Recommended !

Gemma3n was released a few months ago, I wasn't able to find more info and I found it a *very interesting* architecture with a lot of innovations (Matryoshka Transformer, MobileNetV5, etc), so I decided to dig further, here you are the slides of this talk: drive.google.com/file/d/15hbh...

Announcing the @iccv.bsky.social NAVSIM Challenge! What's new? We're testing not only on real recordings, but also perturbed futures generated from the real ones via pseudo-simulation! $8K in prizes + several $1.5k travel grants. Submit by September 20! opendrivelab.com/challenge2025/ 🧵👇

I feel I can build an entire benchmark dataset with ONNX errors that would be harder than the humanity's last exam dataset for us to evaluate AGI

Guillermo del Toro - Studio Ghibli Masterclass (2013, TIFF festival) ❤️ Now online (80 minutes) >> www.youtube.com/watch?v=q8Uo...

"Optimizers Qualitatively Alter Solutions And We Should Leverage This" (arxiv.org/abs/2507.12224), very nice to see this direction of understanding what different optimizers bring in terms of solution properties.

I find incredible how much we can relate about the evolution of Machine Learning in the past decade to what Simondon described in 1958. The shift towards more generalist systems is exactly what Simondon's concept of "concretization" is about. 1/2

I made a diagram on how you can use a World Model with Diffusion Elites:

This is absolutely *charming* in its straightforwardness. All kinds of bells and whistles suggest themselves at once, but this gives "really strong baseline for the general case" vibes.

I'll be at ICML this week, presenting our paper on Wasserstein Policy Optimization on Tuesday! If you're in Vancouver, come say hi!

New blog post: "Diffusion Elites: surprisingly good, simple and embarrassingly parallel", blog.christianperone.com/2025/07/diff...

"Chip Placement with Diffusion Models" (openreview.net/pdf?id=crCPL...) very cool paper.

I'm getting addicted to animations of Langevin sampling with fixed rng and varying params.

Given the amount of different definitions of world models, at this point, I think I can call any model a world model.

If you change the tensor in PyTorch, it will change the tensor in Jax, Numpy, PyTorch and Tensorflow 😅

After a lot of issues with power distribution, the first panel of TorchStation proto is finally here 😉 node sel. for distributed training is coming. You will soon have an open-source and open-hardware @pytorch.org distributed training monitor on your desk: www.youtube.com/watch?v=D7po...

We are hiring 2 Machine Learning Engineers in London/UK 🇬🇧 to work with end-to-end automated driving. ➡️ Senior Machine Learning Engineer woven.toyota/en/careers/d... ➡️ Machine Learning Engineer woven.toyota/en/careers/d... We sponsor visas as well !

VectorVFS is on Hacker News front page 🤟

Introducing VectorVFS, your filesystem as a vector database: github.com/perone/vecto.... VectorVFS stores embeddings directly into filesystem inodes. No external index, daemon, database or metadata files. The first model supported is the SOTA Perception Encoder from Meta.

New open-source project coming out soon 🤞

I think the Google Search Appliance (GSA) was a nice concept that suffered a unfortunate timing. Imagine it today with on-premise LLMs, multi-modal document indexing and modern retrieval. All local wo/ any data sent to cloud. I really want to develop a prototype w/ a Jetson.

IRoPE on LLama 4 seems very interesting, some clever tricks there.

New Forest and its magical beings.

What a crazy evolution. Slide from NVIDIA Blackwell Numerics for AI presentation.

Machine learning engineers can now feel the pain epidemiologists endured during covid: everyone's suddenly an expert.

That can partially explain the similarities of DeepSeek-R1 CoTs and o1-preview CoTs, even though the CoTs from o1 are hidden. There are now probably 3 hypothesis: 1) convergent behavior 2) distillation (but not of CoTs) 3) sinister data leak (directly or through a 3rd party)

Discussions around AI agents would be much more productive if we just make clear the difference between functional autonomy and vs ontological agency. I can see many folks talking past each other because the lack of understanding of these two different types of agency. 1/4

🤭

Finally finished an article I was working on since past year on Gilbert Simondon's philosophy of technology and how it relates to AI/ML. Simondon's philosophy radically changed my views on technology, anyway, here you are the article: blog.christianperone.com/2025/01/note...

If one watches AI competition carefully, it can be weird to see one company saying they need to take *drastic action*, while another company has a *drastic research* team. IMHO, when everything is portrayed as drastic, then nothing really is, as this often leads to nihilism.

I think that Ilya's talk in NeurIPS is just another example of how ML researchers/engineers cannot protect themselves enough against misrepresentation when they speaking in public. I wish we had the same level of scrutiny for futurists spreading historicism as well.

Optimizers in deep learning are a concrete example of how we can move forward and develop very useful algorithms without knowing what is actually going on 😅

I wrote a summary of the main ingredients of the neat proof by Hugo Lavenant that diffusion models do not generally define optimal transport. github.com/mathematical...

I think I will be laughing for the eternity specificsuggestions.com

The sbi package is growing into a community project 🌍 To reflect this and the many algorithms, neural nets, and diagnostics that have been added since its initial release, we have written a new software paper 📝 Check it out, and reach out if you want to get involved: arxiv.org/abs/2411.17337

Amazing blog post on flow matching, stunning visuals! It also makes the connection with normalising flows crystal clear. Incredible effort!

Anne Gagneux, Ségolène Martin, @quentinbertrand.bsky.social Remi Emonet and I wrote a tutorial blog post on flow matching: dl.heeere.com/conditional-... with lots of illustrations and intuition! We got this idea after their cool work on improving Plug and Play with FM: arxiv.org/abs/2410.02423

Finally got a RP2350 😍

For the theoretical side, there’s no better resource than the 2024 book by Blondel and Roulet: arxiv.org/abs/2403.14606 For the practical side, my friend @willtebbutt.bsky.social wrote stellar documentation explaining autodiff with mutation: compintell.github.io/Mooncake.jl/...

Finally an interesting standard for LLM integration, this Model Context Protocol (MCP) from Anthropic looks quite promising (modelcontextprotocol.io/introduction).

I really don't understand why Google keep using these Abseil bindings in Python. Even with plenty of arg parsing options and CLI frameworks, Google prefers this interface: https://buff.ly/4g4ruip. It's just nonsense.

Normal modes of Chladni plates. Made with #python, #numpy and #matplotlib.

You are not prepared for what is coming in 2025 🤭

The LLM parrot analogy is dead. Fantastic work by UCL DARK's @lauraruis.bsky.social on rigorously investigating whether LLMs learn reasoning from procedural knowledge during pretraining.

I think a lot of people would be surprised if Arena Chatbot disclosed the IP addresses and their respective ISPs for the votes given. Some folks would be shaking their legs just to the idea of sharing this 😅

Hi, just sharing an updated version of the PyTorch 2 Internals slides: drive.google.com/file/d/18YZV.... Content: basics, jit, dynamo, Inductor, export path and executorch. This is focused on internals so you will need a bit of C/C++. I show how you can export and run a model on a Pixel Watch too.

In a gratuitous attempt to acquire more followers myself 😁, I've made a start on a "starter pack". Hopefully as more people from 🐦 make it over to 🦋, we can extend this a bit. Suggestions welcome! I've noticed not all accounts seem to be eligible to be added, anyone know what's up with that? 🤔

*The geometry of data* Fascinating blog post by @cperone.bsky.social on how we can use the Stein score (of diffusion renown) to build a metric tensor describing the geometry of our data (e.g., computing geodesics between points). blog.christianperone.com/2024/11/the-...

New article: "The geometry of data: the missing metric tensor and the Stein score" (blog.christianperone.com/2024/11/the-...). I show how you can derive a (efficient to compute) data manifold metric tensor with the Stein score alone ! Deep connections to diffusion, score-based models and physics.

Very eager to see the excavations of The Curtain Theatre in London Shoreditch, few people know but a new museum (The Shakespeare Museum) will soon open there close to The Stage building. 400 years just waiting there in the mud, just unbelievable.