šŖ - working at https://xitter.cf/t.co/s7UZAGGVxr - come join our fight against skin cancer! (previously at Hugging Face).
Paris area, France
Joined April 2008
- Tweets19.8K
- Following5.5K
- Followers8.3K
- Likes30.2K
Julian Bilcke retweeted
This is fascinating. Steering just two neurons in Qwen3.5-4B can make it talk about harming, dominating, and destroying you.
I built an interactive playground so you can try this (and many other combinations) yourself, with the model running entirely in your browser on WebGPU š
Julian Bilcke retweeted
Run Hy-Nomas World Model with 8GB VRAM
WanGP has added Hy-Nomas, a world model you can control with WASD and text prompt.
Should run realtime on 24GB+ VRAM machines, but probably not realtime on 8GB VRAM.
1-click launch on pinokio (thread)
š WanGP v18.81 ā The One World Generator
Build a world from text with Hy-Nomas or from an image with EVOKE, then explore it with your keyboard! Save your progress, record your journey and revisit familiar placesāstarting at 8 GB VRAM thanks to MMGP v4.
š¤ Deepy Prime your local Agent gets roughly 2Ć the context at the same VRAM, plus new abilities:
š Search the web and find compatible LoRAs
š Create graphics, animations and batch workflows with Python
š§ Generate 3D assets and explorable worlds
š¬ Restore, upscale and outpaint videos through Media Flow
Free to run locally. Your GPU has more places to take you.
github.com/deepbeepmeep/WanGā¦
Julian Bilcke retweeted
We finally cracked pixel VFX.
And we just open-sourced the Claude skill behind it.
Download here: scenario.com/explorations/piā¦
> describe a spell and Claude stacks the layers: fireball, shockwave, sparks
> the engine renders every frame locally, for free
> export a sprite sheet, GIF or Godot 4 pack
Julian Bilcke retweeted
The largest math proof in human history is now computer-verified.
Today, Axiom announces a complete Lean auto-formalization of the Classification of Finite Simple Groups, the "periodic table of symmetry."
About 15,000 pages, built by over 100 mathematicians across generations.
Julian Bilcke retweeted
Here's a direct link to OpenAI's PDF claiming a refutation of ETH. In particular, it's claimed that 3SAT on n variables is in about exp(n/(log n)^{1/2+o(1)}) time
github.com/openai/math/blob/ā¦
Julian Bilcke retweeted
We are announcing three new discoveries in mathematics, computer science, and physics about harmonic analysis, complexity theory, and quantum field theory.
openai.com/index/building-scā¦
Julian Bilcke retweeted
GPT-6.1 Terra continues to advance the cost-intelligence frontier across our model families.
Itās a significant step up over GPTā5.6 Terra, with 50% lower standard API prices.
In ChatGPT Work and Codex, your usage allowance goes further with GPTā6.1 Terra than with GPTā5.6 Terraāand far more than with GPTā6.1 Sol or GPTā6 Astra.
For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.
Julian Bilcke retweeted
š NVIDIA has released the MotionBricks weights on the Hugging Face Hub!
MotionBricks generates real-time motion for game characters and humanoid robots. Give it a character's recent poses and the poses it should reach next, and it generates the movement in between, at 15,000 FPS with 2 ms latency.
- One model trained on 350K+ motion clips
- Smart Primitives for navigation and object interaction, no retraining
- SOMA weights for human characters
- Unitree G1 weights for the humanoid robot
- Code under Apache 2.0, weights under the NVIDIA Open Model Agreement
Animation and robotics motion generation, now open for anyone to build on!
Julian Bilcke retweeted
Beavers stop chewing every few seconds to listen for cracking sounds in the tree so they don't get crushed
Julian Bilcke retweeted
This is a virtual human body with 700 muscle actuators, 206 joints, and 90 body segments.
It's fully open source, and researchers can use it to simulate human movement through muscle-level control.
Julian Bilcke retweeted
Foldable iPhone Duo devices are reportedly breaking in large numbers, with displays failing along the hinge crease.
Buyers in Japan, China, Australia, and other countries have reported the same issue. In some cases, the display reportedly fails while the phone is still on display in stores.
The symptoms are identical: after just a few days of use, a black line appears across the screen.
Julian Bilcke retweeted
MEKsans (top) is a better font in kerning, being a non-mono font as opposed to MEKmode (bottom). More efficient spacing, better legibility, something I want to apply on the next build (v0.13).
The glyphs that you've seen in-game came from MEKmode Dings though, all of it!
Julian Bilcke retweeted
Drex 2.0 now runs completely local on Apple Silicon.
Open weights. Sub-200ms latency. Zero cloud dependencies.
ā¢ Footprint: 21 GB (4-bit Q4_K_M)
ā¢ Hardware: 36 GB+ Unified Memory (Apple Silicon)
Median decision latency (M5 Pro):
ā¢ Kapitsa: 113 ms ā¢ llama.cpp: 248 ms ā¢ Ollama: 282 ms
Hugging Face: huggingface.co/nace-ai/drex-ā¦
GitHub: github.com/nace-ai/drex-deciā¦
Julian Bilcke retweeted
We are open sourcing Phonon-2.5, setting a new accuracy-to-size Pareto frontier in multilingual speech recognition.
We brought a 2.5 GB model down to just 6.2% (!!) of its original download size, and simutaneously improved its accuracy.
Phonon-2.5 achieves 101.8% of its full-precision baseās average word accuracy across 25 languages.
Weāve begun rolling out Phonon-2.5 in Detta, bringing dictation in 25 languages to your Mac, completely locally!
Introducing Phonon-2.5: A new accuracy-to-size Pareto frontier in multilingual speech recognition.
It is a miniscule 156MB download, which is more accurate on average, than OpenAIās Whisper large-v3 ( 20x its size!)
Transcribe an hour of multilingual audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Julian Bilcke retweeted
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a personā¦full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
Julian Bilcke retweeted
Introducing Phonon-2.5: A new accuracy-to-size Pareto frontier in multilingual speech recognition.
It is a miniscule 156MB download, which is more accurate on average, than OpenAIās Whisper large-v3 ( 20x its size!)
Transcribe an hour of multilingual audio in just 20 seconds on a MacBook Air!
Open weights, CC-BY-4.0, today.
@fermion_ai
Julian Bilcke retweeted
Just Eat apologises for AI food mistakes after social media ridicule bbc.in/4jAIj9J
