
Currently spinning up GPUs as intern at NVIDIA 💚 PhD candidate in geometric deep learning and subquadratic operators at @AmlabUv w/ @erikjbekkers & @egavves
Amsterdam
Joined November 2010
- Tweets14.3K
- Following328
- Followers478
- Likes674
Pinned Tweet
🥠 Excited to introduce our latest work on Equivariant Neural Fields (ENFs)! Grounding conditioning variables in geometry 🚀
Paper: arxiv.org/abs/2406.05753
Github: github.com/dafidofff/enf-jax
Project Page: dafidofff.github.io/enf-jax/
Details below 👇👇
I'm excited to share that I've started as a Research Scientist Intern at NVIDIA in the BioNeMo team! 💚
A huge thank-you to @sgpaliwal for bringing me on board and making this opportunity happen. I'm really grateful for the support throughout the process.
I'm also super excited to be working with @FarhadRGhorbani and @oliviaviessmann on subquadratic operators for natural data. I'm looking forward to learning a lot, building cool things together and making the most of the next few months!
#NVIDIA
David Wessels retweeted
Our survey on continuous diffusion and flow language models is out! ❤️
A little animated tour of what’s inside 👇 From representations and training to few-step generation, scaling, and open problems.
continuous-lm.github.io
New survey on Continuous Diffusion and Flow Language Models is out! 🌊💬
One of my first papers, back in early 2025, was on continuous simplex flows. After working on discrete diffusion, I wanted to try controllable sequence generation with flow matching, which naturally operates in continuous spaces. But at the time, I found them much harder to train well than discrete diffusion. 🥲
2026 has changed this picture completely! 🚀 It’s been super exciting to watch continuous LMs take off, but with the rapid advancements, it’s also becoming harder to keep up with all the new work across pre-training, few-step generation, guidance, and scaling 🥹
So together with @chandavidlee, @FEijkelboom, and @jw9730, we put together a survey that consolidates the continuous LM literature and its key insights into 37 pages. It covers everything from the early work before the so-called “winter” of continuous LMs to the wave of new papers in 2026 🌟
More in thread! 🧵
I would really recommend people to consider the following position! Nabil is great researcher and a super fun to work with!! Don’t miss the chance!!
I am hiring multiple PhD positions at AMLab at the University of Amsterdam to work on deep learning from a physically inspired point of view.
come work with me!
(link below)
1/n
1
6
793
David Wessels retweeted
I am hiring multiple PhD positions at AMLab at the University of Amsterdam to work on deep learning from a physically inspired point of view.
come work with me!
(link below)
1/n
David Wessels retweeted
Finally able to share what we have been cooking during my time at @cartesia. Tokenization-free models are the future!
We've been working on this on several modalities. I strongly believe that the highest gains will come from natural data modalities such as images and video (see end of video for a small teaser on those results).
I am extremely proud of the team! 🚀 💚
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.
Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?
Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
♾️ LiFT: Loop Flow Transformers 🧵👇
Diffusion models usually spend more inference compute by taking more sampling steps. What if we could also spend compute inside each step?
LiFT turns part of a DiT into a recurrent core, letting us scale depth at inference with only a small architectural change. On ImageNet, it beats dense DiT-XL/2 with 3.34 lower FID, while using 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
Jointly led with Mohammad Mahdi Derakhshani 👑(@mmderakhshani)
And amazing supervisors Cees Snoek (@cgmsnoek), Jan-Willem van de Meent (@jwvdm) and Gertjan J. Burghouts (@gjburghouts)
David Wessels retweeted
1/8 💎 Introducing Packora: a state-of-the-art generative model for molecular crystal structure prediction (CSP).
🧩 Flexible, optional conditioning
🏆 Two-track evaluation + leading performance
🔬 Extensive, systematic ablations
David Wessels retweeted
The flow-based language model agenda continues 🌊
Amazing work by @ayhozade showing how models can learn to reason through iterative denoising.
Can neural networks learn to think by denoising?
Excited to share Thinking with Looped Flows!
We train recurrent reasoning with local denoising objectives. It is a new approach for reasoning that achieves SoTA on ARC-AGI-1 and 2 among looped models.
In collaboration with @chandavidlee, @FEijkelboom, @nmboffi, @ismaililkanc, @jw9730.
1/🧵
David Wessels retweeted
🎉 Excited to share that RoVE was accepted as a spotlight at @TAGinDS!
RoPE made attention scores position-aware but left the value pathway position-blind. RoVE fixes that with a parameter-free tweak, and it turns RoPE attention into an 𝐚𝐭𝐭𝐞𝐧𝐭𝐢𝐯𝐞 𝐜𝐨𝐧𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧🤘
David Wessels retweeted
I have been using Weather Lab to plan my daily activities for some time now, and it's amazing!
We're announcing a major update to Weather Lab, our interactive website for sharing Google’s AI weather models from @GoogleDeepMind and @GoogleResearch: deepmind.google.com/science/…
It’s showtime at #ICML2026! 🚀 Really excited to be presenting our paper, "Platonic Transformers," today. We’ve developed a novel equivariant architecture that is finally, properly scalable. Come find me, @Mohammad_Niaz94 and @SharvVadgama at Hall A poster #2614!
The core idea? By embedding "platonic" structural priors directly into the Transformer, we’ve unlocked a way to leverage geometry without sacrificing computational performance. We’ve successfully bridged the gap to make equivariant models truly robust, efficient, and ready for real-world scaling. 📐✨
I’m around the conference all day to chat about scalable architectures with the right inductive biases, or about the fact that we’re already working on the full scaling laws for this approach! 📈 Swing by my poster at to hear the latest. Paper here: arxiv.org/abs/2510.03511 #MachineLearning #ICML2026 #scalable #architectures
Very proud on my master student Jesse Wiers, presenting our work on ‘Geometry-Aware Cardiac MRI Representation Learning’ which got accepted as an oral on MIDL!
He developed this during his master thesis and just got graduated. I could not attend due to ICML, but if you are interested in geometry grounded methods for the medical domain and attending MIDL let me know and I bring you I contact!
David Wessels retweeted
Drop by our poster @ICML to learn about scalable equivariant Transformers that retain the speed and simplicity of standard Transformers.
Jul 9, '26 • 5:00 – 6:45 PM, HALL A #2614
Here is the core idea behind #PlatonicTransformers and why the results are exciting.🧵 (1/n)
David Wessels retweeted
🇰🇷Excited to present a solid choice for equivariance at #ICML2026
Come say hi to @Mohammad_Niaz94 @Dafidofff and me !
It’s fast! 🐦 It’s equivariant! ✈️
It’s Platonic Transformers 🔷
Due to the powers that be, I am unable to join in person at #ICML2026🇰🇷 next week but go check out our poster on Platonic Transformers!!!
> July 9 (thursday)
> HALL A, #2614
> 5pm - 645pm KST
Look out for my fellow ✨Equivariance Evangelists✨ -- @Mohammad_Niaz94 @Dafidofff @erikjbekkers -- who'll be there to present our work
Will try not to FOMO too much ;)
David Wessels retweeted
Due to the powers that be, I am unable to join in person at #ICML2026🇰🇷 next week but go check out our poster on Platonic Transformers!!!
> July 9 (thursday)
> HALL A, #2614
> 5pm - 645pm KST
Look out for my fellow ✨Equivariance Evangelists✨ -- @Mohammad_Niaz94 @Dafidofff @erikjbekkers -- who'll be there to present our work
Will try not to FOMO too much ;)
Exciting update –– Platonic Transformers has been accepted to #ICML2026 🥳
See you this summer in Seoul 🇰🇷 HMU if you're joining in person, lots of touristing to be done ;)
More details coming soon, stay tuned!
Ever had your coding agent happily fire off a training run, only to have it stuck in a queue because every H100 is booked for the next 4 hours? 😅
I built an MCP server to fix this. Now, your agents get a read-only view of your SLURM clusters to check availability before hitting submit! 🧵
But the real magic: it speaks MCP! 🪄
This means Claude Code, Cursor, or your own custom multi-agent setups can dynamically ask, "Where are the free H100s?" before submitting. If the H100s are jammed, the agent sees the A100s are wide open and pivots automatically. Zero babysitting required. 🙌
2
3
105
Worried about safety? It’s strictly read-only (just parsing squeue/sinfo/nvidia-smi) and never touches your active jobs. 🔒
Quick start: pip install cluster-job-monitorCode & Docs: github.com/Dafidofff/cluster…
Would love to hear your thoughts, and a ⭐ makes my day! 🙏 #SLURM #CodingAgents #MCP

























