thecollabagepatch

more experiments with nemotron3-nano on the dgx spark. it was rly rough yesterday but he's finally got hot reloading set up in his web UI to make changes/add tools and everything in his sandbox. i can even have him edit his own system prompt lol nemo gonna find himself

i feel like we've become too obsessed with the idea of effortlessness. we all need to get comfortable being frustrated again

@RoyalCities @volumetrique "making you more or less creative" is not rly anything i've ever seen software/samples/plugins do. its up to us how creative we wanna be with our toys

I can respect this. On creativity I find everything has degrees. I've used splice but I'm one of those people who resamples them to all hell. 1. Because of issues with content ID and the shared nature of sample marketplaces But more importantly 2. I love taking those samples and running them through things like Infiltrator or Shaperbox so they're totally unrecognizable. Using samples as-is is fine but yeah making them into something entirely new is what's fun to me. Some may see that as more creative while others may see it as just extra work for the same of work. So really I think as with most tools it depends how you use it. If someone was just going to take all these outputs and make a song of ONLY AI - where is the fun in that? But a piano chord progression - post processed with some LFOs to make something like future bass? Could be cool - but I can see why some would be against it. Same reason why some only use hardware over VSTs - people have to use what works best for them and their process.

Technically impressive, legally correct dataset, but still feels like "Optimizing the fun out of making music" to me. It also sounds like crap but eh. It's a very philosophical question - does this make you more creative or less creative? It can probably make you more [...]

i was very bummed about some things in october when we first got it generation time was really poor with all the audio models i use, but since then me and claude have hyper optimized it so that it beats the T4 on every audio task i give it. i can also train models while i keep this docker compose network active lol i'm very happy i went with it over the m4. it's also so nice that it's not my laptop so i can just keep it running experiments while i do other stuff

@thepatch_kev You think it’s worth the jump?

things a dgx spark lets me do sry jeff idk if we need you anymore

After months of work, today I’m releasing Foundation-1. A SOTA text-to-sample model built specifically for music production workflows. It may also be the most advanced AI sample generator currently available - open or closed. • ~7 GB VRAM • Entirely local • 100% free 😁

@TechMDAI just the one for now lol

@thepatch_kev How many do you have?

building non-deterministic software is enjoyable as f until you'd like to share it and show others how it works the unpredictability is both feature and bug

@RoyalCities yeah exactly it never applies to the actual videos i make and is just a slop factory lol

@thepatch_kev Yeah maybe one day. I've seen some automated AI editing tools but alot of them are trash. Very simple clips with some generic AI voiceover. Not great at all.

I decided to wrap the new sample generator into a YT video - sadly I edit slow af.😅 but I did want to share a quick showcase. Yes - this is all 1 model. Yes - it will be entirely free and entirely local. Yes - it knows Acid & Dubstep (+ a ton of other stuff). 😁

@RoyalCities for real lol it's like the only hire i dream about at this point i don't think claude will ever be able to do it sanely but idk... i've been wrong a lot lately about what i thought lil buddy would never be able to handle

@thepatch_kev Sounds great! Honestly I hate editing but it has to be done. I make zero money from any releases but if I could Id hire an editor. I like writing music for all the vids but damn keyframeing animations blows.

@mattbroadstreet oh interesting, yeah i was looking around for gird/bpm alignment and couldn't find it but it's so configurable it might exist somewhere in the depths. kind of a crucial feature for my vst to make a lot of sense but it is still usable

@thepatch_kev Mostly for soundtrack work in this case of course. Vegas had the clip snapping and time/pitch functions from Acid. Not sure what resolve has, if anything, in that department.

so i've been having a blast with da vinci resolve for like 12 hours and just discovered it has access to VSTs... no greater joy than discovering just how many applications your JUCE plugin can run inside

@mattbroadstreet that's so wild. just now migrating away from clip champ for videos lol feel like it would take years to uncover all the features da vinci has

@thepatch_kev I know people who used to compose music in Sony/Sonic Foundry Vegas that use Resolve for music now.

@yikesawjeez @shiraeis sounds very doable. i've seen something like this implemented a few places i wanna say @_lyraaaa_ had one that could literally populate the list of "neighbors" as you're typing, but i might be hallucinating

@shiraeis @thepatch_kev ne thots

both suno and udio will generate 10M+ songs this year. finding the one you actually want is still unsolved, but it shouldn’t be. this is the weirdest gap in AI music right now. the generation problem is approaching solved, but the discovery problem is mostly still tags and feeds. the solution is actually already latent in the problem. the problem: right now if you want to find a specific ambiance or vibe in either platform's library, your options are to either search by text tags or generate a new song entirely and hope it matches what's in your head. there’s no “tracks like this,” no taste profiles, no “why is this similar.” there's no way to navigate by sound. but this is a solved problem in ML. it's called contrastive language-audio pretraining (CLAP). you basically train two encoders (one for audio, one for text), so that matching descriptions and sounds exist near each other in a shared vector space. once you have that, a user can type ANY description and surface sounds that match. they could type “sad lofi with detuned piano and vinyl crackle” or “nostalgic retro 70s pop song,” and you’d just find the nearest audio vectors, no need for tags or a taxonomy. the model will have learned the mapping between language and sound directly. one step further, you can actually do audio-to-audio search. imagine the user uploads a 10 sec reference sound file, you embed it, find the nearest neighbors in the library. recommendation based on taste becomes a basic cosine similarity calculation (and then ranking, deduplication, and guardrails). the reason neither company has built this yet is probably prioritization. when you’re scaling generation models and fighting copyright lawsuits and shipping new features, search infra feels like a nice-to-have, but i’d argue it’s actually the key to unlock the next phase, where these platforms stop being “generate and hope for the best” tools and become actual creative environments where you can navigate a latent space of music. all the technical pieces are there and ready to go. CLAP gives you the joint embedding space; CoLLAP extends it to full length tracks instead of 10 sec clips; sparse concept decomp (Zhang et al, 2025) lets you explain exactly why two tracks are similar, not just that they are; and the actual retrieval infra (approximate nearest neighbor search over a vector database lol) is already proven at scale. suno has the harder version of this problem because their library is bigger and growing faster. udio has the more interesting version because they might use a diffusion-based architecture, which actually produces richer intermediate representations that could double as search embeddings. simply put, the latent space they’re already denoising IS a similarity space. i’m working on something tangential, but the question for @suno and @udiomusic is whether they build this for themselves or a third party builds it for them.

@tinycrops @jogamedev i don't think i wanna be a virus

@jogamedev @thepatch_kev I'm getting stories like this in my feed. It's interesting. Marketing seems like such a chore. I'd rather just effortlessly go viral.

Indie Game Devs: Your potential customer base is much larger than you think. Once I realized how this really works, it resulted in a massive influx of sales. Let me explain: For background, my spellcrafting game, Spellmasons, had plateaued for years and was losing traction when I recommitted myself to marketing. This resulted in my 3rd year, post launch, beating all the other years combined. Here's what I figured out that made this happen: - Your game is an asset you'll own forever - Everybody says launch causes the most sales and that's just not true for indies because indies usually don't reach anywhere near their Total Addressable Market in terms of exposure - chances are, indies will literally never reach every eyeball that might be a potential buyer, which means there are always more potential customers who would love your game if only they hear about it And even more importantly: New gamers are getting old enough to sign up for steam for the first time every. single. day. If I restart marketing in 5 years, there'll be 5 whole years worth of new gamers who might be interested. It's literally millions of new players every year. You'll never run out of people who are hearing about your game for the first time, no matter how viral your content is.

@tinycrops i have now

have you ever seen a vibe coded vtuber avatar before? me either.

@aarondelasy @iannuttall oh dope that's definitely one way to do it

@thepatch_kev @iannuttall well it was actually entire codebase in one giant markdown file – 10,000 lines and I fed into Claude to convert my Next.js app into a React Native application

anybody interested in acquiring codebase.md? it converts any public GitHub repo into a LLM-friendly markdown format with natural language search and gets ~10k human visitors a month not monetised but 10k/mo visits is worth something!

@ViralManager the least musician-facing model one could possibly conceive of lol

im so tired

@iannuttall @aarondelasy idk why there's something ironic to me about a markdown file named codebase lol

@aarondelasy its fate!

@dadabots @deepfates confirmed

@deepfates aphex twin was 3-5 people

Not sure I buy this. The analogy for music seems to point the other way in fact: we don't have as many bands now because you don't need 3 to 5 people to make a whole song. individuals are empowered to make music directly from a personal vision. Writing too: I don't need a magazine to publish me, the tools and distribution are all in my hands. I think getting more people to work together effectively on something has always been the hardest problem in software

@jordiponsdotme big mood. watching codex blast through things that were an epic journey a year ago while i get to practice guitar

glad i’m starting a career as an artist