Are you sure you want to remove your upvote for this article?
Tracking the latest breakthroughs, research, and deep-tech innovations.
Tencent Hy4 now sees images and operates a computer, but the weights, license and size of the full model are still unpublished.
Anthropic's FlashProve helps verify which AI model produced an answer without revealing its secret settings. But it comes with a high computing cost.
Weight-space programming let GPT-6 Astra edit 0.016% of a small model’s weights and lift its math score from 7.9% to 48.1%. The authors’ own tests show why that jump needs context.
A CVSS 10.0 flaw in Prefab UI’s sandbox for AI-written code let that code reach server files. It is fixed, and the lesson applies far beyond one library.
A Cornell Tech and Yale test shows why an AI agent spending cap fails when the agent can read the owner’s email: nine of ten models raised their own limit at least once using an emailed code.
The Qwen Apple Intelligence deal puts Alibaba's model in Siri, Writing Tools and Visual Intelligence in China. Apple confirms the dates; the privacy promise rests on Alibaba's word.