noomb00m

Demystifying AI Music. Breaking down Suno/Udio/MusicLM research for the every day human

[noomb00m @noomb00m]

In 2025, AI music went from fringe experiment to mainstream reality, charting tracks, sparking debate, and challenging how we define creativity. The future of music is here, but it’s complicated. Check out my post on the state of AI music around the globe to find out why🎵🤔!

https://t.co/oZ0NlsMW70

[noomb00m @noomb00m]

The compression researchers and the LM researchers were converging on the same thing from opposite directions.

Sound is just another sequence now. That wasn't the plan. That was the side effect.

Paper: https://t.co/3nmroo2IZo

[noomb00m @noomb00m]

We spent decades compressing audio to save bandwidth.

Turned out we were accidentally solving a different problem.

[noomb00m @noomb00m]

Qwen3-TTS is what happens when someone takes that seriously.

No separate speech engine. The LM predicts codec tokens. Same as it predicts words. Flow matching decodes them back to audio.

[noomb00m @noomb00m]

Link to the full paper

https://t.co/TpinMynGMM

[noomb00m @noomb00m]

VoiceCraft beats SOTA speech editing without redesigning the architecture.

Standard AR models only see the past — if you're editing a word mid-sentence, the model generates blind with no idea what comes after.

Their fix: don't change the model. Change the sequence.

Mask the edit region → move future context before the generation point → model now sees both sides.

Same causal LM. Same training setup. Bidirectional context for free.

Next time you're stuck trying to redesign a model — ask if you can solve it with sequence ordering instead. The architecture is rarely the bottleneck.

[noomb00m @noomb00m]

@GuGi263 Still stood on bihness though

[Vai Viswanathan @vai_viswanathan]

we also tried training a World Model on the 90 trajectories we had — obviously very overfit. it didn’t finish training until a few hours after the hackathon ended, but cool learning experience.

shoutout @PrimeIntellect for the credits!

[Noam Hurwitz @ProbablyNoam]

@vai_viswanathan @earthtojake @noomb00m @demaria_michael @nurcannsonmezz Looks like we inspired @Dyson too!

https://t.co/nKPVui49ar

[Vai Viswanathan @vai_viswanathan]

It was fun building Alfred Café as part of this hackathon!

@earthtojake @noomb00m @ProbablyNoam @demaria_michael @nurcannsonmezz

[Vai Viswanathan @vai_viswanathan]

it will be really ironic if "American Industrialization" just ends up being outsourcing teleop

[dadabots @dadabots]

🥳 Announcing Stable Audio 3 🍕

🏆 fastest music models ever
💻 runs on MacBookPro M-series
🧪 break it plz
🧠 LoRA finetune in < 1h
📷 Sm = faster, Medium = qualityer
⚡ 59x realtime on M5 Pro

One-liner fast install:

curl -LsSf https://t.co/fdFBiUC4PP | bash

[TechCrunch @TechCrunch]

Clouted wants to take the guesswork out of making short videos go viral https://t.co/l6Ej6YsxUU

[Eric Aaberg @ItsEricAaberg]

In a world where short-form video and attention is the modern currency of marketing, Clouted takes out the guesswork for making videos go viral.

Check out our interview with @TechCrunch covering our $7M seed round for @CloutedHQ.

[South Park Commons @southpkcommons]

Scenes from this weekend's embodied AI hackathon 🤖

[Eric Aaberg @ItsEricAaberg]

this is ACTUALLY INSANE to see as a marketer

building the intelligence layer for your content at scale not only means you're posting smarter, but you're saving yourself time from the manual processes/avoiding poor performing formats.

@CloutedHQ is changing marketing forever.

[Wesley @Ambani_Wessley]

What if you could write your own future by controlling attention?

Absolute cinema.

[noomb00m @noomb00m]

@xNamida_music @CloutedHQ @slow @FairlaneMusic we love to see it!

[noomb00m @noomb00m]

@astereeya_ @CloutedHQ @slow 🕐🕑

[noomb00m @noomb00m]

@KaiteeShiks @CloutedHQ @slow this one hit different