hypermedia production and generative ASCII art
Rochester, NY
Joined May 2015
- Tweets16.5K
- Following289
- Followers1.4K
- Likes15.2K
Matthew(マシュー)🥭Hendricks retweeted
1/10 🔮 Can a world model learn to make its own future more predictable?*
In Latent-Foresight, we jointly learn a compact latent representation of visual foundation model (VFM) features and a generative dynamics model that predicts their future.
Here's how 👇
Matthew(マシュー)🥭Hendricks retweeted
"did you read the new @ThinkWiselyMatt post?"
"no, did you?"
"no"
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
Matthew(マシュー)🥭Hendricks retweeted
Uploaded the camera-ready version of our #NeurIPS2026 paper:
"Generalized Wasserstein Flow Matching:
Nested Couplings for Point Cloud Generation"
→ Multisample Flow Matching in Wasserstein space
📄 arxiv.org/abs/2605.08424
with R. Duong & G.Steidl :)
Matthew(マシュー)🥭Hendricks retweeted
Final version is out:
onlinelibrary.wiley.com/doi/…
"Reservoir Computing for Pattern Recognition and Time-Series Prediction Based on Nanofluidic Negative Differential Resistance"
Matthew(マシュー)🥭Hendricks retweeted
Five hundred birds fly at one constant speed. At each step every bird turns a tenth of the way toward the mean heading of the birds within radius R, plus a random wobble of width η.
No bird sees the whole flock. For small η the local rule still settles on one shared heading, and the order φ = |(1/N) ∑ e^(iθₖ)| climbs from 0.02 to 0.99. Past a critical η the order is lost.
Vicsek and coauthors described this phase transition in 1995 naming it Vicsek Model.
Matthew(マシュー)🥭Hendricks retweeted
I have a mild obsession with symmetry and took some time to write-up a blog post about my work on creating an atlas of dynamic simulations of all 2D wallpaper groups. So visually cool and maybe you'll learn something new about symmetry.
diffuse.one/p/b1-001
Matthew(マシュー)🥭Hendricks retweeted
World models that see only through the players' eyes forget what happens out of view.
OverSeer adds a few fixed cameras and generates every view together, so the world stays consistent over time and across players.
📄 arxiv.org/abs/2610.28554
🌐 runjiali-rl.github.io/OverSe…
🧵👇
Matthew(マシュー)🥭Hendricks retweeted
Can a neural network learn ~2^100 LDPC (Low-Density Parity-Check) codewords from just 10-20 examples?
👉 Grokking in LLMs and memorization-to-generalization transition in diffusion models are related phenomena — the model learns true process of data generation from a small number of training samples.
👉 It is hard to study this phenomenon analytically on natural datasets.
👉 LDPC codeword generation is a precise mathematical process in which the codewords are constructed as solutions to parity check equations.
👉 We train a Dense Associative Memory with a tiny fraction of codewords from the full vocabulary, without giving the network the generation algorithm itself. The model manages to learn the entire vocabulary on its own. It turns out, you only need log(n) samples to learn the entire space of 2^n codewords.
👉 The network only needs approximately P=10-20 training examples to correctly learn the full vocabulary containing ~2^100 codewords.
Our new paper with Andy Lucas: arxiv.org/abs/2610.27384.
Let us know what you think!
Matthew(マシュー)🥭Hendricks retweeted
Yann LeCun has changed the game for robotics.
Research he co-authored tackles a problem inside AI world models: their internal representations can make planning harder than it needs to be.
A world model predicts what happens after an action. But even useful visual features can form curved paths in its internal “map,” making distance to a goal misleading.
The researchers introduced Temporal Straightening.
During training, it encourages those internal paths to become straighter, so the planner gets a more useful measure of progress.
The results:
More stable gradient-based planning
Higher success rates across tested goal-reaching tasks
Code available for others to explore
The approach changes how the model represents movement, rather than rewriting every RL algorithm to work in curved space.
For robotics, that’s a useful direction: build world models whose internal maps help an agent choose its next move.
Matthew(マシュー)🥭Hendricks retweeted
f=0;r=240;s=.7*r;b=255// #p5js
draw=_=>{f||createCanvas(W=2*b,W)+noStroke()
background(0,f?8:b);d=.39*cos(f++/W)
for(u=0;q=sin(u),p=cos(u),fill(m=b-p*b,c=b-q*b,b+c-m),u<TAU;u+=.4-d)
for(v=u;j=sin(v+u),k=cos(v+u),v<u+TAU;v+=.4+d)
rect(b+r*q*k,b+s*q*j+s*p,2+q*j)}
Matthew(マシュー)🥭Hendricks retweeted
If you don't know about the Busy Beaver rabbit hole, our new video is for you.
Released yesterday and it's at 100k views already!
Matthew(マシュー)🥭Hendricks retweeted
Two identical sources send out waves. Where a crest from one meets a trough from the other, they cancel; where crest meets crest, they double.
The cancelling lines sit wherever the two path lengths differ by half a wavelength, so moving the sources apart from 0 to 5 wavelengths opens 10 silent lines.
The tone is what a listener at the black dot would hear: loud on the bright bands, silent on the lines.
Matthew(マシュー)🥭Hendricks retweeted
I created a page to gather all the demos I built with Opus 5.5, GPT 6 Astra etc, go check it out if you want to take some inspo or just look through some of them!
leonlin.de/demos
Matthew(マシュー)🥭Hendricks retweeted
The hardest problems in dynamics become ordinary multiplication the moment you enter the s-plane.
In time the output is a convolution of input and impulse response. The Laplace transform collapses that integral into the simple product Y(s) = H(s) X(s).
From the unit step whose transform is 1/s to the sinusoids living on the imaginary axis, each pair maps a location in the complex plane to a precise behavior in time.
What once required solving differential equations now requires only algebra and a table lookup.
Matthew(マシュー)🥭Hendricks retweeted
I had Claude Opus 5.5 make a music video about agent swarms
youtube.com/watch?v=7KWI3f77…
#swarm #musicvideo #ai #pdoom
