I can guarantee to you that a lot of great music today is created and performed entirely without any "AI" involvement. True artists with something to say, and who make music for other reasons than wanting to be rich and famous (haha) still exist.

To give just one example among thousands: Daptone Records is one of the most consistently brilliant record labels, and they don't even allow computers on the studio floor.

I guess I skipped over the "successful".

I still don't think it's generally true. What distinguished the major pop artist from the indie artist is mostly stuff unrelated to the artistic side, I think. Meaning, maybe their managers, promoters, copywriters, cover designers... feel like they have to use LLMs, but not the artists or producers, in the studio.

This is maybe again biased by my preferred genres, but I don't see how an artist using AI in their songwriting or production would even appear on my radar. They wouldn't end up in any of the places where I find my music (local concerts, festivals, human DJs, personal recommendations...) they simply would be disqualified from all the channels that matter (or be unable to appear in them, because they can't perform live). If any slipped through, that would be a Milli Vanilli style situation.

What? I don't think I understand what you're saying. Making music is as easy as it's always been.

1. Pick up a an instrument.
2. Play it.

Some of the best music doesn't involve more than that.

Of course, people without any skills are going to have a hard time, at first. They can either put in the practice, or otherwise they can just get the heck out of music because they're clearly not interested in it for any of the right reasons.

It's easier than ever to get your stuff recorded and published. That used to be almost impossible without a label that would lend you money for studio time and pressing a record. Now you just click record in your basement and upload to a digital distributor. You never needed "AI" for any of that before, and you certainly don't now.

It was fresh when Cher used it in '98. But I also think barely anybody has done anything worthwhile with it since then.

Do the manufactured acts really still work these days? Maybe I'm biased by the genres I listen to, but my impression was that most of the artists who end up becoming reasonably popular start out being ignored by the major labels, build a reputation and fanbase through independent releases, concerts, freeform radio etc., until suddenly they end up on some A&R's radar and suddenly every major wants a piece of the pie.

I'd still call that discovering, even though I think it's mostly trying to react, always lagging behind the tastes and trends people are building elsewhere.

"Purposefully" was a bad choice of word. What I meant is more like, something with intent behind it, an idea, as opposed to something seemingly novel the model generated that somebody could interpret some meaning into afterwards.

There are still going to be limits to that approach. You're either going to have results that stay reasonably in-distribution, or you get the tool to venture far enough out of it where, by nature, it will become erratic and a lot less likely to generate something agreeable. An LLM simply isn't ever going to give you a purposefully novel idea like a human artist could.

> "if it was opt-in, nobody would opt in,"

Classic saying the quiet part out loud, I love it.

> if you publish something for free or upload it to a service, it is in *legal* fact, fair use for AI training

Not automatically. On a service like Twitch or YouTube, perhaps, because they give you limited (if any) control over the license terms with which you can publish your works on their platforms.

But you can certainly upload works of yours for free and public access and put them under a license that forbids use in LLM training. For example, my understanding is that a CC license with the ND clause strictly forbid such use (if you are going to publish your model, at least).

If it's your work, the copyright is yours, and you set the rules. People are of course free to give up some of those rights by uploading to a service like Twitch.

That term is so vague that it's practically meaningless. It would have to be challenged in court what that can and can't include.

Writing purposefully vague terms doesn't give them some carte blanche. Reposting some uploader's video in a way to publicly slander them, for instance, is also a "use", but certainly not legal.

...by spiking it with nicotine to make it addictive.

In the end, every such train of thought will always circle back to the undeniable conclusion of "too many damn people on the planet."

I don't think they're the only one who knows. I think pretty much everybody who uses frontier models daily and extensively (which includes me) notices it.

The hype was real a couple of years ago. Now, the reactions to every new, supposedly groundbreaking model is split about 50:50 between "maybe slightly better" and "worse than the last one". Sounds like plateauing to me.

Most actual recent progress was in harnesses, connecting tools, and stuffing things in loops, not model performance.

The issue is deeper than that, because there is no way to encode a "law" into the inherently fuzzy nature or LLMs. System promots, Agents.md, users putting rules into prompts... they're all just stirred into the same context soup, used to influence which tokens are generated, which then influence which tools are actuated.

To actually encode laws, as mushily and ambiguously worded as they are in natural language -- it can't be done without breaking the fundamemtal value proposition of general applicability of LLMs. You can either have hard rules, or drive a system with an LLM, not both.

Oh, wow. That settles it then, folks! Wind power means that some farmers have to replace a fence every 10 years. Clearly, renewable energy is not worth it, let's roll it back!