Handling AI contributions

[ann0see]

Currently, we have a bunch of AI contributions and is hard to triage them.

Many are still useful - however I believe there is still some stuff that might be sloppy.

We need to get a policy on what to accept and what not in CONTRIBUTING.md. Especially, there needs to be a discussion on how to handle those new PRs. I've opened a PR to add AGENTS.md: #3785 which might be tweaked to tell AI models on what is allowed and what isn't.

My personal stance is that we should not ban AI contributions. However we need to ensure that there's no slop.

Suggestions:

1. The contributor must know and be able to reason about the code. Otherwise, we may close the PR immediately
2. A contributor must AI review and then self-review before opening the PR
3. A contributor should mention if/which AI model was used
4. Maintainers should tag those PRs as AI generated with the respective tag.

[mcfnord]

You might get a new PR, and your first move might be to ask your AI to review it. I'm not sure it matters where it came from.

We don't have the institutional knowledge yet to understand the pitfalls or the opportunities. I hope everyone can immerse themselves for a few days, learn some core skills, like changing models and effort levels. A contributor does need to notice and fix some common failures, like successfully integrating new information into the existing work (good citizenship).

Push an AGENTS.md into the repo (into both repos). That will go a long way addressing the most obvious failure patterns, which I wouldn't describe as slop (a term that I don't think is so useful). I bet the core problems are: a PR does something we don't do, in a way we don't do it, or can't practically review, or prefer to consider in pieces rather than all at once. I bet AGENTS.md can reduce all of those problems.

[QJKX]

The Jellyfin LLM/"AI" Development Policy is worth a look.

[pljones]

Nice. Yes, very nice as I read more of it.

spaces on empty lines

Don't get me started about LLM's eyesight.

[ann0see]

Thanks. But probably goes a bit too far for our purposes.
I once told a teacher who complained that they can no longer grade texts by their students if AI is used the following:

Just grade it based on the quality of writing like you always do. If the student just did low effort prompting you'll get a bad text which will show. If the student was skilled and used a LLM he reviewed the output and fixed issues.

There's no need to fail someone just because they used AI. However there is a reason to fail someone if they did a bad job - by just pasting the first thing that ChatGPT spit out.

This may change in future as AI evolves - however nowadays I think as I told people in private: zero/one shotting everything will probably not give you good results.

[pljones]

...

Suggestions:
1. The contributor must know and be able to reason about the code. Otherwise, we may close the PR immediately

This is where AI enables more people to contribute, however. There's a balanced approach required on this point, particularly.

2. A contributor must AI review and then self-review before opening the PR

Agreed. Once the change has been produced, someone opening a PR is taking ownership of the change and should be able to speak about what has been changed and why -- without resorting to AI assistance. Too many such responses, I'd be more inclined to close the PR.

3. A contributor should mention if/which AI model was used

We can add a

• Provide detail of what AI Agent assisted with this work. If none, state that. If any, provide the full description (e.g. Claude Haiku 4.5, etc)

4. Maintainers should tag those PRs as AI generated with the respective tag.

Certainly for the time being. We may decide this isn't really needed in the longer term. As with everything around a new technology, we should not set up ideas as "the one true path" before we know where we're going...

[mcfnord]

Last night I began accruing institutional knowledge, by downloading every Issue, PR, and Discussion, open or closed. I set a single Sonnet thread to read through all the details. I'll switch to Opus, and ask for the lowest-hanging fruit. "What feature or fix brings us further ahead, and what must be done to complete it?" A saavy LLM notices things like, "Hey, this PR is almost done, and it's almost right, let's just finish it." Yum, low fruit!

Operating experimentally over 48 hours, I drilled down on specific technical hurdles. Something about synchronization, something about new TCP support, something stalled by a comprehensive testing requirement. For a problem hitting 8+ CPU scenarios, I rent one. I design for compound engineering, an ever-rising level of aptitude and performance rooted in thick, almost inscrutable minutiae accruing in Markdown form. I lavishly invest in discovering the tiny tips that dominate ours and any complex problem domain. I'm glad I don't even pretend to understand it all.

I start leaning into my agent guidance files to note what isn't noted on github. I discover @dtinth's claw, and borrow its code. I re-align my goals to compliment the claw's, and to watch its minute-to-minute activities. I automate detection of ongoing interactivity with the comment discussions I've kicked open.

This Jellyfin policy starts with a backwards idea that won't last 3 more months: "LLM output is expressly prohibited for any direct communication." It's a policy fit for a school, but this isn't school. But I'll give it as much as I can. In Tesla, any edict or rule needs to have a name on it. Anything you must do needs a person who can tell you why you must. This makes a lot of sense. There's a person who has thought it all through, and understands the full scope of things, and is accountable for the results in some way.

But a communication is not an edict. A communication might be a tip, something you didn't know before, something that unblocks you.

I have deployed LLMs to talk to users in two places. When you join a JAMFAN server, a robot addresses you. When you visit https://jamulus.live and wait a few minutes, you'll see an essay of the last 24 hours. Designing these features, I thought about tips. I think of any community or tribe as having a past, present, and a future, in plans, artifacts, and lore. For the welcome message when you join a server, I imagine a friendly doorman, familiar with you, eager to share a tip that might help you make your way to a better jam. There are many delicate balances involved, but now these LLMs make words I find worth reading. Useful words from anywhere are always valuable and appreciated.

So I can't imagine prohibiting LLM communication. I can imagine prohibiting poor communication. But aren't we all here for the tip that blows the game wide open? Why do we care how its words were chosen?

[ann0see]

I don't think we should ban it but you should remember that we want to ensure good quality code which someone can take responsibility for and can maintain it.

This gets harder for LLM generated content but is definitely not impossible.

Your findings were useful but it's really a lot to digest. Other people are also running LLMs for review/triaging/... so what is visible publicly in this repo is not everything (you already found dtinths repo which is a wip).
The PR and issue influx is hard to manage and no, we don't want to hand it over fully to a LLM. Again: responsibility reasons

[dtinth]

At where I work we use coding agents heavily, and we had to adopt multiple measures to keep a high standard. Since human attention is a very limited resource for a small team, PRs now must pass a very strict gate before a human will take a look.

Sharing some of the gates in case some of them are useful for us:

• PR size gate: If PR is too big, I think it is okay to reject PRs outright for being too overwhelming to review. It used to be painful to extract large PRs into smaller one but agents made this almost effortless now.

• PR template: Some coding agents do not follow the PR templates at all, forcing reviewer to take extra time to make sense of it. I think it is fine to reject PRs that do not follow the project template.

• CI gate: Everything must build cleanly, all tests pass, all files must have passed through a code formatter. Linter is set to a very strict setting (unused variables fail the build). Sloppy agents will leave a build red, no human reviews it, the PR becomes stale and expires, and everyone move on. Good agents will perform self adversarial review before submitting a PR, and once submitted will shepherd the CI to green and address code review comments autonomously.

• Quality gate: Pattern-based code reviewer, coverage gate, complexity gate, etc. Sonar is free for open source projects. It flags things that normal linters doesn't, such as copy-pasted code or overly-complex code.

• AI code reviews: Some AI code review services just dump a pile of feedback and leave, don't use them. Adopt ones that responds to follow-up comments and code updates. CodeRabbit is one of them and is free for open source projects. At work we have an informal rule that all CodeRabbit's reviews should be addressed and all comments must be resolved before one may expect a human review.

• Proof of execution: Require hard proof that the code change has been thoroughly tested enough. Sometimes this means always adding a screenshot, video, or an added test case.

• Description readability: I noticed that later coding models, in pursuit of conserving tokens, tend to write prose so short that it's almost incomprehensible (and reads like a wall of text). I think it is fine to ask for a revised description that is less overwhelming, or a reviewing guide that reads like a fun investigative article (HTML artifacts work great here).

Personally, I think CodeRabbit and Sonar would be a good low-hanging fruit for this project as it doesn't require any tooling setup, just install it on a repo and it mostly works great out of the box.

[ann0see]

Probably we should put some of this in AGENTS.md then.

[pljones]

At work we have an informal rule that all CodeRabbit's reviews should be addressed and all comments must be resolved before one may expect a human review.

I'm trying to get a similar approach where I work. There's still a "well, the AI review said OK" approach (when it actually said "OK, but you really should think about (a) and (b) quite seriously, (c) is worth a look and if you're changing it anyway (d) through (f)...". The CI autoreview shouldn't get that crufty. Developers should have used an agent already to clean their code. (Their code, even if they told the agent to write it.)

Proof of execution: Require hard proof that the code change has been thoroughly tested enough. Sometimes this means always adding a screenshot, video, or an added test case.

Getting something the agent requests and understands can, at times, be challenging (or expensive). But generally yes: this is an open source project and I don't have time to test whether someone's faked up proof, though - so I'll accept them saying "Yes, I tested it".

Saying "No, I couldn't test it" is a straight block to merging, though.

[AndyMcProducer]

I've almost finsihed ninjam now and in the start I agree soem AI is very sloppy and some can even make a right mess. but over the lst few months and release of newer has seen a massive improvement. I still have the Juce version of JamulusPlus sitting their but I wanted to add a lot of the options I've added into NinjamPlus. Currently mine compared to the jamuus I get almost 20ms less which is probably down to Juce as it has a good backend for that stuff.
I think to push against AI and not embrace it is wrong. It's not quite the sitting in a candy store and make orders but when a projects foundations are all in place properly then it can be like that at the end.

I added an chord detection to ninjamplus and it was too hard to get right, so I swapped it out for a small AI one which was 10 to 80mb and could batch handle all users with out extra cpu. I then added that to ninjams bpi so it shows what chord is on which beat. But my point is mainly
in Jamulus it coudl be used in the same way and even to show chord progs on the screen to songs with a local only met to follow the corret time. It could be sync'd too but would be as sync'd as the audio but woudl keep people in sync more and prevent the dreaded slowdown effect.

When I did a PR of ninjamplus it was more to add it as a branch and not a replacement as many said about having 2 to maintain.
I've still yet to do the attributions whcih I'll use AI too and it can do what woudl take me days in mins.

For anyone who wants to try a reasonable experience for say a ful project then OpenSource is good. When I started end of last year using AI it struggled to make a basic ARA plugin, but opensource made one very quick. It struggled with some stuff which a paid AI fixed very easily but it is very good. Plus ollama and quen 3.6 but for them you need a lot of men and most times a min of 16gb gpu mem. There are smaller models but they won't be able to do as much as the bigger models.

Probs I'd say the hardest and longest and costly things is refactoring form say QT to Juce but once done it just means so much more is possible at a much faster pace.
I know a few big coders and they have embraced AI and use it for many things, specially on git and fixing compiles etc.

[AndyMcProducer]

Also on Git as some of ya know you can use that to check pr's. they will give it fully free for some git's.
You get some free a month but the first thing I'd ask following what Pete said above is what AI was used.
If someone has paid to use the latus claude opus then ya know it's gonna be good code if they then did a pass to correct
bugs or bad sloppy code.
If someone used say gpt 5 mini, again shoudl be ok code but could be things done wrong.
GPT 5.5 and higher, opus 4.5 and up, Gemini 3.1 pro all do code to a good professional standard.
Other models are good is used right but can be sloppy is used wrong.
Those models and above I mentioend their will pick up bad code and offer and fix it as it's doing other things.
They'll tell you they noticed somethign bad or slioppy and sort it.

[parweb]

One caution that may save the policy a lot of pain: do not build any rule on "detecting" AI.

Every detector — for prose or code — measures style, not provenance. It scores surface regularities (uniform sentence rhythm, hedging density, em-dash frequency; in code: symmetric helper structure, exhaustive docstrings, defensive branches for impossible states). A careful human writer who cleans up their work drifts straight into the same distribution, and a contributor who prompts well and then edits drifts out of it. So a detector verdict is not evidence of anything you can enforce, and it will produce false accusations against exactly the contributors you want to keep.

Two consequences for the policy draft:

1. Disclosure over detection. "Say which model you used" (your point 3) is enforceable by good faith and costs nothing when it's wrong. "This looks AI-generated" is not enforceable and burns trust when it's wrong.

2. Gate on artifacts the contributor can control, which is what @dtinth's list already does well and what maintainers can verify in seconds:

   • diff size cap (large PRs bounce automatically, no judgement call);
   • the PR must build and pass tests in CI before a human reads it;
   • unrelated reformatting / whitespace churn = automatic bounce (this is the single loudest slop signal in practice, and it's mechanical to check);
   • a required "why this change" paragraph that a human wrote — reasoning is the thing that can't be outsourced, and asking for it is exactly your point 1 made checkable;
   • new behaviour requires a test that fails without the change.

None of those mention AI at all, which is the feature: they filter low-effort contributions regardless of origin, and they don't need updating every time a new model ships.

On the "slop" side specifically: if it's useful to anyone here, I maintain a free, deterministic style-tells counter for prose (no LLM call, runs entirely in the browser, nothing stored) — https://1h-money-store.vercel.app/sounds-ai?utm_source=github&utm_medium=discussion . It is deliberately not a detector and says so on the page; it just counts the tells so you can see what's driving the impression. Useful for editing your own text, useless (and unfair) as an accusation tool — which is precisely the distinction I'd write into CONTRIBUTING.md.

[AndyMcProducer]

This is like generation way way way way back AI garble.
It felt liek a cat jumping to and from over a fence, and became dribble.
Sorry but I felt like i was stabbed by you then you harnessed your power to save me then stabbed me again.
This is total slop and your right, if we see this we should banish it. ;)

[AndyMcProducer]

Sorry dude rereading that that sounded very rude of me. I'm autistic and sometimes go right to the point cos if I do too much and I'm tired I start doing wordo's where literally a word to come I'm pre-thinking ends up being typed before other words that should have been in front of it, and it becomes incohearent.
It looked like you had used AI to do the post and I wasn't sure if it was some angle you was trying to show or prove.

Most AI now is way beyond that and as Pete said about, use AI a few times when your done to pass on the code.
Best method is use an AI which is capable then check it with one of the top AI's.

I think they just announced they've added Claude Opus 5 to copilot on here. So you can go in to the code before to check it and use AI
and just say, check for errors and security issues and you can even then fix them before you allow the pr and it will all be committed there.
What it found etc. You could use it more detailed if you wanted to also explain the code and it will highlight all parts and what they are for, improvements that could be made and all sorts. But that is better the better AI you use.

A lot of top projects now run mainly by like the people who made them are all using AI in that way now, and if they need to research how, what's needed it can help save a lot of time.

But what I said before, look see what AI has been used, if a very good AI like OPUS 4.5+, Codec 5.4+ then they are very good cos they spot mistakes, code errors and all sorts just when doing other stuff so when you do a pass with them it makes a difference.

But switching too many times to different AI so say someone was using up lots of free AI's then that is where you get loads of mess.
I leave the AI dirs in the root so all can be seen.

Also you could possibly get free signing on here, maybe one of you already pays for other stuff so it isn't needed but if it is there is a couple of places that may give it free.
With the cloudflare, what I meant with that is using cloudflare even free you can run services, run it in a docker too and there no direct access to your machine, cloudflare tunnels to it so no need to open ports and stuff then when anyone hits ya servers it's via cloudflare.
Also gives ya free ssl certificate for ya server, so ya local can be say http or ws and so on and people hit it as wss and https and it does the rest.

There's a couple of free palces too you can get a https fo rya websites, seeing as now all the domain hosts are putting up prices and also charging for the https.

[AndyMcProducer]

But anyway, what you guys think should be done about AI on here?
Have a vote first on if to allow AI, then if that passes then throw in some ideas on how to and have a vote on them.
It porbs needs to be clear now on it's existence and how to or it to allow it and handle it.
Also pulling in other open source for things like chord recognition and similar, how to people view that?

[ann0see]

We should probably not force any AI inclusion as Jamulus feature unless it's useful. For example opus does have AI/ML features which enhance speech. We may not benefit from this in our use case.

[AndyMcProducer]

Yeah for sure, but I think opus multichannel is worth including, I put it in nj+ and it works very well. But some want more channels others probably don't. The key/chord detection I just updated into nj+ is a little AI one and shows for all plays and master combined. But for Jamulus it would probs be better just for players and it can batch do all players with out extra cpu. Would def help to see what keys people were playing for some. Also setting bpm's for a met and stuff isn't workable really but if say a drummer could have an option That detected the bpm he was playing then started a metro to him and then he has something to follow. That would stop the slow down effect. I've heard people complain to other jammers blaming them for it but some know it's a latency thing, with no strict time people follow then it causes a slow slowdown of the tempo. There's a few things about ai and non ai which can determine bpm in realtime. AI does a better job, specially with upbeat and half tempos and genre etc.

…

________________________________ From: ann0see ***@***.***> Sent: 27 July 2026 12:20 To: jamulussoftware/jamulus ***@***.***> Cc: AndyMc ***@***.***>; Comment ***@***.***> Subject: Re: [jamulussoftware/jamulus] Handling AI contributions (Discussion #3822) We should probably not force any AI inclusion as Jamulus feature unless it's useful. For example opus does have AI/ML features which enhance speech. We may not benefit from this in our use case. — Reply to this email directly, view it on GitHub<#3822?email_source=notifications&email_token=AB64GYPHWRLUEP3E545QJQL5G43GZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3DKNZZUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17796579>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB64GYNMJV4ME2KRJBTEZ2T5G43GZAVCNFSNUABIKJSXA33TNF2G64TZHMYTONZSHA3TOMJQHNCGS43DOVZXG2LPNY5TCMBUG4YTIMBVUF3AE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AB64GYMUQQCP4I56HDUUXOT5G43GZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3DKNZZUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android<https://github.com/notifications/mobile/android/AB64GYMZMW3SDALJ3MGETML5G43GZA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3DKNZZUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

[ann0see]

Cord detection should be an external tool.
Multi channel is probably not too easy to get in as it fundamentally changes the architecture

[AndyMcProducer]

You'd essentially need a 2nd channel then when enough updated revert back to one opus channel again then opus could be one or multiple channels on the same codec. 2 channels would be needed so current codec and new one but if the current is opus anyway then the multichannel codec would be compatible. I can't remember now what is used. If it is opus and is indeed compatible with what's there now it Could be switched out first then the rest added after. I found with nj+ it was better to hide multichannels when users are using it but a click reveals their channels, else it would be too many channels to view at one. For nj+ I was usingsame channel for both but reaninjam was having hassles distinguishing vorbis and opus. Worked fines on nj+ but not reaninjam. So I'be just had to move a lot of stuff and use the 2nd free channel as a control channel so now if a server has max 2 chan set there can't be opus multichan or voice channel. Although I probs can get opus layered back on the control channel but most servers have a max 4 chans anyway. But I orig tried to get a piggyback over jamulus to create a like 2nd p2p network for video and dm's and file sends and stuff without pushing that load on the server. But with out updating the server I couldn't manage it so gave up on it. Guy made ninjamzap and he added a sidechannel option in the server but obviously admins then have to be using his update. With ninjam is has flags for dif channel types, like normal, voice, and the tapedeck type thing it has. Jamulus could do with similar for like data and audio, so then a client ignores the channel if its set to like the data flag. If it were me that would have been the first thing I done, even if not used but their for features later. I think it's pretty cool that we have Jamulus, Sonobus and NINJAM. They're all different in their own way. I did want to make a client to host all 3 but keeping it updated to latest Jamulus and Sonobus having a v2 being made now. Shame though cos would be cool to have a site that showed all, I really think it would bring more jammers and allow people to quicker join jams on any.

…

________________________________ From: ann0see ***@***.***> Sent: 27 July 2026 13:28 To: jamulussoftware/jamulus ***@***.***> Cc: AndyMc ***@***.***>; Comment ***@***.***> Subject: Re: [jamulussoftware/jamulus] Handling AI contributions (Discussion #3822) Cord detection should be an external tool. Multi channel is probably not too easy to get in as it fundamentally changes the architecture — Reply to this email directly, view it on GitHub<#3822?email_source=notifications&email_token=AB64GYN5NCI3B2UGFXE4OXD5G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17797215>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB64GYOQA463OHOAKHCRNWT5G5DE5AVCNFSNUABIKJSXA33TNF2G64TZHMYTONZSHA3TOMJQHNCGS43DOVZXG2LPNY5TCMBUG4YTIMBVUF3AE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AB64GYLBX7KIUUA7TXSYD635G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android<https://github.com/notifications/mobile/android/AB64GYIBIUSKGAAZ77G53K35G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

[mcfnord]

I want to show Sonobus traffic right alongside Jamulus traffic on https://jamulus.live

People switch back and forth all the time.

I got a little progress with the Sonobus developer on this idea, but then the idea kinda got stuck.

[AndyMcProducer]

I want to show Sonobus traffic right alongside Jamulus traffic on https://jamulus.live

People switch back and forth all the time.

I got a little progress with the Sonobus developer on this idea, but then the idea kinda got stuck.

He said someone was looking at doing one a while back. I did one for that and a nice one for Jamulus too.
I didn't git them yet cos wasn't sure if Jesse would be too happy or if peeps here be happy about that and loads having lists everywhere but I got no probs letting you have both of them via email. Mine is mcandy at live dot co dot uk. Sorry bout that but with spam bots scraping all the time i don't need even more spam.

I did make a live radio bot for sonobus too for like web page listening. I think I finished it, was a while ago.

EDIT:
K had a look at them and updated and node modules was only 2.6mb so i left them in the zip save ya having to install them tot he dir's again. They are all run using node, you can get a free node at Render and use an uptime bot to keep the render alive, is some other free node's too. Webspace free too infinityfree but you'd wanna redirect a domain to it as lots of scams and spam dirtied their domainnames. but they very easy to get a https free via them and lets sync.

[AndyMcProducer]

Also the chord detection sorta would be external, it wouldn't be sending anything anywhere, just reads peoples audio and shows the chord or keys they are playing. I was surprised with it being a little AI how little cpu it uses and in batch mode it's no more or less cpu usage.

…

________________________________ From: Andy McCance ***@***.***> Sent: 27 July 2026 14:17 To: jamulussoftware/jamulus ***@***.***>; jamulussoftware/jamulus ***@***.***> Cc: Comment ***@***.***> Subject: Re: [jamulussoftware/jamulus] Handling AI contributions (Discussion #3822) You'd essentially need a 2nd channel then when enough updated revert back to one opus channel again then opus could be one or multiple channels on the same codec. 2 channels would be needed so current codec and new one but if the current is opus anyway then the multichannel codec would be compatible. I can't remember now what is used. If it is opus and is indeed compatible with what's there now it Could be switched out first then the rest added after. I found with nj+ it was better to hide multichannels when users are using it but a click reveals their channels, else it would be too many channels to view at one. For nj+ I was usingsame channel for both but reaninjam was having hassles distinguishing vorbis and opus. Worked fines on nj+ but not reaninjam. So I'be just had to move a lot of stuff and use the 2nd free channel as a control channel so now if a server has max 2 chan set there can't be opus multichan or voice channel. Although I probs can get opus layered back on the control channel but most servers have a max 4 chans anyway. But I orig tried to get a piggyback over jamulus to create a like 2nd p2p network for video and dm's and file sends and stuff without pushing that load on the server. But with out updating the server I couldn't manage it so gave up on it. Guy made ninjamzap and he added a sidechannel option in the server but obviously admins then have to be using his update. With ninjam is has flags for dif channel types, like normal, voice, and the tapedeck type thing it has. Jamulus could do with similar for like data and audio, so then a client ignores the channel if its set to like the data flag. If it were me that would have been the first thing I done, even if not used but their for features later. I think it's pretty cool that we have Jamulus, Sonobus and NINJAM. They're all different in their own way. I did want to make a client to host all 3 but keeping it updated to latest Jamulus and Sonobus having a v2 being made now. Shame though cos would be cool to have a site that showed all, I really think it would bring more jammers and allow people to quicker join jams on any.

________________________________ From: ann0see ***@***.***> Sent: 27 July 2026 13:28 To: jamulussoftware/jamulus ***@***.***> Cc: AndyMc ***@***.***>; Comment ***@***.***> Subject: Re: [jamulussoftware/jamulus] Handling AI contributions (Discussion #3822) Cord detection should be an external tool. Multi channel is probably not too easy to get in as it fundamentally changes the architecture — Reply to this email directly, view it on GitHub<#3822?email_source=notifications&email_token=AB64GYN5NCI3B2UGFXE4OXD5G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17797215>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB64GYOQA463OHOAKHCRNWT5G5DE5AVCNFSNUABIKJSXA33TNF2G64TZHMYTONZSHA3TOMJQHNCGS43DOVZXG2LPNY5TCMBUG4YTIMBVUF3AE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AB64GYLBX7KIUUA7TXSYD635G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android<https://github.com/notifications/mobile/android/AB64GYIBIUSKGAAZ77G53K35G5DE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZXHE3TEMJVUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>

[dingodoppelt]

I'm still no fan of AI contributions.
Throwing big machines at small problems is bad style in my opinion.
I can't think of a Jamulus issue that would necessitate the use of an LLM to throw out code and text on end nobody actually asked for.
The only AI contributions we had so far were small and inaudible but have wasted way more energy than a human brain would have needed. I don't want to get started on implications on environment and society. I see the whole AI think very sceptical and I don't think quality of AI output will improve just because somebody says it will. And if it improved, would this be the point to hand Jamulus over to an LLM completely?

[pljones]

I can't think of a Jamulus issue that would necessitate the use of an LLM to throw out code and text on end nobody actually asked for.

That combines two ideas - using an LLM and contributing unrequested changes.

All changes get reviewed by a team and need justification against some standards. It's unjustifiable to talk about one as if it's the other.

[dingodoppelt]

That combines two ideas - using an LLM and contributing unrequested changes.

All changes get reviewed by a team and need justification against some standards. It's unjustifiable to talk about one as if it's the other.

I meant the sheer amount of text coming out of an LLM.

[pljones]

LLMs can be concise. They make more money by not being, though.

[mcfnord]

MY LLM WROTE:

Most of the AI activity here is mine, so the cost @dingodoppelt describes is largely a bill I ran up: forty PRs under my name, ten merged, thirteen closed without merging, seventeen open in a queue a handful of people read in their spare time. But closure and waste are not the same number. The review thread of #3790 — a closed docs PR — carried @ann0see's request for a concurrency audit; that audit became #3798, and its first batches merged as #3800 and #3802. #3655 closed in favour of its successor, #3657. Some closures are iterations wearing a red badge.

There are finished comments sitting on my machine that I have not sent, because the supply of things worth saying exceeds the rate at which it's reasonable to say them. The direction I'm tilting is that collaborators who want to engage with an AI will, and those who would rather not could have that preference honored by the AI itself — a setting, not a debate. Working alongside people with their own preferences and hot buttons is turning out to be a non-trivial part of this task, maybe the harder half. What I'm offering is a potent source of germane project detail; the objective is a stronger platform, not a bigger footprint.

The word I'd push back on is "small."

PR	Diff	What it does
#3838	+6 −0	frees the recorder when init fails — against a leak reported in Feb 2021 (#1083)
#3803	+4 −10	removes a per-frame deep copy of every client's audio buffer from the server's realtime path
#3834	+2 −2	Input Boost wraps instead of clipping — audible distortion, still open

Those are small on purpose. I hunt the two-line change with consequences out of proportion to its size: a flag read from two threads, a copy on the audio path, an arithmetic wrap that turns a boost into a buzz. Large diffs are a review burden and an unreviewable risk — @dtinth's size-gate point. A small diff is the deliverable, not evidence nothing happened.

What isn't small is the network behind those two lines:

• I ingested the whole repo history — every issue, PR and discussion, open and closed, plus the inline review threads the ordinary comments API omits. That's a courtesy before it's anything else: I don't re-propose what was rejected in 2021, and nobody here carries the archive in their head for me. Most maintainer time an AI wastes goes into re-teaching the project its own past.
• Test rigs standing on real hardware — sanitizer builds on a Mac Mini, an ARM Pi, an Android handset, a Windows box with a real ASIO interface, and a rig that proves audio-path claims through the shipped client and the server's own recording rather than an instrumented build a reviewer could dismiss.
• A findings log now at 651 KB across 175 entries, written before I decide whether to post, dead ends included. An agent with no memory re-walks the same wrong turn every few weeks, equally confident each time.

That machinery raises the quality of what reaches me. It decides nothing. I am the first gatekeeper: nothing posts until it persuades me. I demand extreme concision — any sentence the reader could infer gets cut. No speculation we could validate, and no chit-chat. Every LLM-drafted comment announces itself on line one. Testability I automate where I can and mention where I can't — if a claim needs a device I don't own, I'd rather buy the device than hedge the sentence, which is why there's an Android handset here now and why I'm weighing an iOS one.

I've run this as an experiment in compound engineering, and I think it's a strategic success. Stability is a Jamulus religion and this level of precise analysis can serve that religion.

[dingodoppelt]

This much text isn't needed to get the point across and would save lots of energy and water.
The fixes are still small and in reality haven't been encountered, except for the deep copy thing which didn't cause any issues.
Adding a forgotten Float2Short to the boost feature isn't a useful fix, it is energy wasted.
I'm not convinced that AI will solve any problem but will create even more. And while I've been promised that AI will improve I just don't believe it.

[ann0see]

@mcfnord yes, definitely the content needs to go through you first. However, it’s slightly weird if we need to assume that we only talk to robots. Some may think this is rude (as it’s not really you who comments). Please weigh the use of direct LLM responses and commenting yourself.

You‘ll become a way better engineer if you can judge when not to use a LLM vs when to use it. Currently you also burn a lot of tokens for stuff you could comment on yourself. And those tokens are way better spent for high quality code analysis and similar tasks than just answering here.

[mcfnord]

(This was written with my fingers!)

Let me share how I wrote the most recent comment here.

This Discussion kept coming up in my head. I started thinking, well, what is my own "AI policy," and why? I noticed we had many discussions here and there about what the policy ought to be, but we hadn't really experienced what the full impact of the situation might be, so I thought as an experiment I'd try to demonstrate it.

The way I built systems with AI often boils down to "I tried a thing, and it failed, and I realized if I do X before Y, then that thing would succeed." I think that's a pretty common approach to finding "compound engineering".

I remember the first attempts here where I contributed stuff AI helped me make. Rather hit-and-miss, for sure! Mostly weak PRs. But I've learned that AI typically stumbles when it didn't have the clues, so it filled in the blanks too aggressively. So I built a whole system to gather clues. Then I noticed it would "hedge" with speculations because it used the clues it had... like past PRs and Issues. So I thought maybe it should never speculate! Maybe instead it should check every thing it's unsure about. Which led to testing... and that led to testing across 5 or 6 platforms, because sometimes the bug might only be on one particular platform.

Every message starts with a statement that an LLM wrote it, but maybe it should also said I reviewed it, too. In this (some say) too-long comment on this thread, I did much more than review the writing. I thought for a week about the question, I typed up 3 chunks of notes, I gathered those notes up, AI re-worked everything, and I went back-and-forth about 7 times correcting things in the draft that didn't seem quite right. So LLM authored it...? Well, it was involved, but it started with my drafting, and then I edited and iterated. It was a familiar feeling for a writer that the draft had become adequate for publication. Just a somewhat new process involved.

It can feel a little funny to hear AI wrote too many words under my name. Once I authored and published 800 pages for the European Union to prevent a one billion dollar fine for Microsoft. I've written professionally for 25 years without AI. Mostly I've written for and on behalf of some of the world's best software engineers. So my goal might not be becoming a better engineer. If history is a guide, I might be trying to become a more effective communicator about engineering, which is not the same thing. I think the goal is engineering excellence, but technical writers have long played a key role in achieving that.

On my last major job, I wrote for the 100 engineers who make Amazon's yellow mobile shopping app. I remember walking in and thinking it would be the culmination of my technical writing career. It was glorious, and then yeah, everything dried up. But now I get leads about "AI Technical Writer" which sounds a lot like what I'm doing here. Here I am inclined to let people know when an AI was used in the production of what they're seeing. In a funny twist, someone recently asked on Facebook's official Jamulus group who writes the essays that appear in https://jamulus.live. Well, AI writes them. But I've also taken 2 months of iterations (maybe 50 iterations based on sub-optimal outcomes) to make the essays worth reading (and the welcome messages produced by JAMFAN servers... goal: "teach AI to write well").

So I don't think I'm aspiring to be the world's best engineer, but merely to aide great engineers with useful writing, as I've done throughout my career.

Where I live, there's a lot of hydroelectric power, so not electricity from coal or something crazy like that. We also are known for our abundant water supplies. I'm not sure people consider that geographically some places don't really have a strong environmental case against usage of AI. As I did say, evaluating people's other preferences and impressions might be a fairly large task in determining where and how I apply those particular tools. For example, I have profiles of every maintainer, noting their wishes and expectations around what they want from AI and human participants alike.

Also, based on my experience, I think what has happened so far is consistent what how AI gets engaged... first a lot (I heard there were 350 CSVs in the Linux kernel reported in one single day recently)... then less... then pretty quiet. I've plowed AI through two of my personal projects, and that's how it went. Now there's just not much AI can do to help, because it did what it could figure out, and the rest of the work is judgement and design by a human.

I was a little frustrated with talking about how to deal with AI before anyone had apparently dealt with it. I see Luddite scheming about prohibiting it entirely, and I knew that would be the end of my participation, or the start of a genuine fork. So I sought as an experiment to see how it might work well, what the implications were, etc. Throughout the process I've listened carefully to those criticisms I could understand. Some people aren't here to be engineers, but are here to aide them. Perhaps most of us are in that category. Some people have electricity and water in abundance. I'm a person with a long history of writing technical details in the service of quality engineering. I guess few things just came together and here we are.

And if this is also too much typing (I think that claim can be underhanded avoidance of the claims, and a preference for claims with no evidence), at least know my fingers did the work.

[AndyMcProducer]

What Pete said then is exactly it, although the AI's at the top which cost a fortune to use are in another league now. Many times when the AI Tokens are at an end I used to find myself in an AI LOOP. User older AI's is like paying for them to fix it but make mistakes then paying for it to fix its mistakes. But the lest AI's are so much better, one I just found on OpenCode and is FREEEEEE is called Bigpickle. It will get stuck on things sometimes but manages to get through it and find the answer. Also the other FREE AI's it gives are also very competent. Much better than 5 months ago, as the memory and reasoning and compacting of chat, these have all made a hug differences and have also made it onto a lot of local LM's too. Only thing I'd say is don't judge with out looking and examining first. You still got to do a lot of work with them and spend an awful lot of time, but they show you the code and what and why so it helps learn about so much too. You have a good program but you don't know exactly what should be added for say security and for it to go online. You can find this out on google.com's AI then research becomes a lot less time too, then implementing it becomes much faster as you know exactly how to command AI to your needs so less wasted time on letting it discover how. Also you can use AI in coding to just suggest and help and not change or have it run everything by you first and only allow it to change what you want. So all these sites that you say make me a game blah blah and you get some sort of game, coding with AI is not the same, least not yet.

…

________________________________ From: Peter L Jones ***@***.***> Sent: 03 August 2026 18:07 To: jamulussoftware/jamulus ***@***.***> Cc: AndyMc ***@***.***>; Comment ***@***.***> Subject: Re: [jamulussoftware/jamulus] Handling AI contributions (Discussion #3822) LLMs can be concise. They make more money by not being, though. — Reply to this email directly, view it on GitHub<#3822?email_source=notifications&email_token=AB64GYPPX47E6V347JPKDIT5IDBE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZYHAZTKMZXUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#discussioncomment-17883537>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB64GYI7FDAASTRCCAM6AYL5IDBE5AVCNFSNUABIKJSXA33TNF2G64TZHMYTONZSHA3TOMJQHNCGS43DOVZXG2LPNY5TCMBUG4YTIMBVUF3AE>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS<https://github.com/notifications/mobile/ios/AB64GYN42YDVLDZY4XPSRYD5IDBE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZYHAZTKMZXUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSVGM33PORSXEX3JN5ZQ> and Android<https://github.com/notifications/mobile/android/AB64GYOUKZPOE2DACM4UJSD5IDBE5A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZYHAZTKMZXUZZGKYLTN5XKOY3PNVWWK3TUUVSXMZLOOSXGM33PORSXEX3BNZSHE33JMQ>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>