Jan Van Balen
@jvanbalen
Followers
722
Following
3K
Media
137
Statuses
1K
ML researcher @Spotify • previously, recsys & MIR @AppleMusic @UAntwerpen @UUtrecht • also at @[email protected]
BE
Joined September 2012
RT @jpgard: 🐣 🎶 Excited to share LLark: A Multimodal Foundation Model for Music!. LLark is a 12B parameter model that produces text respon….
0
43
0
NL people: there’s a string of shows coming up where Jameszoo and band, plus the amazing Asko|Schönberg ensemble, will perform alongside a generative music system I helped build ✨ VERY excited for this one. More: I’ll be there for Rotterdam and Den Bosch
0
1
7
Most of all though, this lets you play with python ML music tools in real-time🕺. E.g., a version of this, hooked up to some Markov chains and three Disklaviers, ran more or less continuously for about a month at an art gallery last winter:.
3 more weeks of ‘Telephone’ at @Nestruimte together with Philip Akkerman. Works with @jvanbalen Paul van Twist, Cas Dekker and Ruimtetijd. Installation 📷 by Charlott Markus
1
0
0
Under the hood it's mostly a bunch of async python—so that each of your effects has its own event loop, and won’t be (completely) blocked by other effects’ python-paced calculations. Not lightning-fast but still pretty snappy in practice ✨.
1
0
0
I just published a small package I’ve been using to work with real-time MIDI in Python, for a personal project:. I wanted to make it easy to implement custom MIDI effects that support real-time I/O🎹—including controls like knobs and sliders 🎛🎚.
github.com
Approximately real-time MIDI manipulation for Python, macOS - jvbalen/midifx
1
1
23
RT @r4b1tt: I'm proud to share that our team's work at @SpotifyResearch on audio-lyric alignment by Simon Durand, Daniel Stoller, and @seba….
0
10
0
RT @Jameszooo: 3 more weeks of ‘Telephone’ at @Nestruimte together with Philip Akkerman. Works with @jvanbalen Paul van Twist, Cas Dekker….
0
1
0
one design thing that’s very relevant to the algorithm: showing one full-screen video at a time really helps with attributing implicit feedback to individual videos—without that, explore-exploit is a lot more difficult.
0
0
2
“More generally, in AI applications, the sophistication of the algorithm is rarely the limiting factor. The quality of the design, the data, and the people that make up the system all tend to matter more.”.
TikTok users feel its algorithm is eerily accurate. But from everything we know about the algorithm, it's a pretty standard recommender system of the kind that every major social platform uses. TikTok's real innovation is something else. Just posted:
1
0
3
New release!👇 Some really useful things in here, that I’ve been using for every TF project for a while now.
As if today wasn't already fun enough - my team just open-sourced a new library!. Say hello to realbook, a Python library that makes it easier to train audio-based machine learning models with TensorFlow:
0
0
2
Having done some work on large-scale recommender systems, I don’t see this scale to conversations about millions of entities without a more structured representation feeding into or wired into the model, at least not any time soon (4/n).
1
0
1
But as with so many things GPT, it’s often confidently wrong. This one links to the artist page for Drake: (3/n)
1
0
0
Here’s something that already works (this actually links to the right page!): (2/n)
1
0
0
Bitter but true. However there are compelling reasons for KG’s in particular to be folded back into ChatGPT-level models, in some form. (1/n).
A new episode of the “bitter lesson”: almost none of the research from ~2 decades of dialogue publications, conferences and workshops lead to #ChatGPT. Slot filling ❌intent modeling ❌ sentiment detection❌ hybrid symbolic approaches (KGs) ❌.
2
1
2