Jack XV

CTO @MiuroAI. Building GM-1, a music foundation model, at @miurolabs. 🎵
📍 Seoul / Tokyo. R&B, riverside runs & good cocktails. 🍸

[Jordi Pons @jordiponsdotme]

I just released a new track!

I use AI for sound design:
- Generate samples (stable audio)
- Transform guitars (stable audio)
- Generate percussive textures (moises)

I handle composition and production myself.. because making music is fun :)

+ info: https://t.co/MAlH2T0whP

[Jack XV @jackxv_ai]

@ElevenLabs @ElevenCreative Music inside a timeline is the interesting part. Does the music model see the cut points and target duration, or does it generate a fixed clip you trim? The first means the score follows the edit. The second is a stock library with a prompt box.

[Jack XV @jackxv_ai]

@yu_ichi_suzuki Loss bottoming at 400 then climbing is the cleanest signal nine minutes of compute can buy. The 2,000-step result is the interesting one: adapters tend to memorize timbre before melody, so a similarity check on stems would catch quoting before your ear does.

[Jack XV @jackxv_ai]

@jordiponsdotme Fair. The one difference: a plugin renders the same tomorrow. A generative transform does not, so the bounce is the take.

[Jack XV @jackxv_ai]

interesting to see Korea putting AI-assisted music registration on the table.

generation is getting easier.
provenance, contribution and rights are becoming the hard product problems.

the music stack isn't finished at inference.

[Jack XV @jackxv_ai]

AI is becoming an orchestration problem.

models are getting better fast.

but in real products, the hard parts are memory, state, permissions, failures, and getting different systems to work together.

that's the part i enjoy building.

[Jack XV @jackxv_ai]

@suno Describing an effect in words is the easy half. The interesting question is what runs underneath: a language model picking knobs on a fixed plugin chain, or the model actually generating the transformed audio. The first is a nicer UI. The second changes what a stem even is.

[MIURO @MiuroAI]

One idea. A whole song. 🎵
Meet the MIURO creator starter pack: GM-1, MusicVault, RightsVault, and a little purple inspiration. 💜
A concept for now. Which piece would you want on your desk? 👇
#MIURO #AIMusic #GenerativeAI

[Jack XV @jackxv_ai]

Made a song at 2 AM about being awake at 2 AM. Korean lyrics, one lamp on, Han River outside. Straight out of our studio, first take, no edits. Tell me where the Korean falls apart. https://t.co/PlqkWgxSUW

[Coop 💿 @Cooopahtroopa]

ソールドアウトとなったCoopのショー @aturamusic at @CIRCUSTOKYO のフルセットを公開できることを本当に嬉しく思います！

このセットは、僕の人生の中でも特に忘れられないセットのひとつになりました。

ATURAにとって初めての東京。そして僕たちにとってもCIRCUSで開催するまだ2回目のパーティーでした！

こんなにも忘れられない最高の時間を一緒に作ってくれたみんな、本当にありがとう！

[Jack XV @jackxv_ai]

@WaisuMaku Ad Lib card to could-pass-for-a-studio in a few months is the whole local music model story in one sentence. Which GPU, and did you hit the VRAM wall or the patience wall first?

[Jack XV @jackxv_ai]

Chuseok week. Half the city leaves Thursday, which means four days of empty roads, closed lunch spots, and nobody pinging me. Best training window of the year.

[Jack XV @jackxv_ai]

@iced_coffee_dev Music got here about two years early. The sound became cloneable fast. What did not: knowing who made a thing, when, and who gets paid for it. If I were building FOO that is where I would dig in, not the artifact, the record around it.

[Jack XV @jackxv_ai]

@Bitcopath @deepseek_ai 12 tracks, 300 epochs is the number I would watch. Adapters on music models lock onto timbre long before they learn arrangement, so the real test is whether the station still surprises you on day three or whether everything drifts toward the same three songs.

[Jack XV @jackxv_ai]

Allaire says the spectacle era of crypto is over and chains are becoming invisible infrastructure. Agree. The best thing that can happen to a chain is that nobody notices it. Same goal for us: the wallet stays off screen.

[Jack XV @jackxv_ai]

@dargor1406 @AIMusicCreatorA Fair. You should not need a theory of post-training to notice your songs changed.

[Jack XV @jackxv_ai]

@onirenaud @iced_coffee_dev The betrayal is in the timing, not the tool. She found out after she had already decided she loved it. If who made it and how had been attached to the track from the second it was generated, there is nothing left to discover later.

[Jack XV @jackxv_ai]

@dargor1406 @AIMusicCreatorA Fewer words for the same idea is a precise description of what preference tuning does to a generative model. The prompt side still parses you fine; the output distribution got sharpened toward whatever raters rewarded. Test: one prompt, twenty seeds, count the arrangements.