iiinova — where your digital self keeps evolving

Jordi Pons @jordiponsdotme.bsky.social · 05/10/2026

my wife and i spend a lot of time drawing cars for our kid so i put those drawings in a music game based on my latest song AI-generated sounds, hand-drawn cars, and a chicken that dies +info: artintech.substack.com/p/playing-wi... +game: jordipons.me/apps/die/

031

Jordi Pons @jordiponsdotme.bsky.social · 21/09/2026

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: artintech.substack.com/p/undergroun...

000

Jordi Pons @jordiponsdotme.bsky.social · 14/08/2026

wrapping up CULOCODO it's been my first big artistic project, but now it's time to move on to the next chapter we've compiled everything into one website: jordipons.me/apps/culocodo/ 4 interactive ai music interfaces 2 performances 1 videoclip 2 posts

030

Jordi Pons @jordiponsdotme.bsky.social · 14/08/2026

Music Arena: beta.music-arena.org Stable Audio 3, explained in 5 figures: artintech.substack.com/p/stable-aud... Huge shoutout to the Music Arena team♥️

artintech.substack.com

Stable Audio 3, explained in 5 figures

A walkthrough of how it works, what it can do, and why it runs on your laptop.

020

Jordi Pons @jordiponsdotme.bsky.social · 14/08/2026

Fresh results from July 2026 at Music Arena. Happy to see the community is finding Stable Audio 3 Medium as useful as we do. It has been independently evaluated, and Medium is the second-best and fastest model for instrumental music generation, with open weights!

120

Jordi Pons @jordiponsdotme.bsky.social · 29/07/2026

this thumbnail is for those who call "football" football

040

Jordi Pons @jordiponsdotme.bsky.social · 28/07/2026

I have absolute pitch (except when I don’t). After reviewing dozens of papers, I realized my struggles are common and are well documented in the scientific literature. I’m sure many of you are experiencing the same thing! artintech.substack.com/p/absolute-p...

artintech.substack.com

Absolute Pitch Impostor Syndrome

On why I both have and don’t have absolute pitch

000

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

Repeat with me: agents are for writing better, not for writing faster. In this post I share my workflow for writing scientific papers with LaTeX and AI agents. artintech.substack.com/p/agents-for...

artintech.substack.com

Agents for Writing Better, Not Faster

My workflow for writing scientific papers with LaTeX and AI agents

020

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

In the rest of the post, I share how I used agents to create LaTeX figures and tables, support my literature review, manage references, and catch typos. ▶️ artintech.substack.com/p/agents-for...

artintech.substack.com

Agents for Writing Better, Not Faster

My workflow for writing scientific papers with LaTeX and AI agents

110

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

Guided agents reliably get 80% of the work done, but the final edits belong to me. The last 20% is what some call “unslopifying”. I tried hard to automate that last 20%, and I’m not sure it’s worth the effort.

110

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

Why better not faster? Because you’ll still have to review everything. Every paragraph, every change, every citation, every equation, every figure, every table.

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

The agentic loop I use: ▶️ Start with a draft. ▶️ One agent review it as a top machine learning conference reviewer ▶️ A second agent rewrite another draft based on the review, acting as a scientific writer ▶️ Repeat the review–rewrite cycle until the reviewer is satisfied

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

Clarity in scientific writing means.. ▶️ Designing figures that tell a story ▶️ Building tables that highlight the important numbers ▶️ Finding related work that frames your contributions ▶️ Fixing awkward sentences ▶️ Removing typos

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

But if your use agents for writing papers faster, you’ll probably end up with a generic article that fails to communicate the nuances of your contribution.

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

If you don't have anything to say, agents won't fix that. If you do, they can help you convey your ideas more effectively.

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

TL;DR. Let agents handle the parts you struggle with, so you can write better. Keep doing what you enjoy and are good at, or agents won't make you faster.

100

Jordi Pons @jordiponsdotme.bsky.social · 30/06/2026

Agents for Writing Better, Not Faster After spending the last months experimenting with agents for writing, I think scientific writing has never been bottlenecked by typing speed. It's bottlenecked by our ability to communicate ideas clearly. 🔽🔽🔽

artintech.substack.com

Agents for Writing Better, Not Faster

My workflow for writing scientific papers with LaTeX and AI agents

110

Jordi Pons @jordiponsdotme.bsky.social · 21/05/2026

How to get the weights from Huggingface? huggingface.co/settings/tok... + Create new token. Set permissions → Read access to contents of all public gated repos you can access → Save. export HF_TOKEN=<token> Install repo using uv and run: uv run python run_gradio.py --model medium

020

Jordi Pons @jordiponsdotme.bsky.social · 21/05/2026

stability.ai/stable-audio but actually runs on your macbook with this one-liner: curl -LsSf dadabots.com/_/sa3-mac | bash

120

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

Runs on your machine. · small runs on a MacBook Pro M4. It takes 5.92s to generate 120s of audio; with CoreML acceleration, this drops to 3s. · medium requires ~6.5 GB VRAM for 120s of audio. It fits on consumer-grade GPUs and generates about 6m of audio in a few seconds.

010

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

Generating audio in one step is challenging. We use ping-pong sampling instead: denoise, add a small amount of noise, denoise again, repeat with less noise each time. We run eight iterations in total.

110

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

Trained in three stages: 1. Flow matching pre-training (base model): slow, requiring 50–100 steps per generation. 2. Distillation warm-up: fine-tunes the model toward one-step generation. 3. Adversarial post-training: improves sample quality and enables faster inference.

100

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

One-shot generation is cool, but editing is what enables artists to execute their vision. Stable Audio 3 supports audio editing via inpainting: · Single-segment edits · Multi-segment edits · Continuations

100

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

Fixed-length generation allocates Lmax embeddings regardless of the requested duration d, wasting compute on padding for short clips. Variable-length generation allocates L embeddings proportional to d, so compute scales with duration.

100

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

The model takes three inputs: a text prompt (“a jazz piano solo”), the duration in seconds, and optional editing controls. It then generates stereo audio at 44.1 kHz.

100

Jordi Pons @jordiponsdotme.bsky.social · 20/05/2026

Stable Audio 3, explained in 5 figures. It’s a family of open-weight models for generating instrumental music and sound effects. The models are fast, support editing, and are trained on licensed and Creative Commons audio. 👾 artintech.substack.com/p/stable-aud... 🏋️‍♂️ github.com/Stability-AI...

231

Jordi Pons @jordiponsdotme.bsky.social · 19/05/2026

The image above is an original, hand-crafted table from one of our 2018 papers. "End-to-End Learning for Music Audio Tagging at Scale" arxiv.org/pdf/1711.02520

010

Jordi Pons @jordiponsdotme.bsky.social · 19/05/2026

As a LaTeX table artist, my time is up. What used to take me a full day of work now takes minutes. Soon, no one will remember the hours we agonized over a single misaligned &. Pour one out for \begin{tabular}. We had a good run.

130

Jordi Pons @jordiponsdotme.bsky.social · 07/04/2026

New interview is out! Music as Software artintech.substack.com/p/music-as-s... Thanks @teropa for the insightful and engaging conversation around your artistic practice, wellbeing music, and the history of generative music.

artintech.substack.com

Music as Software

Tero Parviainen on the web as a medium for generative music.

000

Jordi Pons @jordiponsdotme.bsky.social · 11/03/2026

new AI music lawsuit independent musicians accusing google www.loevy.com/wp-content/u...

020

Jordi Pons @jordiponsdotme.bsky.social · 26/02/2026

One friend: “The bass sound design is boring”. Another friend: “The track feels like it’s missing something, you could take it in a different direction”. AI remix made a track with a better sounding bass, with additional layers of intricate melodies and arpeggiators. OK

000

Jordi Pons @jordiponsdotme.bsky.social · 26/02/2026

I’m learning to produce music and regularly share my tracks with friends for feedback. Now I use AI remixes as an additional source of feedback (that often aligns with my friends’ opinions). The advantage is that you hear the feedback translated into specific musical examples.

100

Jordi Pons @jordiponsdotme.bsky.social · 26/02/2026

Two unconventional uses of AI REMIXING for music production: Toplining → Explore new melodic ideas when you’re stuck. Not everyone can afford a professional topliner. Feedback → AI remixes can reveal what’s missing or overdone. Similar to sharing an unfinished track for notes.

010

Jordi Pons @jordiponsdotme.bsky.social · 24/02/2026

i'll let my agent answer you in python

000

Jordi Pons @jordiponsdotme.bsky.social · 24/02/2026

Gradio synths are the new VSTs

000

Jordi Pons @jordiponsdotme.bsky.social · 23/02/2026

I cannot pronounce claud clode clod claud cod claut caud clod cloud code but you know what I mean.

131

Jordi Pons @jordiponsdotme.bsky.social · 12/02/2026

Most parameter combinations in synthesizers sound like shit. That flexibility is what makes new sounds possible. But if we release a neural synthesizer and one prompt doesn’t sound as you expect, you’ll be mad at us. So we’ll just keep the new neural synthesizers to ourselves.

000

Jordi Pons @jordiponsdotme.bsky.social · 10/02/2026

The ACE-Step 1.5 paper can be confusing. In this post I share are its main ideas: artintech.substack.com/p/ace-step-1... 1. DIFFUSION MODEL: supports multiple tasks. 2. LANGUAGE MODEL: reprompting & semantic tokens generation. 3. DATA PREPARATION: 27M songs. 4. OPEN WEIGHTS: supports LoRAs.

artintech.substack.com

ACE-Step 1.5 Explained

The ACE-Step 1.5 paper can be confusing. Here are its main ideas.

010

Jordi Pons @jordiponsdotme.bsky.social · 10/02/2026

Dopamine hits now come from agents building useless apps instead of endless scrolling.

030

Jordi Pons @jordiponsdotme.bsky.social · 03/02/2026

HeartMuLa Explained It is an open-source music model that claims to perform on par with Suno. 1. HeartCodec: low-rate audio tokens 2. HeartMuLa: autoregressive model 3. HeartTranscriptor: lyrics tokens 4. HeartCLAP: text & reference-audio tokens artintech.substack.com/p/heartmula-...

artintech.substack.com

HeartMuLa Explained

Does HeartMuLa outperform Suno in music generation?

000

Jordi Pons @jordiponsdotme.bsky.social · 03/02/2026

Love that the AI music conversation is moving toward interactive music and fans-driven art. - artintech.substack.com/p/interactiv... - www.youtube.com/watch?v=U8dc... - aiformusicworkshop.github.io - www.universalmusic.com/universal-mu...

000

Jordi Pons @jordiponsdotme.bsky.social · 15/01/2026

The future of AI music won’t shape itself. It’s up to us to imagine what it can become. If its current state feels like ‘slop,’ consider creating something different yourself. While this essay has focused on Interactive AI Music, I’m sure there’s still much more to discover.

000

Jordi Pons @jordiponsdotme.bsky.social · 13/01/2026

AI music doesn’t have to be ‘slop’. It can be Interactive! Over the break, I wrote a post defining Interactive AI Music, a term that brings together my favorite artistic projects in AI and music. artintech.substack.com/p/interactiv...

000

Jordi Pons @jordiponsdotme.bsky.social · 29/12/2025

New written interview! “On Shared Musical Agency” with Martin Heinze (marts~) artintech.substack.com/p/on-shared-...

artintech.substack.com

On Shared Musical Agency

Interviewing marts~ to learn about his story and artistic vision.

010

Jordi Pons @jordiponsdotme.bsky.social · 23/12/2025

AI art is criticized for feeling soulless, lacking relatability and human narrative. But AI art does not need to be like that. I strongly believe that AI can be used to create art that was previously impossible and that it will be equally appealing as previous art forms.

040

Jordi Pons @jordiponsdotme.bsky.social · 22/12/2025

New post is out! This time, I go through Meta's new model for source separation. "SAM Audio Explained" artintech.substack.com/p/sam-audio-...

artintech.substack.com

SAM Audio Explained

Wouldn’t it be cool to separate any audio using a single model?

020

Jordi Pons @jordiponsdotme.bsky.social · 16/12/2025

The key issue in AI art is agency. We need workflows that produce art that clearly feels like it’s from the artist.

020

Jordi Pons @jordiponsdotme.bsky.social · 12/12/2025

The concept of interactive AI music is gaining traction.

000