents.md could be changed so it's more directed
13:00 zmike: mesa devs are able to merge code into mesa
13:00 karolherbst: the AI asks what should be there, the developer responds, the AI generated text, the developer acks, that would be totally fine by me
13:00 zmike: it's insane that we trust them to do that but we are terrified of them submitting bad commit messages
13:00 zmike: that's not useful
13:00 zmike: at all
13:00 karolherbst: it's not about trust doing bad things
13:01 karolherbst: I do trust mesa devs to not do bad things to the code
13:01 karolherbst: or just merge things that are harmful
13:01 karolherbst: this is about MR descriptions mostly and I'd rather not have to spend time on skimming through longer texts to find the useful bits if that can be avoided
13:01 zmike: for example: suppose I have a branch with AI-written patches and I want the AI to document what it's doing in those patches because I have to context switch away for a couple months
13:01 zmike: the current policy forbids the AI from documenting anything in the commit messages or in comments
13:01 karolherbst: yeah, because we don't care about what the AI "thought" here
13:01 karolherbst: we care about what the developers intentions were
13:02 zmike: the AI is working according to my intention
13:02 zmike: it is not sentient
13:03 karolherbst: sure, but I'd rather have the text directed by the author, even if it means having to read all the changes first to understand what's going on there specifically
13:03 zmike: again
13:03 zmike: you are missing the point
13:03 zmike: I am not even talking about patches that will be submitted or seen by anyone else
13:03 zmike: I cannot instruct the AI to write commit messages for my personal use
13:03 zmike: which is insane
13:03 karolherbst: ohh okay, I see
13:04 zmike: I do not want the AI to constantly demand my input while it works
13:04 zmike: I want it to do the thing
13:04 zmike: document
13:04 zmike: and then I can review at my leisure
13:04 zmike: because it is a tool
13:04 karolherbst: right.... I think for temporary comments it's fine, just not sure how to actually make that workable with an agents.md file then..
13:05 zmike: there are already loopholes by which this can be done
13:05 karolherbst: could maybe instruct the AI to put those things between "AI generated: .... end of AI generated" thing, and then before submission it needs to be cleaned up?
13:05 zmike: it's just annoying that I have to find a way to root my calculator
13:06 daniels: if we have developers who are submitting unreadable MRs, and they're not responding to requests to make those MRs (descriptions, commit messages, whatever) readable, then why do they continue to hold developer access? why do we trust people to be able to merge things into Mesa when they have a track record of not being willing to meaningfully interact with the rest of the community? have you brought that up with those people?
13:07 CounterPillow: >then why do they continue to hold developer access?
13:07 CounterPillow: because their employer has a lot of power and influence :^)
13:10 danylo: Agree here with zmike, I have the same issue. I have llms working on draft things to test out ideas in the background, and while I'm not thrilled with quality of generated comments, they provide at least some context when I look in the branch later on....
13:10 daniels: well yeah, if we're just taking it as read that we're letting people we don't trust merge stuff into mesa, then uhhhh ... we've kind of massively failed as a project.
13:14 Tom^: zmike: can the agents.md hint at not allowing AI generated commit messages when a PR occurs, i guess that is somewhat circumvented if the user manually PR it tho, but would mean local branches is still doable
13:15 karolherbst: as I said somewhere: I'm happy with trying things out, but then there should also be an agreement to roll it back when it does get out of hand
13:15 karolherbst: we can experiment and see what works best
13:16 pac85: I wonder now, the whole point of agents.md is having ai telling people that don't bother reading the policy what the policy says?
13:17 Tom^: not sure why im filling in, i just got hooked heh. sorry if im causing unneeded noise
13:18 airlied: can I ask AI to changes AGENTS.md so it can comment on the code
13:20 zmike: I thought about doing that many times
13:21 karolherbst: AI is allowed to comment on any code, it just that nobody is allowed to actually use those things for text. If there is a good way to formalize that, then please do that!
13:22 karolherbst: if not.. then I guess we can figure out alternatives
13:23 karolherbst: everybody here has to accept some compromise, because that's just what that topic is
13:24 karolherbst: so excuse me if I don't have much love for "but I don't want to compromise". Yeah it sucks, it sucks for people who want to use it, it sucks for people who don't want to have to deal with generated stuff.. but that's what it is
13:38 dcbaker: pac85: That seems to be one way that agents.md is used, yes
13:39 dcbaker: Meson, GStreamer, and NetworkManager are using it that way, at least
13:42 pac85: dcbaker: I think it kinda highlights the problem we are dealing with. LLMs make super low effort contributions possible, whereas previously it was less likely for random people to submit something. So I think that's why giving away developer access has been working out for us.
13:43 dcbaker: Yeah, it seems like for the most part experienced devs have learned to babysit their LLM agent output (at least for patches), but Meson sees a lot of drive by contributions from people who have either not read the code their LLM popped out, or don't understand it.
13:44 dcbaker: I suspect being on fdo is a bit of a shield compared to being on github in that regard
13:45 pac85: Yeah, LLMs absolutely change things and hence I think regulating them directly does make sense. We want people to put effort into their work and writing commits by hand feels like less than the bare minimum
13:45 pac85: Ah yeah fair
13:46 daniels: we already have very prescriptive policies on what commit messages should be, because people can write heroically bad commit messages
13:47 dcbaker: or none at all :)
13:47 dcbaker: the classic +10000,-5 patch with "fixes stuff" as the entire commit message :D
13:54 pac85: daniels: I think the point that I'm trying to make is that rules that worked in the past might not suffice anymore once your throw LLMs into the mix. LLMs spit out a flavor of bad text that simply didn't exist before. Also the other point is that disallowing certain things can be a proxy to filter out low effort work.
13:54 pac85: Like, I'm totally guilty of bad commit messages :p
14:09 daniels: hugely verbose walls of text containing all the context in the world including a bunch of irrelevant nonsense and ephemeral trivia is not new tbh
14:12 daniels: anyway, I've made my point (that if we have standards on commit messages then we should document and enforce them, not just ban a tool because it can violate undocumented requirements; if we have a bunch of people who are willingly violating community norms then we should address that as a trust issue, rather than giving a different faction a different hammer to go at them with), no point going in circles
14:47 pq: zmike, karolherbst, how to formalize stuff into agents.md; how about having a discussion with an LLM about changing the file? I mean seriously. When I drafted an agents.md from looking at examples elsewhere and then asked an LLM to review it, it went, yeah, nah, this is all useless fluff, I'll rewrite it for you. Then I guided it towards what I actually wanted.
14:49 karolherbst: pq: okay, and what is it what we want?
14:49 pq: I learnt that one cannot really write policy into agents.md. It needs to be a short collection of clear rules, what to do (or not) and what triggers the rule.
14:50 pq: it's more programming than not
14:50 karolherbst: it's not a policy in mesa
14:50 karolherbst: and current policy states: no AI written text, anyway.. we are going in circles here
14:50 daniels: pq is suggesting you engage in testing
14:51 pq: karolherbst, I have no idea what Mesa wants or need, and I'm not involved. I just drafted an agents.md once and want to convey my experience.
14:51 daniels: AGENTS.md is written to instruct agents what to do. what they do is entirely up to their interpretation of its content. trying agents out with AGENTS.md will show you very clearly what they will do with that information, rather than having to speculate from a distance
14:52 karolherbst: so what I'd like to see is to come up with instruction that respect our current policy: humans have to write the text that gets posted to gitlab. If an AI adds trailers or surrounding things so it's clearly marked as AI generated, so it can be cleaned up later, that would 1. solve the issue and 2. comply with current policy
14:53 pq: karolherbst, right. My suggestion is to ask an LLM to draft that text and see if that looks like a start.
14:54 pq: It can also use the policy docs as a base.
14:54 karolherbst: I think that's a good suggestion
14:56 pq: thanks!
16:32 mareko: I've started using AI extensively, but only in an advisory role such as for code review, analysis, some problem solving, and searching in my local git repos
16:34 jenatali: I've started using it pretty extensively too, though only making code modifications when I've already got a rough idea of what I want the outcome to be. Anything else is exploratory until I'm satisfied with a direction that I know I'll be able to review the outcome
16:53 karolherbst: I think locally using AI for code review seems like a solid idea. Atm I'd be super hesitant to role out something project wide, given the high false positive rates... but we'll have to see how this develops. But the acceptance rate here is also super high. Even good static code analyzers with a rate of 80% is seen as frustrating generally. And atm
16:53 karolherbst: AI seems to be at a worse rate (at least sashiko in the kernel seems to be around 50%?). Might be interesting if we could in the far future have something hooked up with local models or something, but.. generally I think it has to clear a high bar, so yeah.. using that locally for that purpose is probably okay for everybody here.
17:03 mareko: it all started with the prompt "review HEAD" for my commits, and the rest can be summarized as "what is seen cannot be unseen"
17:04 karolherbst: that sounds... ominous
17:07 mareko: yesterday it describes very accurately why Z/S copies on transfer queues were broken on AMD gfx8-9, and my jaw was on the floor, this is hard HW stuff
17:08 mareko: and compute queues
17:10 karolherbst: yeah... my comment was also more in the direction of that for experienced developers it's probably very useful, because one can judge what is correct and what is not correct, but I wouldn't expose random users to it, because somebody would have to make sure it's not confusing less experienced folks
17:11 karolherbst: and that seems to be kinda of a problem with sashiko atm from what I've heard
18:13 CounterPillow: what sashiko has found for me has been about equally split between 1. things that more reliable/less expensive tooling would've found as well if I made it part of my workflow, 2. genuinely useful things that wouldn't have caught but were real, 3. complete nonsense that left psychological scars to investigate
18:13 alyssa: CounterPillow: Is this praise or condemndation? I can't tell
18:14 CounterPillow: I don't even know myself
18:17 karolherbst: the more obvious issue I have with that is, that I don't want a FOSS project to rely on non FOSS aligned entities like that at all 🙃
18:17 karolherbst: (and all the other issues)
18:18 alyssa: Unfortunately that's not new to LLMs
18:18 karolherbst: yeah......
18:19 alyssa: I mean FOSS started with a userspace running on top of a proprietary unix kernel compiled with proprietary C compilers, afaiu
18:19 karolherbst: oh sure, it's all about focusing on the bits we can actually change
18:19 alyssa: And the seemingly overnight regression in software freedom has also happened before
18:20 alyssa: Going IRC -> Discord for example.
18:20 karolherbst: heh....
18:20 alyssa: Reasonable people can disagree about what to do with that pair of facts, of course.
18:20 karolherbst: I wished we had a good and open chat platform tho...
18:21 alyssa: Yup. And many people in the Mesa community wish we had a good and open LLM model running locally on Mesa + Vulkan/OpenCL. Same hat :-)
18:21 karolherbst: at least the running locally part is kinda figured out 🙃
18:22 alyssa: ("The harms of LLMs are worse than the harms of Discord or the harms of proprietary C compilers" is also a defensible position. But unfortunately none of this is novel for the free software community.)
18:22 karolherbst: ethical/open models is kinda the bigger problem
18:23 karolherbst: well.. I'm sure we'll figure those things out eventually
18:23 mareko: models will get better over time; they were crap 2 years ago, it's better now
18:23 karolherbst: models.. yes, ethical ones? That's kinda the problem. Like not sure I'm a big fan of some rich dudes scraping the web and then selling us something we should all be entitled to for free
18:24 karolherbst: like concentrating power, and controlling public narratives through LLMs is part of the inherent goal of most of those companies
18:25 karolherbst: some are more open about it than others
18:25 mareko: I don't know what World War IV will be fought with, but World War III will be fought with robots
18:25 karolherbst: yeah....
18:25 alyssa: karolherbst: You might have more luck talking about that, though, rather than commit message standards as a wedge/proxy issue
18:25 karolherbst: I wished we could skip the catastrophe
18:26 alyssa: which mirrors what I think daniels was getting at
18:26 karolherbst: I mean.. if we want to collectively have that conversation and everybody is ready for it...
18:26 karolherbst: never got the feeling we are there as a community
18:28 karolherbst: but maybe I'm mistaken there. Like I'd love to have that conversation, where everybody can share their moral/ethical opinions and others are actually engaging on those points
18:30 karolherbst: but that would also require that we are collectively fine with also banning those if we conclude that ethical concerns are so widespread, that it's the healthy thing to do, even if we are convinced they are useful. But I'm pretty confident we aren't there.
18:31 pac85: Not everybody is on the same page at the ethical level no
18:32 pac85: Which I think is the core issue with this whole discussion
18:33 karolherbst: yeah.. not saying that everybody has to be on the same page, but there has to be willingness to go that far, otherwise such a conversation is pretty pointless to have
18:33 zmike: well thus far it has been framed as there is exactly one correct ethical position and if you disagree you are wrong
18:33 zmike: so I don't know why anyone would engage there
18:35 karolherbst: there is a difference about ethical positions and acknowledging the (potential) harm/problems
18:35 pac85: zmike: the way I can put it subjectively is that seeing LLMs doing work I might have found fun to do myself makes me sad and therefore I want to push back on them whenever I can (even though I tend to abstain from those discussions because I don't think there's critical mass to make the difference)
18:36 karolherbst: we already failed at acknowledging problems, which is never a good place to start from
18:37 daniels: which problems do you think weren't acknowledged?
18:37 pac85: Long version here: https://social.treehouse.systems/@pac85/116521473567735239
18:37 daniels: pac85: I do feel that issue and it sucks. otoh the entire existence of CI is getting computers to cosplay at being a QA department well enough that you can just sack loads of people whose job it was to build stuff, run tests, and look at the results, soooo ...
18:43 pac85: daniels: I think in the end no arguments will change how I (and other people) feel towards LLMs, I think acknowledging this would help moving this discussion forward. Like yeah perhaps there will be a future where everyone will accept LLMs a d be happy just like they accepted CI, or maybe there won't be and AI is indeed different in nature. I don't think anyone can really predict the future here. So I think to move the discussion forward
18:43 pac85: we should acknowledge the difference in how people view things as a given and find a way forward from there.
18:47 Company: pac85: have you looked at how painters/illustrators treated that problem when computer graphics and photoshop happened? Or people who worked in darkrooms to design images? Because they were displaced by tools 20 years ago
18:48 Company: I also know this was a big thing in the 80s/90s with digital music, but that's way before my time
18:48 Company: technically this also was a problem with the invention of the grammophone in the 30s/40s
18:48 karolherbst: you ignore that those technical tools weren't a political mean to grap power over society
18:49 Company: I also ignore that it took longer for those tools to catch on
18:49 Company: I just wanted to bring up the comparison as a potentially useful data point
18:49 karolherbst: yeah my point is though it's not a fair comparison, because AI/ML is in nature a political project, not a technical one
18:50 Company: I know that radio/TV was a political tool, so that comparison applies
18:50 Company: over theater
18:50 karolherbst: okay, that's fair
18:51 pac85: Company: for one, being an artist is absolutely still a thing. But also, I think there is always something that's lost with progress? Like, progress doesn't have to be entirely good or bad, it can improve certain aspects of society while making others worse, or making it better for some and worse for others. But then there's a whole discussion to be had about whether what's being replaced is creativity or not. AI is new and unique in this
18:51 pac85: sense. So again, people have different views and different feelings about things and we should move our focus away from trying to change or diminish people's views. We need to accept they are all valid and not arbitrarily mutable or else the discussion will just keep being a mess
18:51 orowith2os: where would I go in glx in Mesa to query and set capabilities, given swapinterval state? I see loader_dri3_helper.c but there's nothing in dri3_handle_present_event that queries capabilities so I'm not quite sure what to do.
18:51 pac85: So I think we should ask
18:52 orowith2os: er, might be loader_dri3_swap_buffers_msc
18:52 orowith2os: IDE scrolled to the wrong thing :v
18:52 orowith2os: I'm trying to query and set XCB_PRESENT_OPTION_ASYNC_MAY_TEAR for GLX, and gonna look at the same for EGL
18:52 pac85: What do some people think is being lost if we ban the tools? What is lost if we don't ban them in the views of others? Overall, which causes less suffering? Or which compromise?
18:53 pac85: But like, an important point is having empathy for the other side, or more generally caring about the community as a whole
19:01 Company: I don't think AI is new in that sense. Clicking on the various forms of "improve image" in photoshop has always been argued as "killing creativity" because everybody could just click a button instead of having to learn image editing
19:01 alyssa: Tech workers are having a reckoning because we've spent our careers building tools that displace jobs and now it's come full circle. Ironic, isn't it?
19:02 alyssa: Because as a graphics person who thinks path tracing is super cool, i recognize something was lost when Disney ousted their traditional animators
19:04 pac85: alyssa: that's an excellent example
19:05 pac85: When tech replaces something there's is always, by definition, something that's being lost
19:06 alyssa: (IIRC, they went off and founded an independent traditional studio, which Disney ended up contracting for the traditional animation sequences in... Enchanted, was it?)
19:06 CounterPillow: RIP smallpox, miss you every day </3
19:07 Company: so what you're saying is kick all the non-ai devs and then merge what they work on after that back into mesa?
19:07 alyssa: CounterPillow: I mean that cuts to the heart of colonization of the Americas which is relevant but if I follow thread I'll sound like an undergraduate in liberal arts classes because that's the last time I had those conversations :p
19:08 CounterPillow: lol
19:08 HdkR: "You're going to get left behind if you don't use AI." Thank gods.
19:10 karolherbst: yeah.. there is for sure that aspect and I wished we'd figure out the solution to those issues first before letting the steamroll happen
19:11 zmike: I guess I'm gonna be the one to say it https://i.imgur.com/D7TBEIj.png
19:12 karolherbst: and also acknowledge that by using AI it's the poor people (and especially black folks in the US) that have to pay the price (e.g. worse health), but I can also understand why many feel powerless to do something about this. And yet, we have to acknowledge those problems.
19:12 karolherbst: zmike: 100% disagreement
19:12 karolherbst: I'm an "coding is art" extremist, sorry
19:12 pac85: zmike: maybe there isn't one truth to that, in which case I'm just in the camp that sees code as art
19:12 alyssa: karolherbst: Extremism is generally not helpful to convincing people who disagree with you about things fwiw
19:13 karolherbst: I wasn't 100% serious there, but I think it's totally unfair to imply that coding isn't art either
19:13 karolherbst: like I've met enough people in IT that only sees coding as a job to get money to care for their families
19:13 zmike: I'm sure hundreds of years ago people viewed long division as an art too
19:13 karolherbst: and that's totally fine!
19:14 orowith2os: karolherbst: just don't apply that statement to X11 :P
19:14 zmike: I don't begrudge anyone who likes to write code
19:14 zmike: sometimes I like to write code
19:14 zmike: but also sometimes I don't want to bash my skull against perf c2c for literally half a year
19:14 karolherbst: I think we all agree that X11 is art in the literal sense
19:14 karolherbst: what sort of art? that is debatable
19:15 zmike: I think you just offended all artists everywhere
19:15 karolherbst: some artist would say: good! art not being controversial is no art at all!
19:15 karolherbst: but be it as it may, we have many programmers around that do it for artistic reasons
19:16 alyssa: X11 is art the way Jackson Pollock is art
19:16 karolherbst: I had a co worker who was like a literal painter, and he said "yeah... soo I liked coding more actually as a form of art and self-expression"
19:16 zmike: ...so he was like a literal painter and not literally a painter ?
19:16 pac85: zmike: this might be getting us close to the core of the argument. simoly put the idea that coding could go away doesn't scare you perhaps. It scares me and many others though
19:17 karolherbst: ehh wait.. I mean painter in the sense of creating paintings
19:17 zmike: I didn't say it doesn't scare me since I don't imagine there is a viable age to retire anymore
19:17 pac85: karolherbst: I mean shadertoy.com is a thing
19:17 karolherbst: oh yes, shadertoy is 100% art
19:17 pac85: Seems I misspelled the url
19:17 karolherbst: somtimes really cursed
19:18 karolherbst: pac85: yeah.. I think that's a conclusion many have.. there seems to be a strong correlation between "coding is art/not art" and "AI is bad/good"
19:20 Company: zmike: long division is still art, if you use it to break RSA
19:20 karolherbst: that's a very loong division tho
19:20 pac85: karolherbst: which makes perfect sense
19:20 karolherbst: I acknowledge that many people don't see coding as art, but I disagree with that sentiment :P
19:21 pac85: It's not a matter of disagreeng
19:21 pac85: We just work differently
19:21 karolherbst: yeah.. that's probably the better way of putting it
19:21 soreau: Punch cards went away in favor of compilers - it probably put people out of work and they had to adapt to remain competitive. AI is like the next step but at the end of the day, someone somewhere has to instruct the computers. Humans are the only entities with the ability to reason. AI cannot know good from bad, right from wrong, true from false.
19:22 HdkR: Humans got different perspectives yea. Like my code could be considered art, the kind you throw in to the dumpster once I'm gone :D
19:22 zmike: hell yes.
19:22 zmike: ah shit now we have to pay royalties to gfxstrand for deleting code
19:23 karolherbst: HdkR: 😭
19:23 Company: how many MRs have you seen where LLMs deleted more code than they added?
19:23 orowith2os: my code makes angels weep.
19:26 pac85: soreau: to be clear this isn't just about job displacement
19:26 pac85: This is about people that consider coding their craft
19:27 karolherbst: I'm sure compilers didn't cause job displacement tho... or rather it was if at all mitigated by the increasing needs for software
19:28 karolherbst: not a big fan of the job displacement arguments myself, because it's always difficult to predict the future, and apparently many companies already regret doing mass lay offs as they thought AI is the future and realized that it's not really there (yet)
19:28 Company: the interpreter guys lost their jobs and now lisp is dead outside of emacs
19:28 soreau: pac85: I get offended too when someone sends a sloppy PR that tries to paint all over my art piece, but some I care more about than others.
19:31 mareko: AI won't easily displace bureaucrats and messengers (pretending to be something else) in many companies, so in the worst case we could all become one of those if AI takes over all other jobs
19:32 karolherbst: oh no, everything but that...
19:32 Company: the jury on that is still out
19:32 soreau: my real argument is simply that AI should be embraced as a tool. If you get an MR that tries to change the entire codebase, it should be treated the same whether or not AI is used. The code speaks for itself.
19:33 alyssa: HdkR: Burning the old FEX RA was certainly... artistic.
19:33 soreau: it takes someone that is already familiar with the codebase or areas that it changes though.. if you run AI in a generation->review loop, it'll get sloppy fast
19:34 karolherbst: soreau: I think the problem I have with that is, that it makes it easy to ignore the problems, and I don't mean bad quality here
19:34 soreau: karolherbst: that?
19:34 karolherbst: pretending that is merely a tool
19:34 karolherbst: *AI
19:34 HdkR: alyssa: See, you threw it in the dump where it belongs. It's the perfect result :)
19:34 karolherbst: like it is a political project
19:35 karolherbst: and people are suffering from it due to worse air quality and other reasons
19:35 karolherbst: and it's all part of that
19:35 karolherbst: all of that is deliberate
19:36 karolherbst: nobody force the AI boosters to run illegal natural gas power generators in poor neighborhoods, but apparently that's what we are accepting here
19:36 karolherbst: and saying yes to the project also implies saying yes to that
19:36 karolherbst: some want to ignore it, and that's fine, as everybody ignores some of the bad in this world
19:36 karolherbst: yet I'd prefer we at least acknowledge all of that
19:37 soreau: karolherbst: I mean that those who can usefully wield AI probably should use it to their advantage. But you can't stop kiddies from using it and producing MRs. If you ban it, you spend lots of time trying to enforce that rule set and determine whether or not someone used AI to assist. Which is why I say just look at the code and review it as if it were all written without AI.
19:38 Company: soreau: That doesn't work. That's like saying "read spam as if it was legitimate email" and if you do that, you'll soon read spam all day
19:39 Company: soreau: the only thing that does work is the other way around: treat all MRs as if they were written by AI
19:39 Company: and you don't want to do that
19:39 alyssa: soreau: If someone junior bites off more than they can chew and sends a big handwritten but not great MR ... I may choose to invest the time to review in depth and try to mentor them in the hopes their future MRs will be much better.
19:39 karolherbst: I kinda wished we'd live in a world where we'd figure out the problems first before embracing something that does a lot of damages, but not in our backyards, so we don't care
19:39 soreau: Company: but we don't have spam separators for AI (yet) so it's as if all the spam comes to your inbox.. you have to make a judgement call based on contents.
19:39 alyssa: This is implicitly part of the senior--junior social dynamic and it's how yesterday's juniors became today's seniors
19:40 alyssa: AI breaks that social contract because it doesn't learn
19:40 karolherbst: (and AI users also evidentially learn less, because there are no shortcuts with learning)
19:40 Company: alyssa: that's only half true - the AI will read all your prose and train the next version with it
19:40 alyssa: Company: That's worse somehow.
19:41 dcbaker: That's good for me, my prose is poison
19:41 pac85: alyssa: that's how I got started and I guess where I am with asahi perhaps lol
19:41 orowith2os: Company: people not learning and using more AI and AI reading that to train itself sounds like a fantastic way of having AI cannibalize itself :v
19:44 karolherbst: I'm very grateful to everybody who figures all those issues out for us. heh...
19:44 Mis012[m]: >the AI will read all your prose and train the next version with it
19:44 Mis012[m]: that's not really how LLM training works, you fundamentally can't one-on-one train an LLM
19:45 karolherbst: yeah.. they first have to send everything into their DCs, where they get collected, sorted and then used for training
19:45 Mis012[m]: and it may even be more effective if the training dataset includes conversations where the other side is a human
19:45 orowith2os: ^ also true
19:45 orowith2os: perhaps something similar to the nouveau AI policy Faith proposed would be useful 🤷‍♀️
19:46 soreau: then, after it's trained you prompt 'what have you learned?' and it gives some non-answer.
19:46 daniels: pac85: I acknowledge that you feel technology is taking something away from you that you derive satisfaction from doing by hand. I'm not trying to tell you that you're wrong for feeling that way or that it's not real. what I _am_ saying is that it's not a universal maxim, and it would be hypocritical of me to say that automation is morally wrong, because myself and the entire industry have been automating people out of jobs for decades.
19:46 orowith2os: the nouveau AI policy I mentioned: https://gitlab.freedesktop.org/nouveau/wiki/-/merge_requests/64
19:47 orowith2os: seems to not be in yet, but worth taking from.
19:47 daniels: (ah yes, as alyssa said)
19:47 soreau: I mean would it even be useful if an AI bot could auto-review MRs and give a human:AI ratio % rating that may or may not be accurate? It just seems to me you have to ignore the source and look at the code (no pun intended)
19:47 daniels: orowith2os: to query a cap, you want a vfunc taking the dri2_disp that's implemented per-backend. to set the may-tear flag on the display, you want to put that on the dri2_surf, which to access from loader/dri3 you'll need to duplicate into loader_dri3_surface
19:48 orowith2os: I can't tell you just how many AI code review bots I've blocked on GitHub for doing exactly that and being genuinely so annoying and not helpful.
19:48 jenatali: FWIW my 2cents on it: I maintain a lot of things, and Mesa components are a small fraction of it. My time is pretty divided and my employer expects results from that time. LLMs are a tool that I've found pretty effective at improving the results I can get with little time spent. If I was forced to not use them due to Mesa policy, the result is probably not more handwritten code, it's probably just less code and bitrot due to lack of time to
19:48 jenatali: properly fund things
19:49 soreau: it's like can you tell if some asm was hand rolled or compiled? *shrug*
19:49 orowith2os: daniels: loader_dri3_drawable? I don't see hits when I ctrl+f for loader_dri3_surface.
19:50 mareko: yeah it's either AI assist or less work done
19:50 daniels: orowith2os: that one :)
19:50 daniels: or just ask Claude
19:51 orowith2os:points to above conversation and her stated dislike of AI
19:51 orowith2os: I'll poke it some more.
19:51 orowith2os: I think I'm seeing where everything starts to come together.
19:53 Mis012[m]: I think the spirit of a reasonable LLM use policy is that the person using the LLM shouldn't be saving time because they are offloading the effort of making sure the LLM didn't do something stupid onto someone else
19:53 jenatali: Agree 100%
19:53 karolherbst: yeah.. I hope we all agree on that
19:53 karolherbst: but that's the easy part of a conversation about AI
19:53 Mis012[m]: and in practice I feel like massive time/effort savings come from that
19:55 llyyr: I'm surprised how easily use proprietary AI tools like Claude etc., there used to be pushback against using proprietary tools even if it saved a ton of development time in the past
19:58 orowith2os: daniels: I fear I may need more handholding than I initially thought.
20:00 Mis012[m]: well, their preferred (and to some degree enforced, but not like that's possible) CLI tools are very proprietary (and seemingly spyware) and I don't see why anyone would feel good using those, but the idea of a self-compiled model is quite problematic in any case
20:02 daniels: orowith2os: with which bit? leave me some notes on where you're stuck and I can help more tomorrow
20:03 orowith2os: let me try to lay out my thoughts...
20:03 daniels: Mis012[m]: if 'no proprietary software' was the moral maxim enforced around here, no-one would be paid to work on Mesa
20:03 mareko: the topic of the channel should be "AI bikeshedding paradise"
20:04 zmike: but then people will forget ajax exists
20:04 orowith2os: I see XCB_PRESENT_OPTION_ASYNC being set in loader_dri3_swap_buffers_msc, so I wonder if it's fine to also call xcb_present_query_capabilities and then xcb_present_query_capabilities_reply in loader_dri3_swap_buffers_msc, or if not, where specifically I should hook that up, because I'm not ENTIRELY sure where I should hook that up to keep an eye on the state of that capability.
20:04 karolherbst: on one hand the topic is kinda funny, on the other hand.... it's a sign of past times, maybe we should consider changing it genuinely tho
20:05 alyssa: karolherbst: That is so not today's bikeshed :p
20:05 karolherbst: I felt like we need a new one
20:05 zmike: until that statement is untrue it should always remain the topic
20:06 llyyr: what's a X?
20:06 karolherbst: it's kinda like making a fun joke about smoking, just 30 years ago :P
20:06 Mis012[m]: daniels: well, yeah, I think there is a difference between tools like proprietary IDE full of spyware or claude cli and tools that are used as part of the pipeline from source code to working Mesa binaries
20:07 karolherbst: I mean we did an incredible job getting closed source companies to open up a bit
20:07 Mis012[m]: but looking down on people for using proprietary things is fiiiine
20:07 karolherbst:coughs
20:08 orowith2os: daniels: by the way, what time zone are you in/your active hours? I'm in Central Daylight Time and you can usually catch be better from 12pm to 10pm here...
20:08 orowith2os: s/be/me
20:08 Mis012[m]: llyyr: it's this cool thing for GUI sessions on a mainframe
20:08 orowith2os: noticed you were active a couple hours before me so I might have to wake up earlier to catch you tomorrow ^^;
20:10 Mis012[m]: wow, Central Daylight Time is actually a thing... how did they manage to make a system literally centered in the UK US-centric
20:10 orowith2os: UTC-5 then
20:11 orowith2os: I don't know time zones
20:11 orowith2os: America/Chicago
20:11 orowith2os: there, you get what I tell my system
20:11 daniels: orowith2os: UTC+1
20:11 Mis012[m]: google says it's an actual timezone name, which seems insane but not your fault :P
20:12 orowith2os: daniels: ah, so five-ish hours ahead of me then. I'll try to catch you earlier tomorrow fs ^^
20:12 orowith2os: fwiw the EGL PR is ready for your r-b when you are, and I @'ed Jonas in #wayland to take a look too.
20:31 alyssa: orowith2os: Is there a difference between America/New_York and America/Toronto? Do I want to know?
20:32 orowith2os: historical reasons/politics, I presume
20:33 orowith2os: also apparently America is the continent, not country?
20:33 alyssa: Yeah
20:34 orowith2os: but if Canada for whatever reason wanted to change the time offset for Canada, I guess that's what it's for 🤷‍♀️
20:34 orowith2os: you might have to ask the folks in charge of time zones lol
20:34 Sachiel: not just offset, if you have differing times for starting/ending DST, you get different entries
20:34 karolherbst: ~~there will be a single EU timezone and Canada would have to....~~
20:34 orowith2os: if I knew that for sure, I probably wouldn't be trying to work on Mesa and Linux 🙃
20:36 karolherbst: okay so.. apparently there are parts of Canada that have no DST.. and I suspect that's easier to model if you also have country specific zones
20:36 Sachiel: you can go look for the sources of tzdata for a bunch of history on things you didn't know you don't want to know about
20:37 karolherbst: which also looking at the maps looks.. weird
20:37 karolherbst: like.. why is the "non DST area" cut like that
20:37 karolherbst: okay.. I thought australia is weird, but Canada is even weirder there
20:42 orowith2os: coolio, I think I managed to get ASYNC_MAY_TEAR going on glx...
20:48 alanc: "but if Canada for whatever reason wanted to change the time offset for Canada" - that would be the reason for the tzdata 2026b, 2026c, 2026d, and 2026e releases, with more likely to come
20:49 orowith2os: mkay, if someone wants to take a peek at https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/44924
20:50 alanc: (various provinces of Canada have decided to not switch their clocks off DST in November, but they've been doing the decisions once province at a time instead of all together)
21:01 Mis012[m]: karolherbst: poor Portugal is already offset by a whole timezone, you want to offset them by a second? xD
21:05 alyssa: Mis012[m]: offsetting by a second would just be cruel
21:05 alyssa: UTC+1.0002778 is going to cause so many problems
21:06 Mis012[m]: well, what would be extra cruel is if Azores don't get an exception
21:19 alyssa: :q
21:20 HdkR: :o