beat · 154 stories
Aleph Alpha's new open-weight language model Kolibri — 78B total / 3B active per inference pass via a Mixture-of-Experts design — is what decides whether a regulated European buyer can self-host a frontier-class model on its own hardware.
Musk says he is in talks with TSMC about a dedicated chip fab, his 'Terafab' initiative, to supply Tesla, SpaceX, and xAI. Intel, the only other named partner contributing its most advanced announced 14A chipmaking process, may not be in the room.
The 24GB M4 Pro MacBook Pro runs the lower 40 layers (stacked model stages) of the 64-layer Qwen3.8-27B; the iPhone 17 Pro Max's A19 Pro GPU runs the upper 24.
A new measurement study shows why one Adam run (the default neural-network training optimizer), in a neural network under ill-conditioning (a sharply uneven loss surface), can drift a thousandfold from the theoretical ideal and still reach the
An arXiv preprint proposes sampling tailored modifications to a model's own settings, rather than just longer streams of answers, as a way to buy more capability per fixed compute budget.
The 85% cache share — re-reads of text the model has already processed in the same session to keep its working memory coherent — points to high-bandwidth memory, not GPUs, as the next build constraint.
University of Science and Technology of China (USTC) and MIT researchers used quantum vacuum fluctuations to lift an ultrathin superconductor's transition temperature — the temperature at which the film loses all electrical resistance — by up to 5.
Muse runs an hourly job to build a page on every person in your life, then offers to show you the file. The user did not ask for it to be built, and the cadence is something the user did not set.
Fatigue, not raw strength, is the constraint five 2026 deployments are now buying around, from mountain rescue to artillery resupply.
Walking, general-purpose robots built to work alongside people are now shipping in roughly one-meter form factors, expanding who can actually use them.