Dario Amodei says we're close to human-level AI, yet society doesn't see the tsunami coming Instead of preparing, people dismiss it as a trick of the light, while some push to accelerate blindly without realizing the risks "technical control has improved more than expected, but societal awareness has gotten worse"

In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away. vincentsitzmann.com/blog/bit…

my favourite citation yet

Glad to have our #ICCV2025 work out in open. It was a very rewarding experience learning and setting up the lightstage, capturing data and then creating Gaussian Avatars. First of its kind for developing relighting human avatars

We present UniGAHA: Audio-Driven Universal Gaussian Head Avatars at #SIGGRAPHAsia2025 From speech, UniGAHA generates photoreal 3D head avatars that lip-sync, shift gaze, and capture subtle expressions. Works across identities + adapts from a short video. bit.ly/46z7p0G

🧵 PractiLight: Practical Light Control Using Foundational Diffusion Models 🔗 yoterel.github.io/PractiLigh… Diffusion models already “know” a lot about light transport. No need for massive finetuning that hurts generalization. {1/6} 👇

Breaking news: A large-scale, publicly accessible dataset of multi-view OLAT full-body human captures now exists. Accepted at #ICCV2025! A collaboration between MPI-INF and NVIDIA led by Timo Teufel. For @jankautz and me, this is our second joint work. vcai.mpi-inf.mpg.de/projects…

Google just released a brilliant research, a new active learning method for curating high-quality data that reduces training data requirements for fine-tuning LLMs by orders of magnitude. You get the same or better model quality with 250 to 450 expert labels instead of 100K, thanks to focusing expert effort only on confusing boundary cases. That means dramatically lower labeling cost, much faster iteration, up to 10,000x less data to collect, and easier updates when policies or abuse patterns change. Targeted expert labeling around the decision boundary lets a small fine‑tuning set of 250 to 450 examples match or beat models trained on 100K crowdsourced labels, reaching up to 10,000x training data reduction without hurting quality. The usual practice, throwing massive noisy datasets at the model, is slow, expensive, and brittle when ad safety rules or abuse patterns shift. 🧠 The Core idea The team uses an LLM as a scout to sweep a huge pool of ads, then asks experts to judge only the handful of examples that truly confuse the model. Those expert decisions train the next model iteration, and the loop repeats until model–expert agreement stops improving. It is classic active learning, but adapted to large, noisy, imbalanced traffic like ads, where only about 1% of items are actually problematic. @GoogleResearch

🎉 LeRobot Hackathon Wrap-Up + Winners Announcement! What a weekend! 🌍✨Over 3,000 participants from 44 countries, 250+ submissions, and 55+ hours of non-stop robotics hacking ,we’re still buzzing from your incredible energy! A massive THANK YOU to everyone who joined, shared ideas, and made this community so special. 👉 Meet the 30 winning teams! 🏆 huggingface.co/spaces/LeRobo… Keep tinkering, keep collaborating, and stay tuned for what’s next. Until then, keep those robots dancing! 🤖💃