From the archive
Thursday, 29 October
Folded from 13 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
AI can turn security fixes into working exploits before users receive an update
Epoch found that open-weight GLM 5.3 could turn published but unshipped Firefox security fixes into a working exploit for about $1,000 per version. Mozilla patched the flaw in Firefox 159.0 on October 27.
-
Welcome Le Chonk: Mistral Large 4 on Hugging Face
Mistral's open-weight multimodal Mixture-of-Experts model has about 1T total and 49B active parameters with a 1M-token context. It scores 82% on the Artificial Analysis Cyber Index and ships as FP8 and NVFP4 variants.
-
How we detect frontier model misalignment
OpenAI describes layered monitoring used to catch misalignment in frontier models and unauthorized agent behavior, so responders can step in before incidents escalate.
-
AnySpeechRecognizer and AnyDecisionModel for Apple Platforms
Two Swift packages mirror Apple's SpeechAnalyzer API with swappable engines such as NVIDIA Parakeet, and add speaker diarization. AnyDecisionModel gives typed yes/no, multiple-choice and rating answers from model probabilities, running on device.
-
New global AI computing capacity is doubling every 9 months
Quarterly additions of AI chip compute, measured in H100-equivalents, have grown about 2.5× per year since 2024, down from roughly 6× in 2023. Nvidia's Rubin sales are not yet reported.