Christian S. Perone

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/ 🧵👇

Thanks for sharing, very interesting, will read it !

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.

Not to mention this passage below from his "Du mode d’existence des objets techniques", models today are proving the viability of a structure similar to a natural structure, and we are now submitting models more and more to inductive study because they bear scientific value. 2/2

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 😅

YouTube video by Christian S. Perone TorchStation - Development proto v1

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...

Very expensive 4 points though 😅

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 🤟

GitHub - perone/vectorvfs: Your filesystem is a vector database Your filesystem is a vector database. Contribute to perone/vectorvfs development by creating an account on GitHub.

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)

Now, independently of regulation, states and legislators need to make very clear that: make no mistake, since these systems are derived and relational, *humans remain ultimately responsible* for its actions and oversight. 4/4.