🚨President Putin sends message to the US about India and PM Modi:
“The US is trying to pressure India, for example, when it comes to cooperation with Russia. But putting pressure on Narendra Modi's democracy is detrimental for international relations. It doesn't matter where this pressure comes from.”
China silver imports hit record high – nearly 800 tonnes – in January and February
China's silver imports hit an eight-year high in the first two months of 2026 as burgeoning demand from the industrial and investment sectors boosted local price premiums, drained domestic silver stockpiles, and drove unprecedented foreign purchases, according to the latest customs data released Friday.
In January and February, #China imported over 790 tonnes of #silver, the data showed, with February alone accounting for a year-over-year record of 470 tonnes in imports...
Full story at Kitco:
https://t.co/oVRnmgN3S6
🚨 BREAKING: ByteDance just dropped a paper that should make every GPU infra team and NVIDIA investor do a double-take.
An AI agent that now writes faster CUDA kernels than torch.compile the gold-standard compiler every ML engineer relies on.
Not close. 2.11× faster overall geometric mean.
Claude Opus 4.5 hits 1.46×. Gemini 3 Pro hits 1.42×.
ByteDance’s CUDA Agent beats both by ~40% on the hardest benchmarks.
Here’s the part that’s actually wild:
The same base model started at 0.69× (slower than the compiler).
Then they ran pure reinforcement learning on real GPU profiling data— no human-labeled examples, just the model writing kernels, running them, measuring wall-clock time, and learning. Weeks later it had done what human compiler teams took years to perfect.
The implication nobody wants to say out loud yet:
The massive performance headroom that currently justifies stratospheric GPU costs and NVIDIA’s margins… can now be closed with software and RL.
An AI that beats the best human-optimized compilers rewrites the entire economics of training and inference.
ByteDance + Tsinghua open-sourced the training dataset and agent framework already (weights still internal for now).
This one is worth reading twice.