MIT is not copyleft. The copyright notice must be included for those incorporated elements, but other downstream code it remains part of can be licensed however it wants.

AGPL and GPL are, on the other hand, as you describe.

Modifications can be licensed differently but that takes extra work. If I release a project with the MIT license at the top of each file and you download my project and make a 1-line change which you then redistribute, you need to explicitly mark that line as having a different license from the rest of the file otherwise it could be interpreted as also being MIT licensed.

You also could not legally remove the MIT license from those files and distribute with all rights reserved. My original granting of permission to modify and redistribute continues downstream.

Same.

If you’re able to self host and run the tool for any use, it’s effectively a free, extensible, modifiable software solution.

Copyleft licenses are as restrictive as the license DHH put out with Fizzy. I’m an Apache 2.0 or MIT licensing OSS advocate myself, but it’s difficult to argue that it’s worse or equal to a fully closed SaaS solution.

It’s not even remotely close to one of these bullshit “ee” OSS licenses

Amen.

I can assure you that inference is not profitable if the user is paying nothing.

DAU/MAU stats of free users have already carved out multi-millionaire and billionaire fortunes for employees and executives, all paid out with VC money. Plenty of people are profiting, even if the corporation is deep in the red.

Subjective. Most in creative industries regularly still use SDXL.

Once Z-image base comes out and some real tuning can be done, I think it has a chance of replacing it for the function SDXL has

I don't think that's fair. SDXL is crap at composition. It's really good with LoRAs to stylize/inpaint though.

Source?

Most of the people I know doing local AI prefer SDXL to Flux. Lots of people are still using SDXL, even today.

Flux has largely been met with a collective yawn.

The only thing Flux had going for it was photorealism and prompt adherence. But the skin and jaws of the humans it generated looked weird, it was difficult to fine tune, and the licensing was weird. Furthermore, Flux never had good aesthetics. It always felt plain.

Nobody doing anime or cartoons used Flux. SDXL continues to shine here. People doing photoreal kept using Midjourney.

> it was difficult to fine tune

Yep. It's pretty difficult to fine tune, mostly because it's a distilled model. You can fine tune it a little bit, but it will quickly collapse and start producing garbage, even though fundamentally it should have been an easier architecture to fine-tune compared to SDXL (since it uses the much more modern flow matching paradigm).

I think that's probably the reason why we never really got any good anime Flux models (at least not as good as they were for SDXL). You just don't have enough leeway to be able to train the model for long enough to make the model great for a domain it's currently suboptimal for without completely collapsing it.

> It's pretty difficult to fine tune, mostly because it's a distilled model.

What about being distilled makes it harder to fine-tune?

AFAIK a big part of it is that they distilled the guidance into the model.

I'm going to simplify all of this a lot so please bear with me, but normally the equation to denoise an image would look something like this:

    pos = model(latent, positive_prompt_emb)
    neg = model(latent, negative_prompt_emb)
    next_latent = latent + dt * (neg + cfg_scale * (pos - neg))

So what this does - you trigger the model once with a negative prompt (which can be empty) to get the "starting point" for the prediction, and then you run the model again with a positive prompt to get the direction in which you want to go, and then you combine them.

So, for example, let's assume your positive prompt is "dog", and your negative prompt is empty. So triggering the model with your empty prompt with generate a "neutral" latent, and then you nudge it into the direction of your positive prompt, in the direction of a "dog". And you do this for 20 steps, and you get an image of a dog.

Now, for Flux the equation looks like this:

    next_latent = latent + dt * model(latent, positive_prompt_emb)

The guidance here was distilled into the model. It's cheaper to do inference with, but now we can't really train the model too much without destroying this embedded guidance (the model will just forget it and collapse).

There's also an issue of training dynamics. We don't know exactly how they trained their models, so it's impossible for us to jerry rig our training runs in a similar way. And if you don't match the original training dynamics when finetuning it also negatively affects the model.

So you might ask here - what if we just train the model for a really long time - will it be able to recover? And the answer is - yes, but at this point the most of the original model will essentially be overwritten. People actually did this for Flux Schnell, but you need way more resources to pull it off and the results can be disappointing: https://huggingface.co/lodestones/Chroma

Thanks for the extended reply, very illuminating. So the core issue is how they distilled it, ie that they "baked in the offset" so to speak.

I did try Chroma and I was quite disappointed, what I got out looked nowhere near as good as what was advertised. Now I have a better understanding why.

How much would it cost the community to pretrain something with a more modern architecture?

Assuming it was carefully done in stages (more compute) to make sure no mistakes are made?

I suppose we won't need to with the Chinese gifting so much open source recently?

> How much would it cost the community to pretrain something with a more modern architecture?

Quite a lot. Search for "Chroma" (which was a partial-ish retraining of Flux Schnell) or Pony (which was a partial-ish retraining of SDXL). You're probably looking at a cost of at least tens of thousands or even hundred of thousands of dollars. Even bigger SDXL community finetunes like bigASP cost thousands.

And it's not only the compute that's the issue. You also need a ton of data. You need a big dataset, with millions of images, and you need it cleaned, filtered, and labeled.

And of course you need someone who knows what they're doing. Training these state-of-art models takes quite a bit of skill, especially since a lot of it is pretty much a black art.

> Search for "Chroma" (which was a partial-ish retraining of Flux Schnell)

Chroma is not simply a "partial-ish" retraining of Schnell, its a retraining of Schnell after rearchitecting part of the model (replacing a 3.3B parameter portion of the model with a 250M parameter replacement with different architecture.)

> You're probably looking at a cost of at least tens of thousands or even hundred of thousands of dollars.

For reference here, Chroma involved 105,000 hours of H100 GPU time [0]. Doing a quick search, $2/hr seems to be about the low end of pricing for H100 time per hour, so hundreds of thousands seems right for that model, and still probably lower for a base model from scratch.

[0] https://www.reddit.com/r/StableDiffusion/comments/1mxwr4e/up...

Nah.

Categorizing all AI as slop lacks nuance and demonstrates a shallow understanding of art

Not AI, suno.

What is the difference between

1.) Sampling a real snare sound 2.) Suno generating a couple random snare hits for you to choose from

? There are ways to use it that aren't far-removed from how real producers work.

It's different to say "Generate me top-40 Sounding pop song"... But Suno has more uses than that.

There are already royalty free drum samples for days, and infinite possibilities of drum synthesis. No producer is hurting for snare samples. Why would I pay money for something I already have.

Tweaking drum sounds until they sound good is one of the more satisfying activities for a beat producer, and one of the primary ways I demonstrate my value. I’m not paying to turn that over to someone else.

I don't oppose synthesis or samples. But arguably Suno songs typically goes way little beyond that. If you didn't chose the samples, didn't chose the words, didn't chose the voice, but vibed for a few mins what you did is not made a song, you heard a song. And the songs suck.

Don't bother, the slop line has been drawn in the sand and people will stay on their side. I also think that slop is low effort and/or low/zero iteration output, but beyond that the sloppiness fades away, but good luck convincing the hardcore slop hunters to see things in a nuanced way (in before ocean boiling, that's usually where that argument goes next).

It’s a very useful line in the sand. No Gen AI anywhere near art.

Once you add in a level of nuance, youre bickering about the degree which is far worse argument to be having, tactically.

So you're saying your perception of what is or isn't art is ideologically motivated? That seems like the worse "tactical" argument.

Why don't we scrap anything that uses ableton because it makes art sterile [1][2]. Or maybe anything that uses autotune [3]. Maybe we can have stickers that say "AI free". Or maybe the fact that suno is a distribution platform that doesn't encourage creation of the _form_ of art that that I like is the problem [4].

It's a tool. Your view that art exists in some purist state and isn't for people to enjoy is extremist. This war has been fought and lost, continuously, about every innovation in music. People want to enjoy things. You can tell by their pattern of consumption.

[1] https://www.wired.com/2002/05/laptop-2 [2] https://web.uvic.ca/~aschloss/publications/JNMR02_Dilemma_of... [3] https://www.soundonsound.com/techniques/vocal-fixes [4] https://www.salon.com/2003/06/18/itunes_innovation

Yes, I oppose AI ideologically.

If you are against most uses but want to introduce nuance, my argument is that doing so normalizes the use for the majority of cases you do oppose and makes it harder to organize opposition.

If you’re cool with AI in all cases we don’t have much to talk about.

Who knew that nuance was the enemy of rational thought.

It’s the enemy of successful organization.

I have thoughts on when AI is appropriate, but the conversations I want to be having is ‘how do we oppose AI’ and not ‘why is my specific definition of what is Ok better than your very similar one’.

Also, once Any AI is allowed, each step beyond that will be barely worth fighting for because it’s only just beyond acceptability.

Turning conversations away from nuance and towards black/white thinking is precisely why we're in an increasingly polarized society.

You may see it as a means towards the collective action you'd prefer, but your argument is that sheep are easier to herd towards a goal you've already determined is "right" without inviting critical thought or analysis when you don't allow for nuance.

Even if true, I disagree with both your assertion that "organizing against AI" is the right path forward, and the approach to engaging in discourse.

I mean, you have AI in your name, of course you don’t :p

I also don’t really think the problem with society is that we’re ‘polarized’, a viewpoint which believes that there are merits to “both sides” of human rights issues.

Must be nice to live in a world without trolley problems.

I don't think I'm a hardcore slop hunter really. Suno was fun to play with, what comes out feels icky though. Perhaps that is because it is low effort though, my attempts certainly were.

I'm very bullish on AI, but low effort outputs just lobbed over the fence in my direction (be that code, music, text) irk me to no end. And that's not even considering if something is tasteful or kitsch. So yeah, for me it's definitely that. Someone who puts in real work, and produces output with intent, I might still not like because it's not my taste, but at least I don't consider it slop.

Funnily enough, slop AI video I still get entertainment value out of, just because it's often so bizarre and absurd.

That is an interesting point. I am not sure how to define Slop. I am 100% sure you could use Suno for something that doesn't feel horrible. I tried and failed bad. But maybe it is possible. Typically I just hear on the surface song like audio but without anything to make me care about the song at all.

I won't disagree that most use of AI is slop - Just like most people in a 'sips and strokes' class make slop.

The issue I raise is that we can critique the users of it without discarding the tool as used by an artist who produces something that is not slop with it.

You were going to claim the “win” on this one no matter what anyone says. That’s one of the features of being arrogant.

I have zero desire to get into a semantic argument over this. That would be very boring and is a poor refuge for anyone trying to have an honest discussion.

It’s not a coincidence that nearly all the people today with actual artistic talent universally despise AI. Meanwhile it’s all the talentless tech bros who won’t shut up about how they’re now incredible “artists” that love AI.

I don’t see prompting an AI as creating art in the same way that commissioning a painting doesn’t make you an artist. In this example, it’s the AI model that is the artist creating the “art”, but since AI models aren’t sentient (yet) then what they create isn’t art anymore than a sunset is despite being aesthetically beautiful.

Your “it’s just a tool” argument is especially ridiculous when you consider that the “tool” can create the same “art” in its entirety without you. It would be like if I googled the Mona Lisa and copy pasted it into Paint and then called myself an artist because I used Google and Paint as my tools.

In the case of AI models anything you can think to prompt is already embedded in the model so it’s not like you’re even creating anything. It’s already there. If you have infinite monkeys on a keyboard prompting AI they can generate every single possible image an AI is able to generate. Where is the artistry again?

I'll note I used no pejorative terms to describe you in order to make my points. I don't need to cast you as a villain or "unwise" individual to argue against the statements themselves.

--

Your framing of AI presumes that my proposal is prompt in >> art out. False frame.

You also state "near universal" disdain of AI from people with "actual" talent. A bold claim to make with little data, and at least for me, clear examples of being false.

I know award-winning directors, iconic creatives, and career/professional artists who are all excited by the technology, exploring ways to use it, and learning how to composite it into their work.

Perhaps you would propose that they have no talent and are not "true" artists! This is why I ask for the definition of art. It's not a "poor refuge" for an honest discussion, it's something that is the fundamental term upon which your argument is hinged. You can't gatekeep "art" without defining what the gate is.

The OP talks about Suno Studio - A DAW experience with recording, editing, generative tools, and generative restyling. This is, objectively, a different level of tool than a "prompt to music" generation experience.

Something that is composited through multiple decisions made by a human, toward some end that only they see, through an iterative process...

Well, I'd say that's where the artistry is.

No, there is nothing artistic about tools that help you modify or organize your prompts aka commissions for all the reasons I already stated above.

Let’s play a game. I’ll pick an artist who hates AI and you reply with an artist that loves AI. Let’s see who runs out first.

https://www.wired.com/story/guillermo-del-toro-hopes-hes-dea...

Modify or organize your prompts? lol - I don't think you have any familiarity with the tooling landscape in AI media gen.

I'll do you one better - I'll cite someone who has changed their mind.

https://deadline.com/2025/04/james-cameron-use-ai-cut-cost-f...

I think I finally see what we’re clashing on. Where do you draw the line between “this is a tool assisting me in creating my art” and “here is my AI film I stitched together with 10 Sora prompts”?

The songwriter for K-Pop Demon Hunter’s Soda Pop used AI to help them write the lyrics - fine. The hacks on X calling themselves incredible artists because they can prompt Dall-E - slop.

Yes - If you're willing to allow that AI might be a tool in the artistic process if it is used in conjunction with intent and effort, then we are at a reasonable position -- That stating "all AI is slop" is as inarticulate and insufficient as the claim of "look at what I made, I'm an artist!" coming from someone who typed in a four word prompt.

The majority of people using AI believe that production itself is art. They produce slop because they have no craft, no taste, and no innate talent.

Artists recognize that art requires the development of craft and taste - and there are many that see AI as a tool that can be used in that process.

So you must really hate collage artists. All they are doing is cutting out images and pasting them into new images. Picasso really produced some slot I guess. You seem to have a very rigid definition of "art".

That has nothing to do with AI, which itself is categorically and noncontingently "slop".

The issue here is that you are trying to bridge two disparate goals - making money and helping kids.

The fact that this isn’t open source, as it stands, means the latter is not a primary goal - which is not an indictment, just an observation.

The complaints will come, regardless, for that reason alone, given the marketing/narrative.

You’re selling a product to parents/educators who want to gamify the technical education of their children. That market, small as it is, despises micro transactions.

A sustainable business has the capacity to help a lot more kids than an unfinished open source project that never gets released on iOS because no one wants to pay the developer fee.

This isn’t “HackVille by Zynga,” it’s an indie dev trying to make a product they believe in. I hope it succeeds and inspires more high quality edutainment.

My point is that packaging the app in such a way as to put off your target audience is inherently unsustainable business.

I agree with that criticism, and I'd encourage the dev to iterate on non-micro-transaction monetization schemes. The part I disagree with is that a profit motive is antithetical to helping kids.

It'd be nice if we had robust, no-strings attached funding streams to make this kind of content, but we don't, so if we want it to exist, consumers need to pay for it.

You're not arguing against the GP but for the same thing from different angles. They're saying the approach is fighting the goal, while you're just saying "I hope they're successful".

I was responding to the claim that making money is in tension with helping kids learn.

I think it’s fair to claim that a large enterprise will eventually crank the money dial to maximum extraction. But a solo dev is free to follow their conscience and make money in a responsible way.

I don’t like the “pay per hint” model as currently implemented, but I’m willing to give the developer the benefit of the doubt that they didn’t think it all the way through.

What does open sourcing an application have to do with helping kids?

There are plenty of arguments for open sourcing things. “Closed source apps necessarily deprioritize helping children” is not an obvious argument to me. Can you draw the connection more explicitly?

Scale and accessibility - Eliminating any barriers for children to get access to education, etc.

Not to mention, it’s an app trying to help kids get exposed to underpinning technologies - seeing how the game itself is made would be optimizing for that end.

It’s not that closed source deprioritizes, but the “helping kids” were the sole and primary goal sought, there’s a clear answer to what would align with that.

All said, it’s not a critique of the OP - reconciling ideals and practical reality often require trade offs that would allow for a project like this to happen in the first place.

I think it's hugely important to eliminate barriers to get access to education, which is why there's a free, web-based version of Hacktivate that is already being used 350+ schools around the world.

I also think there's a lot of people out there who would pay to have Hacktivate running offline, using the full power of their device, and with no external resources being required, so I made that too.

Suggesting that I need to make them open source to prove I want to help kids learn is really strange, particularly when literally thousands of students around the world are benefitting from my work without paying a cent.

As mentioned, no indictment, and you don’t need to prove anything - helping kids learns is clearly a goal.

But so too is making money off the iOS app, correct?

It's not DNS

There's no way it's DNS

It was DNS

It's "DNS" because the problem is that at the very top of the abstraction hierarchy in any system is a bit of manual configuration.

As it happens, that naturally maps to the bootstrapping process on hardware needing to know how to find the external services it needs, which is what "DNS" is for. So "DNS" ends up being the top level of manual configuration.

But it's the inevitability of the manual process that's the issue here, not the technology. We're at a spot now where the rest of the system reliability is so good that the only things that bring it down are the spots where human beings make mistakes on the tiny handful of places where human operation is (inevitably!) required.

> hardware needing to know how to find the external services it needs, which is what "DNS" is for. So "DNS" ends up being the top level of manual configuration.

Unless DNS configuration propagates over DHCP?

DHCP can only tell you who the local DNS server is. That's not what's failed, nor what needs human configuration.

At the top of the stack someone needs to say "This is the cluster that controls boot storage", "This is the IP to ask for auth tokens", etc... You can automatically configure almost everything but there still has to be some way to get started.

I just did it.

You were right.

It is, in fact, that good.

You could have recorded, found it to be good, and didn't shared the news. Only used for your self. But you decided to share only the news, not the recording. That tells me something.

To be more clear, I can move this argument further. I promise you that if you share the recording that led you to believe that, I will not judge it. In fact, I will do the opposite and focus on people who judge it, trying my best to make the recording look good and point out whoever is nitpicking.

Generally, this seems an obvious and correct decision for Tor.

Barring integration with a locally run LLM, AI doesn't make sense for the Tor security posture - you don't want to be routing content to unintended/insecure third parties, period.

and for any malicious mfs, you dont want AI in your browser.