Sepehr Mousavi (@sepmsv) header
Sepehr Mousavi (@sepmsv) profile picture

PhD student at @eth_en AI for Science, Operator Learning

Zurich, Switzerland
Joined May 2026
Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)Photo by Sepehr Mousavi (@sepmsv)
[1/4] 🧵 What does a latent space look like when no single task shapes it? Our functional autoencoder is trained on random synthetic functions, frozen, and shared by operators, so its latent space belongs to no task or semantics and can be explored on its own: pick points and paths in it freely and decode them. Every embedding lies on a hypersphere and every point of that sphere decodes to a valid, smooth field, not only the regions near training data. 🎮 Play the Tiles game to get a feeling of it and see if you can break the record: daedalus.ethz.ch/arcade/tile…
[1/6] 🧵 What if physics had its own DINO? Daedalus is a self-supervised backbone for functions on point clouds, trained once on synthetic data and frozen. Operators for fluids, solids and waves learn on its embeddings, faster and at far larger scale. Language and vision moved to shared, pre-trained representations years ago: one backbone, reused by every downstream task. Neural operators mostly still learn their representation from scratch, end to end, for every PDE, geometry and dataset. We think that is a big part of what holds them back at scale, and Daedalus is our take on the alternative: daedalus.ethz.ch #RepresentationLearning #SciML #OperatorLearning
1
1
99
[3/4] Directions carry meanings. Take two fields that differ by one operation, such as a stretch, a blur or the split of one bump into two, and subtract their embeddings. Add that difference to the embedding of a completely unrelated field, and the same edit happens to it. A larger coefficient applies the edit more strongly, and a negative one undoes it. An operation therefore occupies a consistent direction across the whole sphere, not just near the pair it was read from. It is a small hint that meaningful transformations act on embeddings in a simple way, which is what you'd hope for in a space meant to host operators.
1
34
[4/4] Explore the latent space yourself. Four demos let you move between two of your own fields along the sphere, apply an edit to your own field, search a bank of functions by sketching one, and turn an embedding in random directions to see that every variant stays plausible. Demos: daedalus.ethz.ch/#demos Article: daedalus.ethz.ch/interpretab…
30
[1/6] 🧵 What if physics had its own DINO? Daedalus is a self-supervised backbone for functions on point clouds, trained once on synthetic data and frozen. Operators for fluids, solids and waves learn on its embeddings, faster and at far larger scale. Language and vision moved to shared, pre-trained representations years ago: one backbone, reused by every downstream task. Neural operators mostly still learn their representation from scratch, end to end, for every PDE, geometry and dataset. We think that is a big part of what holds them back at scale, and Daedalus is our take on the alternative: daedalus.ethz.ch #RepresentationLearning #SciML #OperatorLearning
1
2
12
1,404
[5/6] On a single RTX 4090 (24 GB), compared with GAOT, the state-of-the-art baseline: • 64x larger training problems (4.2M points at a batch of 32), and 13.6x faster on the largest problem GAOT can still train on • 92x larger models (101M parameters on dense problems), and 4.9x faster at the largest GAOT that fits • inference with a 1.6B-parameter processor on 32 dense problems at once, or on 537M points in one go with chunked inference • latent rollouts up to 10x faster than decoding and re-encoding at every step And accuracy holds: on a benchmark of 28 problems across fluids, solids, acoustics and elliptic PDEs, Daedalus is more accurate than the SOTA baseline on 23, while training significantly faster.
1
2
76
[6/6] You can try Daedalus yourself on our website. Nine live demos run the frozen backbone on your own formula, sketch or photo: reconstruction, upsampling, inpainting, outpainting, exact derivatives, smooth transitions, latent edits, latent search over a bank of 2,300 functions, and latent perturbations. Try our three small games to test your eyes on functions: Freaks, Tiles and Divide and Conquer, and see where you land on the leaderboards. 🎮 Site: daedalus.ethz.ch Blog post: daedalus.ethz.ch/overview.ht… Demos: daedalus.ethz.ch/#demos Games: daedalus.ethz.ch/#arcade
1
1
2
132
😳
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
1
1
80
For the first time I’m scared of AI, no joking
1
43
caching has always been the key
56
GPUs finally used in the interest of the gaming industry again 🤝
This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
54
Honestly if they wake me up now and tell me the whole Claude thing was just a dream I won’t be surprised. I’ll be like oh right I felt something was not right all the time.
47
Suddenly it sounds like a cool idea to stay in the academia
40
What’s really great about this democratization of knowledge through LLMs is that scientists will again be able to have a good overview of multiple disciplines like 200 years ago.
1
2
49
Thinking about it this way, I expect to see a major shift in disciplines and the working system of academia and a lot of mind-blowing advancements in all areas of research in the next 10 years; if we can survive the transition.
34
sadly not my N
Memory I/O are slow, float ops are fast, so matrix product is awesome: O(N^2) floats to transfer and O(N^3) operations to do, there will be a N where both memory I/O and the floating point units will be perfectly busy.
1
2
1,037
Sepehr Mousavi retweeted
by contrast, this is what packing circles into a circle looks like
i'm consistently surprised this square packing stuff elicits such a strong reaction among non-mathematicians. to me this falls under "ask ugly questions, get ugly answers." there's an art to asking good questions and there's nothing particularly natural about square packing
90
210
38
7,162
658,666