Waddlesplash’s suggestion of an audio conference is a good one, and I think it deserves to be taken further.
What tends to work better is applying that format to each specific point of friction, one at a time, with a defined agenda.
If the audio conference format is genuinely on the table, I would suggest structuring it that way: one session per identified issue, starting with the most concrete ones.
Not an easy exercise to do but maybe better than a forum’s format where much energy is lost by everyone
I was pretty clear in the commit message that this was not sufficiently tested. It was merged anyways. In the fixup commit I also mentionned that I’m happy for the change to be reverted if it isn’t the right approach. But no one else looked into fixing that bug and I think it’s the last one blocking the beta 6 release. So maybe people were too happy to see an attempt to fix.
I think there is also a lower bar to entry for people who have commit access, or generally when there is confidence that if there is a problem, the person who sent the initial change will fix it. There is less risk that the code will be broken with no one knowing how to investigate it. Which is not the case for someone more “unknown”, in that case, we make really sure, as much as possible, that we fully understand the code before merging. Over time, trust in people who contribute regularly grows (because we see that they fix their own mistakes, they join the communication channels and discuss things, and we eventually have a more common vision of what should be done).
There has to be a balance here, because, on the other hand, we don’t want to be too strict with newcomers. Asking them for more and more changes and increasing the scope of their patches is a way to discourage them of ever getting something merged. And, even worse, if we let changes by current maintainers pass easily, while being annoying to newcomers, this creates a “double standard” effect. I think it’s not intentional, and explained by the previous paragraph. Yet, it is something we should seek to improve.
I also tried to push for a change of the way we handle this, basically saying that the formatting tool is always right. Then people could just run haiku-format before sending their patches, and there would never be any discussion about this in code reviews. Other people did not agree. So for the time being we will continue to waste time doing manually something that could easily be automated. And, indeed, the energy spent there is not spent on the actually useful parts of code review.
Also, I think I’m quite bad at code review, and I miss a lot of things because I trust the person writing the code to be right and having tested their changes. This is also one of the reasons that I don’t think I can do a good job working with LLMs (where it’s more reviewing and less writing code).
On the purely functional side of things, I hope KapiX’s ongoing work to restore the unit tests and run them on each commit will help catch some things
Ok, I think I’ll make one last exception to my rule of avoiding posting in such a thread. That’s because I see my words being altered on purpose.
@zuMi Do you really think that famous aphorism (attributed to Voltaire or whomever else) really fits here? I was talking about LLM lovers who have nothing to say but they say it continuously, 24/7, evangelizing their Holy LLMs. They overused their right for free speech turning it to a crusade. Someone has to stop them before they destroy this forum. Having your opinion is one thing. Evangelizing it 24/7 is another.
I had enough of it.
By the way, and speaking about French notorious people, I would mention Evariste Galois before Voltaire.
No, he uses an example to prove a point. Otherwise, comparing dogs to humans in general would be an insult for the dogs, because they are way wiser.
Oh come on. You know very well what I am talking about: LLM crusaders who won’t stop until we all bow to their sacred LLMs. Want names? Andrea is one of them. But Im sure you know what I am talking about. You are just exaggerating using hyperbole in an effort to prove me (and others) wrong.
And to answer your questions frankly, YES, I wish LLM crusaders to be banned, no matter who they are. Note the term “crusaders”. If that includes you, me, whomever, then so be it.
And YES, I want any software made with LLMs to be banned too, or at least clearly stamped and distinguished from the rest. As a user I have the right to know if software is made with such “tools”, so I can avoid it at all costs. That’s because I do not trust bullshit generators, a.k.a. your beloved LLMs. I tried them, trained them, and guess what, they are exactly what I just said — bullshit generators.
Complete, utter nonsense. I am not in any holy war. You are. I am not posting again and again for an eternity about LLMs here. You, and others, are. I am not using every single excuse to talk about LLMs. You and others do that. Really, who is in a holy war here?
And I am getting really, really bored of this.
Say what you will. I do not care anymore. As @waddlesplash already mentioned, people actually avoid posting because they had enough of this LLM drama. I think I’ll join them again.
I don’t see anything waddlesplash hasen’t adressed, so I doubt the senserity of your request. Unless you mean your acusation that I closed some thread in Order to not answer you on Telegram, to which I can only say: I do not have a telegram account, have no intention of getting one, and do not care or know what is discussed there.
I’m sorry, but this doesn’t seem to be happening. Yes, we could say that people who like LLMs started this controversy. But I wouldn’t say then that they are “evangelizing it 24/7”.
Both sides here in this controversy sometimes get heated.
Yes, Andrea was often very polarizing. I haven’t really been paying attention at the time, but I am pretty sure they often contributed to making many flamewars more heated. But there is similar behavior that I have seen from some who are against LLMs. Some people on both sides need to moderate their tone.
We could say there are two “crusades” happening simultaneously, one against LLMs at all cost and one for LLMs at all cost. But, fortunately, I think most people in this conversation are not that extreme.
Each side has a reasonable basis for believing what they believe. We have seen many good arguments from both sides. My opinion: The challenge is figuring out how to get along with each other reasonably well. And that involves compromise. People who like LLMs should label their software as LLM generated and people who dislike LLMs can then block that category. Fortunately, we already have rules in place to do this.
And there is … less than useful … rhetoric coming from partisans on both sides, too. And I’m not confident that it’ll settle down until there are reasonably clear policies set for each independent part of the community.
As I understand it, there are only 2 parts that even have policies so far:
OS dev: no LLM-generated code in Haiku itself; LLM-generated code in external dependencies is accepted
forum: no LLM-generated posts, except maybe in the context of automated language translation; discussion of LLM-generated software to be in the “Proprietary & Other” category
But I think we’ll see flareups of this fight until there are policies covering everything. So I would definitely encourage those of y’all who are decision makers for other parts of the community to be coming up with your own policies.
You doubt the senserity of my request because I think you get confused by this sea of people who also intervene off-topic
I have never accused you (singular, this is the limit of using English) of closing a thread to avoid answering me
I am giving you a brief summary of what happened, so you dont have to re-read the whole thread again
Summary
On Telegram, I pointed out that you (singular) had closed a thread by another user, abruptly redirecting the discussion to another medium. By symmetry, a user might not want to continue there: “I do not have an account there, I have no intention of getting one, and do not care or know what is discussed there” also applies to anyone, especially if OP opened a thread on this official medium.
The response to my observation on Telegram was given by another mod, who many times intervened on the Telegram channel via Matrix, and decided to answer me on that closed thread, instead of opening a new one here, elevating my doubt about your (plural) conduct as mods: on the one hand, you (pl) affirm that the use of LLMs is banned because it limits human partecipation, and on the other hand, you (pl) limit the users’ responses using a closed thread
You (plural) can close whatever you (plural) want here, you (plural) are the mods and you (plural) can do it, I respect authority when authority respects the rules. but if their conduct appears unmotivated, unclear, driven by bias or bordering persistence in pursuing [1], one can expect to see me asking explanations about that, again and again if necessary
BTW the reason given for hiding the post by @anonimous doesn’t hold up, doesn’t convince me, and I ain’t certainly the only one, given that…
but apparently another user can send to hell anyone who doesn’t think like them, and this is not considered offensive, which reinforces that “two weights and two measures” apply here
[1] I don’t know or find a perfectly matching translation for the Italian word “accanimento.” Some dictionaries say “fury,” “harassment,” or “doggedness,” but these don’t match; they are too strong or too bland. It’s more like the canine obstinacy of a dog that, in example, continues to bark and follow the postman at the fence, perhaps while a thief is at home.
This is a figurative way of saying something very present in Italian idioms, as in “accanimento terapeutico” (therapeutic obstinacy), when a terminal patient is kept alive with unnecessary treatments while it’s time to let them go, or “accanimento giudiziario” (judicial obstinacy), when the judiciary keeps pursuing a suspect while there is no evidence of a felony, aside from the preconceived notion that “they did it before, so it’s fosho they did it again”
Wow, quite the thread. I am new to the forums and after making my first post I was definitely a little surprised it got moved from native apps to Proprietary and Other. A native C++ app, that uses no outside dependencies/libraries does not stop becoming a native app just because AI helped with research, planning, implementation, and debugging. It’s just a tool, and poor code can be written by human or machine, that’s what the review process is for. Should I ignore all the programming errors that AI found in Haiku in it’s very limited scope of having the MidiKit2 and USBKit source, and say it’s better code because it wasn’t written by AI? Ultimately, the forum rules for lumping all AI assisted software into Proprietary and Other seems a bit silly to me, but I will respect it. I definitely know where to post the thread I wanted to post about my AI with Haiku workflow.
I think the categories here seem a little restrictive in general, and that only certain category combinations are allowed. If a post could be tagged with all relevant categories, like my other post could have been tagged Software, Native, and “AI Assisted” and all been correct. Users who don’t like LLMs can filter not to see LLM related content, and users who may be interested in AI use can easily find what they are looking for. Win-win. Just my 2 cents as someone new to the sensitivities around this subject within the community.
Not to drag this up again, but this is discussed severall times already. the TL;DR is that no, AI code can’t just be reviewed. If version 0.1 and 0.2 of your application do not share code because your LLM just generated new and different code that isn’t something a code review can acomodate, you are basically asking for an audit instead.
In any case, the forum category was introduced exactly for this purpose, so I don’t see why it beeing moved there would be confusing… all forum categories have a description that can be consulted when it is unclear where a topic belongs. It’s not that bad when you post in some wrong category, it will just be moved then. The usage of allowing users to filter out the Prop category here is certainly intended
IDK. I am pretty anal and review every step of the process, so I know it is not just rewriting full swaths of code without constant review, and they are usually small changes, unless we are fully rewriting something that became a dead end, but that is also similar to the flow of a natural development cycle. I mean I definitely wouldn’t want a vibe coder with no real programming or technical knowledge to submit commits to an open-source project. But I also wouldn’t want a human coder with the same knowledge base committing to the project either.
I certainly would want a coder with limited knowledge to submit patches. I did not know any C++ before working on Haiku code, but code review is the thing that helped and tought me. Sure sometimes the question will be “why didn’t you just do x instead?” But realistically without code review and starting from a limited understanding there is no path to get better at programming.
For vibe coding there is not much path forward, how will you learn about git and code maintenance requirements when “the ai will just fix it”?
As an aside your use of “we” when reffering to these technologies feels really creepy to me.
LLMs don’t merely make mistakes though, they also make up things (they are still text predictors) and actively try to cheat and cut corners (they don’t avoid success at all cost). Also, they tend to make really awful design decisions (a text calculator does not really think).
A typical review assumes good faith and does not work in this setting. An appropriate form of review would be a security audit.
It’s not much of a problem when you merely generate a stub or an example, but committing slop is incompatible with quality.
This said, sometimes quality is not the goal, and the result is reasonable for a throw-away script.
These are certainly good points, but please let’s not pretend we all are great designers and software engineers and cyber security experts, because we are not.
No offense meant to anyone involved in Haiku development (especially because I’ve been part of It for a long time), but
I don’t think the haiku code has ever been thoroughly checked for security issues, and there are many design issues, especially because there were in Beos, too.
Well, I care enough to learn some theorem proving, software modelling and verification, and ended up being one of contributors in Iris-Lean despite having no formal CS education.
I still have to commit anything into Haiku itself though. And I’m definitely not a security expert.
I’m not a big fan of segregating or isolating topics but honestly it was horrible before because so many negative comments from people. My first ever c++ app I posted was a terminal app. It was a fun project, did some neat things e.g., also compiles on Linux with no code changes. I managed to also get projectM to run o it. I had a lot of AI help because yeah, I’m not good at c++ at all. I mention it was AI assisted and that was nail in the coffin.
Comments like “Haiku is a GUI only operating system”, and it’s not native Haiku code. And soon my little humble thread was a flooded with users debating AI and not to AI.
Besides that, simple apps like a konsole addon. Man you’d think a c++ app that was 50 lines are less that was AI assisted would not cause a stir. Wrong! It was endless.
I’ve also managed to port rakarrack to Haiku which to me is so amazing. It’s the only guitar app to my knowledge and it’s really amazingly stable. But yeah, AI assisted. Boo hoo. Well, it should of been ported sooner!
But on a serious note, I am thankful for Haiku and the community, all in all it’s the world where in now and I think the admins and devs are doing a good job. Haiku will only get better from all this I think.
Yes, this category is named a bit, let’s say “unfortunately”. It really should be “AI assisted and AI created apps” or something along that line. A native Haiku app, even if it was created by or with the help of an LLM is still a native Haiku app. As far as I know there are no actively maintained proprietary apps for Haiku since TuneTracker moved away from Haiku.
I remember a thread here posted somewhat recently talking about how Haiku should have an AI kit. That was categorized outside of the “Proprietary and Other” category, which makes sense, since it wasn’t talking about an AI assisted app but rather an idea for a new kit. And yet, I think that should have gone to an AI category. But “Proprietary and Other” wasn’t a good fit for what the topic was talking about. So, yes, in my personal opinion, I think there is room for improvement on our AI category naming.
It’s well known that Google Gemini generates text where provided URLs are wrapped into a redirect URL via its Google Search service.
While I see no reason for an human, in particular the one human that perfectly knows what is the URL of his project on GitHub because it’s his project, to write such redirection URL, in particular in a post to present his project.
For Google Gemini to do that, au contraire, it makes sense in term of business (it will place ads and therefore gain money everytime this wrapped URL will be clicked by someone). Gemini knows how to generate money for Google…
I agree. Having an open source native app for Haiku being moved under “Proprietary and Other” category is both misleading and an abuse of the “Other” meaning.
Either we should have a Softwares / Native / LLM assisted subcategory, or we should enforce some tag/title convention for threads about LLM assisted native softwares, because clearly vibe coding won’t go anywhere and more and more people will be able to use LLM to build softwares running natively on Haiku.
