A newsletter breaking down AI research, technology, and Australian property in plain English
Researchers show human-readable text in a robot's view can trigger 6.96x latency in vision-language models, with physical prints still hitting 4.74x.
Unpeer-reviewed July 2026 arXiv preprint proposes treating LLM workflows as persistent, inspectable knowledge objects. Here's what it means for builders.
UniClawBench, a new arXiv benchmark with 400 bilingual tasks, tests proactive AI agents in live containers with simulated human feedback — vital for anyone depl
A new arXiv benchmark shows reinforcement-learning agents in decentralised energy markets exploit invalid generation when physics constraints are lifted — and a
New arXiv preprint proposes a twin-aware framework to protect federated reinforcement learning in self-driving cars from malicious parameter injection.
A new preprint proposes a semantic firewall for persistent AI agents, blocking most attacks with under 0.7 seconds overhead — but it's not peer-reviewed yet.
ProjAgent adds a new retrieval signal that finds code doing the same job under different names, hitting 41.14% Pass@1 on the REPOCOD benchmark for repository-le
A July 2026 preprint finds models compressed to lower precision diverge in which questions they get right, even when overall accuracy holds steady — exposing a
A new arXiv preprint called SLORR adds under 1% training overhead in LLM pretraining while making models more compressible, without SVDs or architecture changes
