Kyle Sargent

Vision-language models are getting better every day. Can we use them to improve image compression? Yes! For my internship, working w/ @GoogleDeepMind, @GoogleResearch, we designed VLIC, a diffusion autoencoder post-trained with VLM preferences. Our preprint is out today! A🧵:

My personal benchmark for voice AI is: when am I going to be able to improve my Chinese by talking to it in a mixture of bad Chinese and English, interrupting it frequently for clarifications and definitions, switching often between languages, etc.? So far nothing works at all

@volokuleshov @chenhao_chao Same question, seems like kind of a free lunch?

@chenhao_chao Question: what happens if you apply the subtokenizer to the AR model? Does NLL improve similarly? Is it possible that the experiments are effectively comparing models with different tokenizers?

(1/7) We introduce MDM-Prime-v2 which scales 21.8× better than autoregressive models (ARMs) in compute-optimal comparisons. 📎 Paper: arxiv.org/abs/2603.16077 🌟 Blog: chen-hao-chao.github.io/mdm-prime-v2 ⌨️ Github: github.com/chen-hao-chao/… Here’s how we did it👇:

@Logo_Daedalus Eddington was an astonishing film. Insane snub

So they didn’t nominate Eddington for anything but we’re supposed to take this seriously as some ceremony of cinematic legitimacy granting

Spatial reconstruction is a long-context problem: real scenes come with hundreds of images. But O(N²) transformer-based models don’t scale efficiently. Introducing: 🤐ZipMap (CVPR ’26): Linear-Time, Stateful 3D Reconstruction via Test-Time Training (TTT). ZipMap “zips” a large image collection into an implicit TTT scene state in a single linear-time operation. The state will then be decoded into spatial outputs, and can be queried efficiently for novel-view geometry and appearance (~100 FPS) ZipMap is not only much faster (>20× faster than VGGT), but also matches or surpasses the accuracy of all SOTA models.

Can we not do this please lol? Can we just stick with artificial NNs? They're pretty good now!

#CVPR2026 🤩 PerpetualWonder: interactive 4D scene generation with long-horizon actions. From a single image, build a world you can interact with a sequence of ✨3D actions (point force / wind / gravity). Project: johnzhan2023.github.io/PerpetualWonde… More details in 🧵: 1/5

@Taesung Whoa

Reve's new text-to-image model is here. Really proud of the team to rank at #3 with the big labs. To my knowledge, we are the first lab to use native pixel space diffusion without latent autoencoder at 4k (16MP) resolution for production level image generation.

"b (gh gw) (ph pw c) -> b c (gh ph) (gw pw)"

@TheGregYang @NotionHQ Notion is great but for some reason their desktop version is the only usable version

I'm trying out @NotionHQ for organizing my learnings on lyme & test results feedback so far 1. android app is kinda jank and lags everywhere 2. notion ai is not bad so far but very annoying it can read files I upload in the chat window but can't attach those files to pages

@holynski_ Looks amazing

hey guys, you wanted to try Genie... now's your chance!!!

I’m broadly optimistic about the ability of AI to automate most coding, but there’s something about human care and attention in software that still feels so valuable and important to me. The other day I flipped open my Game Boy Advance SP, which my parents got me in 2003-ish, and started playing Super Mario Bros 3. After 23 years, it still works perfectly. I’ve played the whole game through maybe a dozen times – hundreds or low thousands of hours of overall play time – and have yet to encounter a single bug. There’s maybe one sharp edge, which is that in the Ice World stage, level 5 or 6 (forget which) has a button combination that’s a little challenging to press on the Game Boy Advance SP compared with other platforms on which I assume the game was released. But that’s all I can remember struggling with. AI can write thousands of lines of code for you, but what software do we have now, in 2026, that still works this way? Trying to boot up my PS5 to play a triple-A title is an aggravating bore because you have to deal with mandatory console updates, followed by mandatory game updates, etc.

Maybe but N-dimensional hyperparameter sweeps can get pricey

I love AI coding tools but often feel they are way too aggressive about suppressing errors or introducing footguns like my_dict.get(key, dumb_default). Silent failures are bad! If there's a bug in my code I want to know - show me the traceback!

@HadiZayer Thanks Hadi :)

What I appreciate about this work is how we can optimize compression models on metrics beyond L2 reconstruction error. VLMs provide a neat tool to allow us to optimize for arbitrary goals and objectives

✨ Any static 3D assets ➡️ 4D dynamic worlds. Introducing CHORD, a universal framework for generating scene-level 4D dynamic motion from any static 3D inputs. It generalizes surprisingly well across a wide range of objects 🤯 and can even be used to learn robotics manipulation policy 🤖! Project page: yanzhelyu.github.io/chord. Dive deeper in a 🧵: 1/n

@__nmca__ Are you hinting zoph has a 360 bench? That’s insanely impressive if so

some consider a 360lb bench press unethical

IMO it should be considered quite rude in most contexts to post or send someone a wall of 100% AI-generated text. “Here, read this thing I didn’t care enough about to express myself.”

What if we can simulate an *interactive 3D world*, from a single image, in the wild, in real time? Introducing PointWorld-1B: a large pre-trained 3D world model that predicts env dynamics given RGB-D capture and robot actions. 🌐 point-world.github.io from @Stanford @nvidia

@NotionHQ mobile app really needs some love wow