phd student @PrincetonVL. formerly: bse @princetoncs, research intern @elorianai
Joined March 2025
- Tweets161
- Following439
- Followers661
- Likes619
david yan retweeted
I spent a decade building foundation models at Google and Apple. But what our small team at Elorian accomplished with just a few hundred GPUs might be the most exciting leap yet. Meet Glimpse -- Our frontier-level multimodal model at platform.elorian.ai
🚀🚀🚀 @ElorianAI 🚀🚀🚀
david yan retweeted
Today we're thrilled to announce the release of our first frontier visual reasoning model
amazing team, amazing model :)
david yan retweeted
On results, Glimpse is competitive on the frontier across multimodal benchmarks and dominates Flash-sized models.
david yan retweeted
Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that.
Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks.
david yan retweeted
1/ I’m excited to share our work on “Depth as Time in One-Step Generative Models”!
alphaxiv.org/abs/2610.03626
TL;DR: We ask, when a model learns to generate an image in one forward pass, what happens to the denoising process? We observe that process unfolding over depth!
kids these days will never understand the pain of
RuntimeError: Operator xyz failed, context is incorrect
It's insane these days that the way 3D will likely end up being solved is through LLMs.
For the last decade+, there's been a whole line of work building specialized models to generate 3D objects (see GET3D, Magic3d). But now, language models and coding models are powerful enough that you can just ask it to write code in Blender, and it'll figure out how to create the objects, scenes, and textures.
It's also quite mind blowing that you don't need to use any texture files and they can do this from scratch. Models can just rely on code that generates randomness in color patterns and in the lighting, which is enough to make the objects look textured and realistic.
All of these scenes are generated by language models writing code, positioning the camera, lighting, and applying the materials to the objects to get good views, each in a couple hundred lines of Blender Python code!
david yan retweeted
We’re introducing Humanity’s Sixth Sense, a new benchmark testing the intuitive visual reasoning people rely on every day.
Across 522 open-ended tasks spanning images and video, HSS tests everything from spatial and causal reasoning to social understanding.
The gap is significant. Humans score 93.1%, while the strongest model, GPT-6-astra, reaches 53.6%. The median model scores just 30.9%.
david yan retweeted
The most profound paper of my PhD so far. We truly did something special to make LMs play/explain chess, needing both architectural & algorithmic innovation (natural-language analogue of the Alphazero algorithm) + applicable to many other domains. Please read on and share!
1/10
david yan retweeted
Task and motion planning (TAMP) excels at multi-step, semantic, and geometric reasoning. VLAs excel at deformable object manipulation, which TAMP can't do. Can one VLA do both?
DATAFARM fine-tunes VLAs on TAMP demos aligned to the pre-training distribution.
(1/9)
We traditionally put value on being "person who came up with x method that solves age-old task y." So people spent most of their time coming up with better x's. AI has left us in a shortage of y, so we ought to spend time converting ever more ambitious goals into concrete y's.
Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.
Perhaps counterintuitively, I think it's a great time to do so.
I wrote up some thoughts on this here: web.mit.edu/phillipi/www/wri…
