𝔐ᴜᴛᴜᴀʟɪsᴍ𝔛 is officially out on @_otherpeople via all streaming services and as 2xLP ! The making of this album been a planetary-spanning journey where a generative music collaboration emerged through time and space via technologies of “sonic becoming” with @semilla_ai

Live Music Diffusion Models (Novack et al. ) Can the diffusion ecosystem, offline by nature, be transmuted into streaming? They used it live as a “generative delay” transforming a musician’s improvisation on a gaming laptop. 🔗 github.com/ZacharyNovack/liv…

LiveBand (Sony CSL × QMUL, June 2026) @SonyCSL @QMUL Listens to your live audio and plays along, causal, zero lookahead, runs on consumer hardware. A causal transformer trained adversarially so that training and inference are the same computation. 🔗 arxiv.org/abs/2606.0380

The big release: Magenta RealTime 2 from @GoogleMagenta @Ilaria__Manco @jesseengel An open-weights live music model you play like an instrument: ❉ steered by text, audio examples AND MIDI, injected every 40ms frame ❉ ships with apps DAW plugin 🔗 magenta.withgoogle.com/mrt2

Firstly, our research: 𝕮𝖔𝕯𝖎𝕮𝖔𝖉𝖊𝖈-𝕱𝖑𝖔𝖜 ❉ A Flow Matching DiT that synthesizes audio inside CoDiCodec’s continuous latent space, block-causal, for musical and improvisation. It trains & runs in realtime on an Apple Silicon laptop. 🔗 github.com/moiseshorta/CoDiC…

We wrote some thoughts and a survey on the new wave of realtime AI music tools. Check it out here 🔊

Meet LiveBand: a real-time AI jamming companion! 🎸 It generates live music accompaniments with zero perceived latency ⚡️ It runs locally on Macbooks, can generate any instrument (more than one at a time), is wildly robust, and is trained from scratch on a single GPU! 🧵👇

I'm also so so happy to see a growing community around local & real-time models that let us turn new tech into genuinely exciting and fresh creative ideas - @dadabots @zacknovack @hexorcismos to name just a few

Super interesting paper, the key innovation here being that there's absolutely no 'hidden latent coordinates' but just natural language serving as an encoder and then decoded back into sound with Python scripts via coding LLM's...

Real-Time Neural Audio Synthesis with CoDiCodec-Flow, generating from a custom model, trained in about 4 hrs. and generating on an M4 Macbook Pro. For those wanting to experiment with it , the code is live performance ready! New software coming up via @semilla_ai 🔜🔜🔜

Just added a Google Colab notebook to the CoDiCodec-Flow repo. You can train your own real-time neural audio synthesis model with about 2-3 hrs. of audio and get results in about 3 hours of training time. Small Data models > Corporate Big Tech models github.com/moiseshorta/CoDiC…

🔥New @ISMIRConf paper alert: CoDiCodec – a unified neural audio codec producing both continuous embeddings and discrete tokens from the same model. 👉 pip install codicodec It outperforms existing continuous and discrete codecs in audio quality (FAD, FAD_clap)! ..by the great and only one @marco_ppasini 💪 CoDiCodec offers - Continuous (~11 Hz) discrete (2.38 kbps) latents - FSQ-dropout: improves continuous decoding while keeping discrete tokens useful - Autoregressive & parallel decoding @SonyCSLParis @SonyCSLMusic #codec

Why is no one talking about relation between LLM coding agents and ADHD?

In an alternate timeline we’d be using Evangelion GUI designs rather than CLIs

this started with a striking PC1 falling out of persona space my main insights from the past few months: ⊹ “distance from the Assistant” is the main axis of persona variation across these models e.g. the most relevant thing seems to be “how Assistant-like is this persona” ⊹ this axis already exists in base models and steering with it makes them speak from the POV of helpful archetypes like therapists, coaches, and consultants ⊹ not all personas far from the Assistant are bad! the risk comes from departing the more predictable territory of post-trained behaviour still have a lot of questions about what to anthropomorphize, what to treat as fundamentally alien…