Master AI (Artificial Intelligence) Discussion/News Thread

[xanalog]

I see we are now in the minority so all those counts of “1m million users per month” to justify that one website is better than another is rather chin strokinly suspect

Would there be an oncoming threat to our beloved ASR forum's traffic #s"
I mean: Why go thru the laborious effort to post an audio-related question here, when you can readily fetch answers from you-know-who?
Competition?

 

[JustAnandaDourEyedDude]

As a teen, I knew author Roald Dahl principally for some horror and fantasy and a little bit of raunch. Only many years later did I learn that he was famous for children's stories.

In 1954, was Roald Dahl hallucinating the future? Was he prescient about the possibility of creating LLMs in computer memory that stores all words? About how to deal with the loss of writers' jobs that this might entail? The YouTuber Tibees tries to persuade us in a 20-minute-long video. You be the judge.

 

[JustAnandaDourEyedDude]

Terence Tao explains that the math behind LLMs is simple enough that an undergrad should be able to perform the calculations. Yet, he says, we cannot yet fully understand why the LLM's output has the characteristics it does. This is a six-minute clip from an interview by Brian Keating. At the end of the video is a link to the full hour-plus interview.

 

[Somafunk]

This is a six-minute clip from an interview by Brian Keating. At the end of the video is a link to the full hour-plus interview.

Interesting vid but annoying to follow with the editing so I’ll watch the full 1hr interview

 

[MattHooper]

I gather the following is based on Cloudflare network traffic data:
Agentic/bot web traffic is said to be "57.4% of all web activity", as of May 2026. They have officially surpassed the web-based human activity, spiking by 187% in 2025.
According to a CNBC report, 'automated AI programs for human users' also rose by a whopping 8,000%.
Projections cum predictions were that this flip would not occur until the end of 2027 but, here we are!

The meaning of these spikes are not exactly known at the current time for humans and/or the internet.
I have found successful work-arounds for the internet's present 'advertisement' biz-model but that model may be in peril (at best) or becomes AI-infused (at worst).

The old two-tier internet concept is resurfacing again and may even be making a comeback already: In 2025, Cloudflare launched "Pay per Crawl", which allows publishers to charge AI scrapers for access to their content, via "toll roads". Cloudflare CEO -Mathew Prince- puts it this way:

Dorothy was right!

Hideous.

Whenever I start reading replies to a post on X I have to remind myself that a large percentage of the replies are bots. So I’ve mostly stopped reading them.

Just recently I got duped - I was on a philosophical reddit forum and end up engaging in debate with a maddening character, and after a number of exchanges, somebody else pointed out to me “ you’re talking to an LLM bot constructed to troll and keep engagement.” Ugh ! Should’ve been so obvious in retrospect.

When my friends and I talk about it, we all agree that the Internet sure doesn’t feel the same as it used to. Websites have become almost useless not just due to advertising, but due to AI taking over.

It’s one reason why I hold on dearly to forums like this for my hobbies.

I wonder when we’re gonna start having bots intruding here and fooling the staff, they are real people. Has that happened yet?

 

[entsnack]

Thought I'd share my recent experiment here! I have two clustered DGX Sparks running the Ace-Step 1.5 XL music generation models and saving the results to my Plex server. I fine-tune these on my organic FLAC collection to give the model some direction. This gives me an effectively infinite supply of private and personalized music.

It's not perfect, but I can imagine a future where musicians become pure data-providers for generative AI. I find it immensely satisfying listening to Shrimp Bizkit and Phlegminem on repeat, and much prefer this to the organic music created after 2011. Though I am sad about the loss of community and live performances.

I am now working on adding a reinforcement learning interface where I can upvote/downvote generations and have the model continually learn from my preferences.

 

[xanalog]

Thought I'd share my recent experiment here! I have two clustered DGX Sparks running the Ace-Step 1.5 XL music generation models and saving the results to my Plex server. I fine-tune these on my organic FLAC collection to give the model some direction. This gives me an effectively infinite supply of private and personalized music.

It's not perfect, but I can imagine a future where musicians become pure data-providers for generative AI. I find it immensely satisfying listening to Shrimp Bizkit and Phlegminem on repeat, and much prefer this to the organic music created after 2011. Though I am sad about the loss of community and live performances.

I am now working on adding a reinforcement learning interface where I can upvote/downvote generations and have the model continually learn from my preferences.

So sad!
I think those of us who value our music libraries may have to consider the use of 'chastity belts' on our organic music collections, so as to keep them virgins, unmolested by AI and away from the ultra-processed variety that music is about to become.
Although, I realize that my above sentiments are probably similar to those made against the "electric" guitar when it raised -the proverbial musical- 'bar'.

 

[MattHooper]

Wasn’t sure where to put this, but I figured since this is an AI thread….

I found this new video interesting from Jay’s Iyagi.

Jay tries out what is apparently the world’s first amplifier with AI features. The Luxsin X8 from China.

It’s a headphone amplifier.

Using prompts, Jay tries to get the amplifier to duplicate all sorts of different sound characteristics - added spaciousness, added punch and dynamics, the sensation of listening to loudspeakers, and he also tries to get it to duplicate the sound of his tube headphone amp:

 

[JustAnandaDourEyedDude]

Interesting vid but annoying to follow with the editing so I’ll watch the full 1hr interview

I do not recall there being much more about AI in the full interview than Keating collected in the clip. They covered several other topics. I have watched a couple of hour-long lectures by Terence Tao in YT videos, related to the use of AI to solve research math problems. I could not follow the math, but they were inspiring nevertheless, Tao being a Fields Medalist and in the top echelon of present-day mathematicians. He is an excellent communicator, able to explain advanced math in an easy-breezy way for audiences without the strongest math background.

 

[Axo1989]

View attachment 537636

Thought I'd share my recent experiment here! I have two clustered DGX Sparks running the Ace-Step 1.5 XL music generation models and saving the results to my Plex server. I fine-tune these on my organic FLAC collection to give the model some direction. This gives me an effectively infinite supply of private and personalized music.

It's not perfect, but I can imagine a future where musicians become pure data-providers for generative AI. I find it immensely satisfying listening to Shrimp Bizkit and Phlegminem on repeat, and much prefer this to the organic music created after 2011. Though I am sad about the loss of community and live performances.

I am now working on adding a reinforcement learning interface where I can upvote/downvote generations and have the model continually learn from my preferences.

I’d hope you were using the nice hardware for something more than spinning up variations of other people’s old music? Music stopped in 2011 eh? Crikey …

Anyway there’s an ‘AI and the Future of Music Production’ thread you might be looking for.

AI and Future of Music Production

In a recent thread there was a discussion around what the technology would be like in 2074. I commented that by then AI will be able to on-demand generate any music we like. We could for example give it a seed song, and then modify it with our mood, desires, and even fidelity levels and let...

www.audiosciencereview.com

 

[Somafunk]

View attachment 537636

Thought I'd share my recent experiment here! I have two clustered DGX Sparks running the Ace-Step 1.5 XL music generation models and saving the results to my Plex server. I fine-tune these on my organic FLAC collection to give the model some direction. This gives me an effectively infinite supply of private and personalized music.

It's not perfect, but I can imagine a future where musicians become pure data-providers for generative AI. I find it immensely satisfying listening to Shrimp Bizkit and Phlegminem on repeat, and much prefer this to the organic music created after 2011. Though I am sad about the loss of community and live performances.

I am now working on adding a reinforcement learning interface where I can upvote/downvote generations and have the model continually learn from my preferences.

Thats a piss take?.....right?

 

[Axo1989]

Wasn’t sure where to put this, but I figured since this is an AI thread….

I found this new video interesting from Jay’s Iyagi.

Jay tries out what is apparently the world’s first amplifier with AI features. The Luxsin X8 from China.

It’s a headphone amplifier.

Using prompts, Jay tries to get the amplifier to duplicate all sorts of different sound characteristics - added spaciousness, added punch and dynamics, the sensation of listening to loudspeakers, and he also tries to get it to duplicate the sound of his tube headphone amp:

Hmmm. Interesting. He’s wrong about big soundstage actually. The biggest rendition I get are video soundtracks on Apple Vision Pro. Which starts with a binaural downmix of the Atmos. Really, sounds coming from everywhere in just the manner you’d like. And from quite small transducers close to your ears. So it’s entirely possible. But AVP has a few extra tricks like LiDAR scanning the room to factor that into the presentation.

 

[Snarfie]

View attachment 537636

Thought I'd share my recent experiment here! I have two clustered DGX Sparks running the Ace-Step 1.5 XL music generation models and saving the results to my Plex server. I fine-tune these on my organic FLAC collection to give the model some direction. This gives me an effectively infinite supply of private and personalized music.

It's not perfect, but I can imagine a future where musicians become pure data-providers for generative AI. I find it immensely satisfying listening to Shrimp Bizkit and Phlegminem on repeat, and much prefer this to the organic music created after 2011. Though I am sad about the loss of community and live performances.

I am now working on adding a reinforcement learning interface where I can upvote/downvote generations and have the model continually learn from my preferences.

Interesting workflow. My approach is a bit different. Instead of fine tuning on existing music, I put a lot of arrangement knowledge directly into the prompts. I specify things like style, mood, ghost notes, micro-timing variations in drum patterns, exact chord choices, where chord progressions should change/appear, the drum feel (for example, a shuffle groove), bass playing style, and even detailed characteristics of brass and other instruments.

In a way, I'm treating the AI more like an arranger than a composer. It follows the harmonic and structural directions I give it, then does what it is particularly good at reinterpreting, rearranging, and developing chord progressions and musical ideas in unexpected ways.

IMO (subjective) the output is often surprisingly interesting from a creative standpoint, especially when you get very specific with the musical details like the creative interpretations from single instrument produced over chord progressions.

 

[entsnack]

Interesting workflow. My approach is a bit different. Instead of fine tuning on existing music, I put a lot of arrangement knowledge directly into the prompts. I specify things like style, mood, ghost notes, micro-timing variations in drum patterns, exact chord choices, where chord progressions should change/appear, the drum feel (for example, a shuffle groove), bass playing style, and even detailed characteristics of brass and other instruments.

In a way, I'm treating the AI more like an arranger than a composer. It follows the harmonic and structural directions I give it, then does what it is particularly good at reinterpreting, rearranging, and developing chord progressions and musical ideas in unexpected ways.

IMO (subjective) the output is often surprisingly interesting from a creative standpoint, especially when you get very specific with the musical details like the creative interpretations from single instrument produced over chord progressions.

This is amazing. You're an expert, so you can accurately articulate what you want, and that's key to eliciting quality outputs from these models. I'm far from an expert, but I can provide examples of what I like, which helps to some extent but it's not as good.

The private context in your head is the final
barrier between today's AI and AGI. One of my research projects is actually on designing markets for agents to elicit this context from humans, because (my hypothesis) agents cannot synthesize tacit knowledge themselves so they cannot self-prompt themselves.

 

[Blake Klondike]

Hideous.

Whenever I start reading replies to a post on X I have to remind myself that a large percentage of the replies are bots. So I’ve mostly stopped reading them.

Just recently I got duped - I was on a philosophical reddit forum and end up engaging in debate with a maddening character, and after a number of exchanges, somebody else pointed out to me “ you’re talking to an LLM bot constructed to troll and keep engagement.” Ugh ! Should’ve been so obvious in retrospect.

When my friends and I talk about it, we all agree that the Internet sure doesn’t feel the same as it used to. Websites have become almost useless not just due to advertising, but due to AI taking over.

It’s one reason why I hold on dearly to forums like this for my hobbies.

I wonder when we’re gonna start having bots intruding here and fooling the staff, they are real people. Has that happened yet?

it seems like nobody ever answered this question. are we getting blasted by bots here?

 

[Snarfie]

This is amazing. You're an expert, so you can accurately articulate what you want, and that's key to eliciting quality outputs from these models. I'm far from an expert, but I can provide examples of what I like, which helps to some extent but it's not as good.

The private context in your head is the final
barrier between today's AI and AGI. One of my research projects is actually on designing markets for agents to elicit this context from humans, because (my hypothesis) agents cannot synthesize tacit knowledge themselves so they cannot self-prompt themselves.

I wouldn’t really call myself a expert or musician. I play a bit of bass, but more importantly I have a general understanding of how arrangements work. The kind of music that resonates with me is mainly jazz-oriented stuff.

For me it’s ultimately about what you feel and recognize in the music, which is why I use the word “resonate.” I’m 67 now, and I’ve been listening seriously to music for well over 57 years later becoming more focused on jazz in particular, which is where I really found my reference point.

Over the years I’ve been to a lot of live performances and watched countless interviews video's with jazz legends, so I’ve been exposed to the language of the genre in a fairly direct way.

I’d say I’ve developed a pretty good ear for balanced, groove based music jazz, jazz-rock, neo-jazz, funk, and related styles. That’s probably where my speciality, if you can call it that, lies, recognizing how these elements come together more or less rather than performing them myself.

 

[Peterinvan]

I will be sad when all newly produced "music" is produced by AI or loops from a synth. I get the best emotional reaction listening to a small group "in the groove" or "flow" being in sync with other, playing real instruments. Watch this Bela Fleck group on Tidal.

Béla Fleck, Sierra Hull & Molly Tuttle - Wheels Up

Watch Wheels Up on TIDAL

tidal.com

 

[entsnack]

Thats a piss take?.....right?

Not at all. Not very different from setting up solar panels or wind turbines to generate your own electricity because the state decided your property was not worth builiding the infrastructure out to. Musicians don't make music for you, they make it for labels, who need profits and mass market appeal. My setup fixes that problem for me. I do wish my music tastes remained mainstream over the ages, so I didn't have to do this.

 

[Somafunk]

Not at all. Not very different from setting up solar panels or wind turbines to generate your own electricity because the state decided your property was not worth builiding the infrastructure out to. Musicians don't make music for you, they make it for labels, who need profits and mass market appeal. My setup fixes that problem for me. I do wish my music tastes remained mainstream over the ages, so I didn't have to do this.

word salad but fair enough,

by mainstream?, what do you mean with regard to genre?, artist?

 

[entsnack]

word salad but fair enough,

by mainstream?, what do you mean with regard to genre?, artist?

The music itself.

For example: I'd like to listen to a male vocalist with a whiny nasal voice rapping over a rock band with a DJ in it. That kind of music isn't mainstream anymore, it doesn't sell today, so it is rarely created if at all.