00:28 karolherbst: well.. we already have some sort of middle layer with the vulkan runtime as well
01:00 mareko: when a game is GPU-bound, optimizing for CPU overhead is worth it if the goal is to reduce power consumption
01:05 HdkR: Optimizing for CPU overhead makes my life better. Please think of your local emulator developer.
01:10 HdkR: One dollar a day can help save millions of starving emulator authors around the planet.
04:30 alyssa: HdkR: *Millions of dollars a day can help save one emulator author
04:38 HdkR: I mean, I wouldn't say no :P
05:26 CounterPillow: mareko: do those types of optimisations actually allow a CPU's power budgeting to use a lower OPP in gameplay scenarios?
05:31 mareko: the CPU will see lower load, the rest is up to power management
07:29 linkmauve: CounterPillow, also there can be huge variations in CPU used by the people playing the game. My game might run at 2000 fps on my laptop, but it will barely keep 60 fps on my phone for instance.
07:29 linkmauve: And there are even worse CPUs than this Cortex-A53.
07:36 HdkR: Indeed, think of the people with Cortex-A34.
08:13 karolherbst: Yeah... like.. sadly lowering CPU overhead on the applications thread and increasing power efficiency aren't the same thing :')
17:17 zmike: for discussion https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/44878
17:25 karolherbst: you are really on a roll today to provide examples on why we need governance, aren't you :P
17:27 zmike: I think it's just common sense updates
17:27 zmike: the current policy is truly pointless
17:28 zmike: this is like pre-optimizing some huge new feature before it even works
17:31 karolherbst: you are introducing a contradiction there
17:34 zmike: I don't like that we are telling the calculator it can't use division because humans might abuse it
17:35 karolherbst: this isn't about abusing it
17:36 karolherbst: We can just say that those AI dumps are super annoying and a pain to read and that most here rather wouldn't want to have to read it and write it down as an expectation
17:36 zmike: again
17:36 zmike: long AI dumps are not the fault of the AI
17:36 zmike: they are the human's fault
17:36 karolherbst: why then allow it for some and not others?
17:36 zmike: I'd be fine allowing it for all?
17:36 zmike: but I assumed that would receive more pushback
17:36 karolherbst: and I wouldn't be
17:37 karolherbst: it's up to the submitter to prepare it accordingly
17:37 karolherbst: if somebody takes an AI generated thing, and cleans it up properly, I wouldn't mind either
17:37 karolherbst: any 1:1 copy I would mind
17:37 zmike: I'm fine with that as a policy too
17:37 zmike: feel free to propose different wording
17:37 karolherbst: okay
17:38 karolherbst: I'll think about how to phrase it
17:38 zmike: if you're worried about long messages, you could even introduce a word limit
17:38 zmike: right now I can't even instruct AI to directly add a one-line description of a patch as a reminder
17:38 kisak: functionally, it's not about the tool that bypassed thoughtful writing, it's the 90% word padding that's bad signal to noise.
17:39 zmike: but again
17:39 zmike: this is a human problem
17:39 zmike: if the human is too stupid to have the AI generate useful text, that's the human being stupid
17:41 Company: so you want to say "make sure it doesn't look AI generated" but phrase it better?
17:42 zmike: I want it to be up to the human to submit good commit logs
17:42 zmike: the same as it's always been
17:43 karolherbst: Maybe we could turn it around a bit. The AI could ask the submitter for some text that is used to write a terse description from, but it shouldn't be the AI just generating text on its own
17:44 zmike: that's functionally what we have now
17:44 karolherbst: then what we have now is okay
17:44 zmike: it is not
17:45 karolherbst: why not?
17:45 zmike: because right now the policy encourages users of AI to just copy/paste the same text the AI would've added directly as the commit log
17:45 karolherbst: I want to read about the intent of the submitter, and what the plan is, and why we want to do this. Not what a AI came up with
17:45 karolherbst: the AI wouldn't have added the same text
17:46 zmike: again
17:46 zmike: I don't know how many times I have to say this
17:46 zmike: it is up to the user to make sure the commit log is good
17:46 zmike: that has always been the policy
17:46 zmike: it didn't magically change when AI tools came into being
17:46 Sachiel: the hope is that when a user asks the ai to generate the commit, the ai says "I can't because here are the project rules", so the user goes on to write the commit themselves
17:46 Sachiel: is their first instinct is to work around the rules, we'll see it and ban them
17:47 zmike: ...which is what we always would've done for anyone who submits patches with very bad commit logs
17:47 zmike: so why does it only matter in this case?
17:47 karolherbst: sadly AI did made this worse
17:47 Sachiel: because the bar is too low now
17:47 Company: what matters is what the AI tells the submitter
17:48 karolherbst: we are seeing more and more studies on this, and AI sycophancy is an issue, and that's one of the core issues with AI generated text. Nobody is immune to this.
17:49 karolherbst: people can use AI to polish up the wording for all I care, but no 1:1 copies from AI
17:49 karolherbst: no matter who
17:49 zmike: are you really going to make me say it again
17:56 karolherbst: I mean.. I don't want to point fingers, but I haven't seen "cleaned up AI text" that was actually properly cleaned up yet...
17:57 karolherbst: maybe you have good counter examples
17:59 zmike: I'll look at it over the next week for specifics, but thus far when I've had it write sample commit logs they've been brief, informative, and perfectly readable
18:04 karolherbst: mhhh.. looking over some of them, it does appear they are indeed better for at least commit messages, though I'd like to see more rephrasing there because some of the language certainly feels odd sometimes.
18:05 karolherbst: I wished they would be more terse tho
20:29 daniels: my take is the same as zmike: if you have standards about commit messages, then enforce those standards. making it about the tool means you give humans carte blanche to write awful commit messages. if the argument is really about just wanting to ban LLM contributions regardless, then have that argument rather than a proxy for it
20:33 mareko: good models can produce good commit messages, but the developer needs to understand the subject and guide the model to produce accurate text over several iterations
20:33 karolherbst: my issue is mostly about those 1:1 copies for MR descriptions, and even the "cleaned up" bits I've seen was rather on the bad end of the scale. Sure, we can try to formalize more what we expect there, but I also don't just ignore that many will put in less effort as I've seen there already. Some of the text there is just straight up misleading. And
20:33 karolherbst: that from developers I normally would trust to not do that if they write it themselves
20:34 karolherbst: and to be fair, commit messages do seem to look a lot better compared to the MR descriptions
20:35 karolherbst: but I also wouldn't mind adding a checklist what not to do when it's AI generated
20:39 mareko: I think the difference between good AI output and bad AI output is that the bad output can be after 1 iteration, while the good output is 10 iterations of AI interactions and 5 iterations of manual development and testing
20:39 karolherbst: mareko: and yeah, I'd also be fine with that in principle, because that goes into the "I have some ideas what it should write, just need the AI to write it down in a better language then I'd do myself", which.. sure, might not save that much time, but also sometimes writing up stuff is sometimes difficult
20:39 karolherbst: yeah....
20:40 karolherbst: I do want to emphasize I have mostly an issue with people 1:1 copying it and maybe only remove a bit, because it shows
20:40 mareko: absolutely
20:45 dwfreed: I work for a company doing technical support for a product where the customer's employees that open tickets are themselves generally supposed to be pretty technical people (in the case I'm about to mention, the person works for a company that helps run the internet)
20:46 dwfreed: Some customers have started sending us AI "summaries" far longer than text they would have written themselves about a "problem" they're having
20:47 dwfreed: I once spent 8 hours going through one such "summary", which required far more jumping back and forth on what was an almost 10 page document, only to determine the "problem" was 2 softwares (one ours, one not) working exactly as they were respectively designed
20:47 karolherbst: Those tools do make it very trivial to paste low quality text, where a human written text would be a lot better, even if worse in language quality, but higher quality in the actual content. So I'm really not available to discuss why that's not the fault of the tools anyway
20:48 karolherbst: maybe not "fault", but it's something that just comes alongside their usage
20:49 dwfreed: the only actual problem was the software not ours behaved in a way it really shouldn't have (not complying with a draft spec by trying to be smarter than it should be) and our software just did what it was told according to the draft spec (that we wrote)
20:50 dwfreed: The AI tried to claim that file mtime being less than birth or ctime was non-sensical
20:51 dwfreed: (meanwhile, 'rsync -a' does this for every file transfered)
20:51 karolherbst: yeah I mean.. I would trust mesa developers to sort those things out
20:52 karolherbst: it's mostly a thing from random contributors that they don't understand it
20:52 karolherbst: but I also really don't know how to formalize it in a way that it could work out in an Agents.md file
20:54 dwfreed: this particular case, the person is in fact highly technical, and his English skills, while probably not native, would have been totally fine (but I used to have to read the output of Google Translate on Mandarin from 2012, so I'm good at figuring out meaning when the literal translation is a bit broken)
20:55 dwfreed: "meat proxy" is the in-vogue insult people like to use to refer to such people
20:55 karolherbst: I mostly have an issue with claiming it's a personal thing regardless of the tools while research shows it absolutely isn't
20:56 karolherbst: and I've seen that from experience as well
20:57 karolherbst: anyway.. reading those 1:1 copies isn't fun at all, but I also have no great solution there anyway
20:57 dwfreed: yeah, if he'd written the ticket himself, I estimate it would have taken me at least 50% less, if not even better, time to parse his ticket and figure out the answer
20:57 dwfreed: and I wouldn't be outright exhausted by the end of it
20:58 karolherbst: mood
20:59 karolherbst: maybe it would be enough to remind people... or tell the AI to skip phrasing things like "I verified/checked", because no, the AI can't do those things...