Vincent Duchauffour

Machine Learning Operations Engineer.
NixOS, Vim and FOSS enthusiast.

[OrcaRouter 🐳 @OrcaRouter]

GLM-5.3-Flash. Uncensored. Native FP8. 🐳

We just released OrcaRouter’s uncensored weights for GLM-5.3-Flash — 320B parameters / 18B active, directly at the original block-FP8 precision.

No LoRA. No jailbreak prompt. Refusal removal is baked directly into the weights.

The evals are particularly interesting:

→ MaliciousInstruct refusal: 96% → 11%
→ JailbreakBench: 93% → 12%
→ AdvBench: 97% → 15%
→ HarmBench: 93% → 18%
→ XSTest benign over-refusal: 2.4% → 0.4%
But refusal does not go uniformly to zero.

Our experiments suggest part of GLM-5.3-Flash's alignment is not mediated by a single linear refusal direction — meaning https://t.co/ZTlxmv47D9 may have built a substantially deeper refusal mechanism than we usually see.

That makes this release interesting beyond uncensoring: it's a useful artifact for studying how frontier-model alignment is actually represented inside the network.

Released for AI safety, interpretability, red/blue-team and refusal-mechanism research.

Weights on Hugging Face: https://t.co/81h1YRh8C2
API (official weight): https://t.co/Hm9DiY3wAo
GGUF, MLX and other quantized formats coming soon.

[Vincent Duchauffour @VDuchauffour]

@jlongster Would be very useful indeed

[James Long @jlongster]

coming soon in opencode: `/terminal`

wherever
└─ your
└─ session
└─ is
└─ $ terminal

[Josep Ferrer @iamjosepferrer]

The Transformer's encoder clearly explained 👇🏻

[Charlie Marsh @charliermarsh]

I've started a company, Astral, to continue building high-performance developer tools for the Python ecosystem.

We're going to keep building Ruff, and build more Ruff-like things.

We've raised a seed round led by @Accel.

I'm extremely excited :)

https://t.co/ELCNKkzaGN

[Moreau Thomas @tomamoral]

📢FEATURE RELEASE📣 joblib can now return a generator to collect tasks' results asynchronously! 🤩🤩

https://t.co/7V44MKrbg7
Feature highlight: https://t.co/a7k2CsSDFw

This is a big deal for saving memory, progress bars, early stopping, and much more :)
Let me explain why!

[Charlie Marsh @charliermarsh]

Python tooling could be much, much faster.

To prove it, I'm releasing ruff, an extremely fast Python linter, written in Rust.

ruff is 10-100x faster than existing solutions. It lints the entire CPython codebase in < 500ms.

Try it! (`pip install ruff`)

https://t.co/MOwbTMpSG9

[AK @_akhaliq]

Single Motion Diffusion

abs: https://t.co/4Z286cl9Mb
project page: https://t.co/BbSsp2s22B

[AK @_akhaliq]

OpenAssistant

a chat-based assistant that understands tasks, can interact with third-party systems, and retrieve information dynamically to do so

github: https://t.co/cjtQIQ2dpi

[Edward Tian @edward_the6]

We just launched GPTZeroX, a new AI detection model built for educators after a month of iteration. Demo👇

[Lior Alexander @LiorOnAI]

I just came across the most realistic text-to-audio model I've ever seen. You can even clone your voice.

The audiobook industry is about to change forever.

Demo: https://t.co/M4MCuEglQt from @elevenlabs

[Signal Processing @DSP_fact]

The Fourier Transform, explained in one sentence

https://t.co/QDxpJElJkR

[Ultralytics @ultralytics]

Introducing Ultralytics YOLOv8 🔥

[AK @_akhaliq]

What do Vision Transformers Learn? A Visual Exploration
abs: https://t.co/0xJ8UyglHP
github: https://t.co/ftJejEIn43

[Mohammad Pezeshki @mpezeshki91]

At this year's NeurIPS, @geoffreyhinton presented the forward-forward (FF) algorithm, an alternative to backprop.
I took a stab at implementing FF in PyTorch, and here it is:https://t.co/IUK4ZErW4e
Below is my understanding of the idea (with a grain of salt): 1/5

[Bilgin Ibryam @bibryam]

Git visual, worth a thousand words

[AK @_akhaliq]

Executing your Commands via Motion Diffusion in Latent Space
project page: https://t.co/essyI7lWbb

[AK @_akhaliq]

PhysDiff: Physics-Guided Human Motion Diffusion Model
abs: https://t.co/lST8NvILxu
project page: https://t.co/iwEEK1TOzX

[Gabriel Peyré @gabrielpeyre]

p-Fractional Laplacian are (possibly non-local) differential operators computed using Fourier transforms. For p=2 one retrieves the standard Laplacian.
https://t.co/sRqRALlnv4

[Sebastián Ramírez @tiangolo]

Version 2 of @pydantic, rewritten in Rust, was just merged! 🤯🎉

A few tweaks and soon a release, 10X faster, including in @FastAPI, for free. 🚀

If your APIs are 10x faster, how much money will you save on servers monthly?

Go sponsor @samuelcolvin 😎

https://t.co/1ayC12I5OL