I'm late to the game but... making crazy combos in @playbalatro.com is awesome!

Very impressive!

LLMs are largely memory bound! Tokenizers can eat >30% of on‑device memory for small (<1B) models, because we reuse the huge vocabularies from big models. Our method lets you distill large models into small ones with much smaller vocabulary, reducing overall memory footprint.

We can't fall into right-wing populism’s lie that the most vulnerable in society are to blame for wealth inequality in our countries. We need to build movements that tell the truth: the story of wealth inequality is not a cultural one, but a class one.

Emails Show Epstein Scheming That Environmental Destruction Could Solve “Overpopulation” | Article by Joe Wilkins (@joeonhere.bsky.social) for @futurism.com: futurism.com/science-ener...

This is Edward Niedermeyer explaining, in plain terms, why Tesla stock stays float but above the $400 mark, despite the car sales going down. I encourage anyone to watch it from the marker onwards I've set as Edward explains that #$TSLA is built on momentum and what we can do to bring him down

Small Language Models (SLMs) don’t have the capacity to remember everything in their training data. Which tokens should they learn to predict, and when should they ask for help? We tackle this question in our new preprint. You can check it out on arxiv: arxiv.org/abs/2602.12005 🧵1/7

Absolutely breathtaking!

Billionaires who lobbied Trump to halt Canadian bridge praise EZ-Bribe App

OneVision Encoder Currently, multimodal models sample frames at a very low FPS, reducing continuous motion to a handful of static snapshots. OneVision Encoder uses video codecs - native motion vectors and residual signals to select patches,

The new ByteDance SeeDance 2.0 video model is VERY good. Each video is the very first output of the prompt. There are four, worth seeing them all to get a sense of the range (and potential issues) "A nature documentary about an otter flying an airplane"

xAI founding members are jumping ship.

You can now train MoE models 12× faster with 35% less VRAM via our new Triton kernels (no accuracy loss). Train gpt-oss locally on 12.8GB VRAM. In collab with @hf.co, Unsloth trains DeepSeek, Qwen3, GLM faster. Repo: github.com/unslothai/un... Blog: unsloth.ai/docs/new/fas...

✨ New 3D pose estimation method from my lab! #FMPose3D allows for monocular (i.e. single camera) 2D➡️3D 🔥 Led by Ti Wang & w/ Xiaohang Yu #FMPose3D is SOTA on human & animal 3D benchmarks, & will be integrated into @deeplabcut.bsky.social ⬇️ 📝 arxiv.org/abs/2602.05755 ➡️ xiu-cs.github.io/FMPose3D/

Google DeepMind's Mining Generalizable Activation Functions They argue that evolutionary search is a powerful framework for discovering new activation functions, showing that LLM-driven pipelines like AlphaEvolve replace manually designed search spaces with flexible,

the fact that early-stage automated vehicles that still occasionally require remote operation assistance are like several orders of magnitude safer than the average American driver should be the real wakeup call here, not the idea that these vehicles still sometimes need remote operation assistance

Pydantic's Monty A minimal, secure Python interpreter written in Rust for use by AI. github.com/pydantic/monty

RayRoPE: Projective Ray Positional Encoding for Multi-view Attention Yu Wu, Minsik Jeon, Jen-Hao Rick Chang, @onceltuzel.bsky.social Shubham Tulsiani tl;dr: even more projective geometry based PE, if you know your camera pose. arxiv.org/abs/2601.15275

They propose a new paradigm called Drifting Models, which evolve the pushforward distribution during training and naturally admit one-step inference. We introduce a drifting field that governs the sample movement and achieves equilibrium when the distributions match.

In case it wasn’t ALREADY blatantly clear!

They develop TinyLoRA, a new ft method. with TinyLoRA + RL, models learn well with dozens or hundreds of params. For example, they use only 13 parameters to train 7B Qwen model from 76 to 91% on GSM8K. "Learning to Reason in 13 Parameters" Paper: arxiv.org/abs/2602.04118

It is possible to get Genie 3 to make worlds that are visually interesting I have been using images I generated in Midjourney of vast megastructures & odd cities in various styles. After 20 seconds I can freely wander around them for a minute or so. (Yes, I controlled the cat in the first scene)

Two new speech-to-text models (similar to Whisper) from Mistral today - one of them is API-only, the other is a 8.9GB Apache-2.0 licensed open weights model for "realtime" transcription. They're both very good! simonwillison.net/2026/Feb/4/v...

Nvidia’s Nemotron is the closest thing the U.S. has to a Qwen approach to open models, but most people don’t know it yet. I’m very bullish on Nvidia’s open model efforts in 2026. Interconnects interview #17 on the past, present, and future of the Nemotron project. www.youtube.com/watch?v=Y3Vb...

Rectified LpJEPA A JEPA architecture that learns sparse, non-negative, informative representations through principled distributional regularization.

The dreamcoder boys are back https://arxiv.org/abs/2602.00929

What if position encodings were designed for vision from scratch? We introduce PaPE—Parabolic Position Encoding. Outperforms RoPE on 7/8 datasets and extrapolates to higher resolutions without fine-tuning or position interpolation. Paper, code, and website in thread 🧵

Today we’re releasing the International AI Safety Report 2026: the most comprehensive evidence-based assessment of AI capabilities, emerging risks, and safety measures to date. 🧵 (1/19)

I was in a zoom call with Brock Pierce a couple years ago, along with the startup CEO I worked for. Brock was introduced to us as potential investor. Brock was a weird creep that oozed dishonesty. I told the CEO we didn’t want that guy in our cap table or anywhere near. We dodged a bullet.

I'm finally writing up how Nanite Tessellation works. The first few blogs posts are up. More will be coming. graphicrants.blogspot.com/2026/02/nani...

Next time someone asks you about energy and land use, maybe remind them that the ~35 million acres that currently grow corn for ethanol in the US could produce ~15 PWh per year of electricity from solar photovoltaics. That's ~3.5 times more than *total annual generation from all US power plants*.

I love this… I worked on the realtime renderer that shipped with Maya 4! Early programmable shading, even getting a good per-pixel specular was challenging!

Donald Trump speed-running USA towards dictatorship and civil war

NVIDIA announced "Cosmos Cookoff", a virtual, 4-week physical AI challenge for robotics, AV, and vision AI builders. 🏆 Prizes include $5,000, an NVIDIA DGX Spark, and more! nvda.ws/3Z9lrCO

📉25% LOWER ALL-CAUSE MORTALITY! Wowzers—one of the largest long-term safety studies ever undertaken: “Among 28 mil adults, those who received an mRNA-based COVID-19 vaccine were less likely to die in subsequent 4 years than unvaccinated, corresponding to a 25% lower risk of death from all causes.”

A data collection farm from Fourier robotics. The operators appear to be collecting data for brain-computer interfaces while also controlling the robots to do a wide variety of tasks. Video from RoboHub on X (originally from Fourier)

This project looks awesome. Are you using any plugin for this, e.g. dungeon architect?

This is great news! Too few ML practitioners know what a gamechanger nbdev can be. Years ago at AMD I used nbdev (version 1?) to implement an early prototype of FSR4 realtime super resolution. It was self-explanatory (e.g. with diagrams showing examples of jittering and temporal super resolution)…

NVIDIA launched an ultra-efficient NVFP4 precision version of Nemotron 3 Nano that delivers up to 4x higher throughput on Blackwell B200. "Using our new Quantization Aware Distillation method, the NVFP4 version achieves up to 99.4% accuracy of BF16."

100%. I honestly believe political leaders ask too little of people. The thinking is you need an extraordinarily low bar - so ask people for money or to sign something. People want to be part of something and they're eager to do real meaningful work in defense of their community and their rights.

Stoked to share my latest podcast as Arcee launches their 400B MoE model today — Trinity Large — open weights, good license, strong performance. A small startup decided to take on the many established AI labs by pretraining their own model to release openly and monetize later, and they’re succeeding

Man who wouldn’t wear mask to protect people fine wearing mask to shoot them

NVIDIA released FastGen, a flexible, unified interface to perform model distillation, e.g. to speed up image and video models. developer.nvidia.com/blog/acceler... github.com/NVlabs/FastGen

Thank you WEF for the golf cart ride through Davos, but also (more importantly!) for the excellent questions about the future of AI. 🛺 www.weforum.org/videos/davos...

Board of Peace - Season 1

Can we compress human speech for AI without losing its nuances? FocalCodec proves it’s possible. By simplifying how tokens are learned, this model paves the way for more efficient multimodal LLMs. Research explained by Luca Della Libera, Francesco Paissan, Cem Subakan, and Mirco Ravanelli.

That’s the timeline we should all live in - one in which NASA’s patents lead to nuclear fusion and electric vehicles in the 1990s. I will re-subscribe to Apple TV just for that show.

For the LeJEPA/SIGReg/SSL demo, I added a summary of the algorithm (scroll down) and links to more resources: www.scotthawley.com/ssltoy

Wow, this is an impressive milestone in LLM progress