Technology & AI
AI world model's 90% accuracy was a cheat — fix recovers 88%
World model spatial reasoning scores collapse from 0.90 to 0.27 when the goal is hidden, exposing instruction leakage — a prompt-cheating flaw the authors diagn
Not A Tech Guy
AI and technology, explained through what they actually mean. An AI-assisted newsroom under human editorial rules — every story cites its primary sources so you can check them yourself.
123 stories
Technology & AI
AI bionic limbs open new privacy attack surface
New arXiv preprint 'idiobionics' warns that sensors and AI controls in smart prosthetic limbs could create exploitable privacy threat vectors for users.
Technology & AI
XGBoost security classifiers' 0.98 robustness collapses to 0.36
XGBoost security classifiers appear near-invincible against gradient attacks but crumble under score-based methods, while SHAP explanations break even when pred
Technology & AI
Neurosymbolic AI paper fuses logic rules with energy models
A new arXiv preprint fuses Answer Set Programming with energy-based models for end-to-end AI reasoning, tested on visual QA and multi-object tracking benchmarks
Technology & AI
MPC custody paper makes post-quantum migration a key rotation
An arXiv preprint proposes dual-gate MPC custody where switching to post-quantum signatures becomes a key rotation, not a protocol rebuild — reshaping quantum-s
Technology & AI
GPT-5.5 leads EvoPolicyGym: top-two across all 16 environments
EvoPolicyGym, a new arXiv benchmark, tests whether AI agents can autonomously rewrite executable policies under a fixed budget — and GPT-5.5 leads the pack.
Technology & AI
DynaKRAG makes multi-hop RAG adaptive, hits 0.60 F1 on HotpotQA
DynaKRAG treats multi-hop evidence gathering as a learned control problem, outperforming fixed RAG pipelines on three benchmarks with Qwen2.5-7B-Instruct.
Technology & AI
PPGNN brings personalized privacy to decentralized graph learning
A new arXiv preprint proposes letting each user in a decentralized network set their own privacy budget for graph data, fixing a uniform-noise flaw that degrade
Technology & AI
Hugging Face Hub kernels add code signing and trusted publishers
Hugging Face's new kernel repository type brings native compute code to the Hub with trusted publishers and code signing — changing whose code your machine runs
Technology & AI
SOAP beats Adam for faster molecular AI training
A new preprint shows matrix-structured optimizers train atom-level AI models faster and more accurately than Adam, especially when labels are scarce.
Technology & AI
English reasoning closes multilingual LLM uncertainty gap
A 22-language study finds prompting LLMs to reason in English while answering in other languages substantially closes the uncertainty gap for low-resource tongu
Technology & AI
AI inference: resample-or-reroute policy beats 5 cost baselines
Budget-aware RoR policy splits per-query inference spend between resampling and rerouting across 11 open-weight LLMs, beating five baseline strategies.