https://t.co/YsdnkNc8xL - A hackable @PyTorch script to visualize emergent structures in n-interacting-particle systems.

New 𝗩𝗮𝗹𝘂𝗮𝗯𝗹𝗲 𝗙𝗿𝗲𝗲𝗕𝗦𝗗 𝟭𝟱.𝟬-𝗥𝗘𝗟𝗘𝗔𝗦𝗘 𝗨𝗽𝗱𝗮𝘁𝗲𝘀 [Valuable FreeBSD 15.0-RELEASE Updates] article. https://t.co/qraS629fmU #verblog #freebsd #improvements #release #updates

@d0rc The egg is hatching so fast 🥲

what should you do if the transformer generates the first readable word and it is "mother"...?

I took some time to flesh out the documentation of the Fil-C compiler, with special attention paid to the clang driver changes! https://t.co/vPSHXqe5GT

Memory safe Linux now has pretty fonts because I ported freetype, fontconfig, harfbuzz, and graphite to Fil-C. This is the Deja Vu monospace font. This entire desktop is memory safe because everything is compiled with Fil-C. It's snappy - you could use it as a daily driver

We just made an app that walks you through designing a novel protein with AI from scratch. Takes about 5 minutes, requires zero biology knowledge. ➡️ https://t.co/JEJD5sd7A7 The best part: we will actually synthesize 1000 of those protein designs in the lab and test their real

*C* is on the rise:)

GigaAI just dropped a major release for the Embodied AI community: The GigaBrain-0 VLA Model is officially open-source, with performance that reportedly matches the leading π0.5 model! This model is a huge win for researchers, featuring: ➤ Data Power: Trained on 1000+ hours

You want baselines for the int8 EGG model? Here you go! Standard transformer (fp32 + backprop + SGD) vs EGG in the data-efficient regime. EGG Outperforms at large data batch sizes?!

Good news, everyone The European Parliament has decided that the government, and large tech companies, are responsible for baby sitting YOUR children. They have voted to work on legislation which makes social media platforms such as TikTok, X, and Meta (Facebook, Instagram,

EGGROLL 8xB200 distributed training with P2P weight averaging! https://t.co/WimvgcOLj4 Thank you @NorthflankWill and @northflank for the B200s.

I am increasingly glad to have left this continent behind. God bless the United States, cherish the republic you have kept.

@ReginaDo @EPPGroup You're an imbecile. This won't stop anything. The kids on social media don't have parents who actually "parent". All they're going to do is find a bypass or convince their parent (or some other adult) to allow them access to the social media platform. You've just given large

GRPO at hyper-scale at home is unlocked:)

Soo, I cannot register for NeurIPS - it is sold out 😅 I really want to go because: 1. BoltzGen meetings 2. I am applying for faculty positions and have some meetings surrounding that set up. Does someone have a registration but can no longer come? :) ❤️

Replacing rigid, old-school models with neural nets paired with "split conformal prediction", this tech predicts potential errors and wraps the robot’s motion in Lyapunov safety certificates. Flexible movement with a mathematical promise that the robot won't crash or drift.

@j_foerst I'm still trying to generate readable text with it, but the implementation is "almost certainly" correct It reaches a loss of ~3.3, runs at 4.5M tps on a 5090, and can be optimized further it has 64-bit accumulators to allow for larger populations https://t.co/7BVAbNqfjo

In-N-On: a new recipe for scaling robot manipulation using human egocentric videos and it’s a game-changer.

I genuinely believe that EGGROLL can lead to the "Cambrian Explosion" of AI systems

@Leik0w0 @d0rc @j_foerst I just published my attempt at this code, I have an int8 TC hotpath using cublasLT, although I still need to change it to use int8 activations, because there is quant/dequant overhead. Something I'll try later. https://t.co/4fIBv7OrVp