Transfer learning survives three adversarial attacks
Mina Remeli's dissertation tested whether strong source-task performance guarantees transferability, and three attack types couldn't break the link.
The top 0.01% of AI knowledge, and what it changes for your work
Daily AI and technology coverage — models, chips, agents and research, and what each development actually means for you and your business.
732 stories
Mina Remeli's dissertation tested whether strong source-task performance guarantees transferability, and three attack types couldn't break the link.
Flide, a new inference engine from NUS and Oxford researchers, runs decision-model queries up to 9.5 times faster than vLLM and SGLang.
Apple researchers show that observable batch sizes break DP-SGD privacy guarantees and propose a dummy-point mechanism to fix it.
Impeccable, a Rust-based open-source tool, uses 62 deterministic rules to catch AI coding agents' design mistakes before they ship.
Screenpipe, claiming YC S26 backing, captures screens, audio, and keystrokes to feed AI agents, but its claims remain unverified.
Locally Uncensored, a free open-source desktop app for local AI chat, image and video generation, trends on GitHub as v3.0.8 ships.
OpenMontage turns AI coding assistants into video studios, trending on GitHub with 74,127 stars and example projects costing under $5.
Google uplinked a haiku prompt to a Gemma 4 model on a TPU in orbit and got a response, but no one outside the company has verified the milestone.
Matt Pocock's AI agent skills repo hit GitHub's daily trending list with 307,711 stars, pushing disciplined engineering over vibe coding.
CocoIndex, trending on GitHub with 11,781 stars, claims to give AI agents fresh context from codebases and Slack using incremental processing.
LLMs produce near-optimal solutions to optimization problems but persistent numerical errors limit their use as standalone solvers.
A preprint centers distributionally robust optimization on a learned density model, lowering cost under distribution shift when data is scarce.