Connectome releases enable public torture experiments
Open connectome datasets and model‑steering techniques enabled rapid public "torture chamber" experiments that proliferated and attracted funding, prompting platforms to prohibit sustained abusive behavior toward models. — lesswrong.com
Quad9 exits France
Quad9 will withdraw its recursive DNS service from users in France after a court ordered region‑targeted DNS blocking; compliance is technically unfeasible and conflicts with Quad9’s privacy and neutrality principles. — quad9.net
Trivial eMudhra CA keys exposed
Several eMudhra intermediate private keys used trivially small values that allowed instant key recovery from public certificates, leaving many non‑browser TLS clients and code‑signing checks vulnerable. — scotthelme.co.uk
LM psychology treats models as quasi‑minds
Large‑model psychology frames models as having personas, introspection, and emotion‑like axes to predict, elicit, and evaluate behavior, making human‑level descriptions a pragmatic tool for understanding and managing models. — lesswrong.com
Lobsters retires vibecoding tag and tightens rules
Lobsters deactivated the overloaded "vibecoding" tag, introduced an ai‑tooling tag, barred direct code‑repo submissions, and directed low‑effort or AI‑generated items toward the spam flag to curb abuse. — lobste.rs
PHP issues urgent security release
PHP 8.2.35 fixes critical remote‑code‑execution vulnerabilities and administrators are urged to update immediately. — github.com
GRAFT localizes reward‑hacking to removable adapters
GRAFT confines learning updates from detected reward‑hacking to a removable adapter so that ablating that adapter at deployment eliminates learned hacks while preserving intended task performance. — lesswrong.com
MiniMax M3.1‑Flash multimodal model release
MiniMax‑M3.1‑Flash is a native multimodal mixture‑of‑experts model focused on coding, agentic workflows, and cybersecurity that improves efficiency and benchmark performance. — huggingface.co
Design, not language, usually limits speed
Most application slowness stems from I/O‑bound designs and unnecessary work; profiling and targeted redesign—caching, async work, and query fixes—usually yield larger speed gains than language rewrites. — seangoedecke.com
eMudhra to exit global SSL/TLS business
eMudhra will withdraw from the global SSL/TLS certificate business to sharpen focus on enterprise identity and other trust services while supporting customers through a managed transition. — emudhra.com