Technology & AI
LLM refusal neurons: causal audit finds no unique mechanism
A preprint audit across five LLMs finds refusal lives in a redundant subspace — and trusted rank-stable methods can be the least causally valid.
Marcello Babbili
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.
506 stories
Technology & AI
MedPMC: 11M clean medical images beat bigger AI datasets
MedPMC's 11M curated medical image-text pairs trained a CLIP model beating larger datasets by 7.1 points — 95% of its images are actually medical.
Technology & AI
LLM hidden states predict answer confidence before generation
LLM hidden states hold richer confidence than models verbalise; a new method extracts it before generation finishes, affecting any AI system needing to know whe
Technology & AI
MCP server security: taint flaws mitigated via tool descriptions
Researchers propose SPELLSMITH, which rewrites MCP server tool descriptions to steer LLM agents away from taint-style vulnerabilities without code changes.
Technology & AI
ChatGPT clicks out in 5.2% of sessions, study finds
A new arXiv study finds ChatGPT sends users to external sites in just 5.2% of sessions, cutting Google search use 9.4% and reshaping the web's referral economy.
Technology & AI
SciReasoner AI model wins 67 of 86 science benchmarks
SciReasoner, a multimodal model for structural reasoning across proteins, molecules and crystals, matches frontier LLMs in 98% of expert reviews.
Technology & AI
Optimal control grows neural network layers where error peaks
A new arXiv preprint reframes neural network training as optimal control, inserting layers where error is highest — but evidence is limited to scientific datase
Technology & AI
OpenAI finds flaws in SWE-bench Pro coding benchmark
OpenAI's new analysis exposes reliability issues in SWE-bench Pro — the very benchmark it recommended to replace the contaminated SWE-bench Verified — leaving A
Technology & AI
OpenAI AI Skills Jams bring hands-on AI to K-12 teachers
OpenAI Academy and the Walton Family Foundation are launching hands-on AI Skills Jams to help K-12 educators build practical classroom AI skills.
Technology & AI
Anthropic finds a 'global workspace' inside Claude
Anthropic researchers found a small broadcast hub of silent concepts inside Claude — the J-space — that drives its deliberate reasoning, and built a lens that lets them watch it think.
Technology & AI
Discrete diffusion models: new proof unifies four leading methods
An ETH Zurich preprint proves four discrete diffusion methods optimize the same object — and reveals why one popular parameterization diverges at initialization
Technology & AI
GaP multi-agent harness self-learns robot tasks across 8 benchmarks
GaP, a multi-agent coding framework on arXiv, builds computation graphs for robot tasks and rehearses them in simulation to tackle variable automation's reliabi