10:19 PM
Today in brief.
OpenAI dumps 722 AI-written math manuscripts on GitHub
An unreleased OpenAI model produced 722 math manuscripts across 372 result families, and OpenAI put the whole pile, plus reasoning traces and Lean proofs, in a public repo. Anthropic and Claude had their own math results the same day.
EmbeddingGemma 2 searches your video and audio on-device
Google's new open embedding model maps text, images, audio, video and code into one space at 740M parameters, small enough to run fully on a phone or laptop.
Boston Dynamics names Rohit Prasad CEO to push Physical AI
Boston Dynamics hired Rohit Prasad as CEO to speed up AI plus robotics, the same day it pitched Spot for pharma plant monitoring.
2 sources 2:35 PMCritical bugs jumped from 100 to 600 a month, and labs arm defenders
a16z data shows reported critical vulnerabilities at 21 big software companies spiked sixfold since spring, and most exploited bugs are now attacked by the day they go public. Anthropic is widening access to its top models for security pros.
4 sources 3:27 AMTry today: Opus 5.5 UGC ads, an AI-built TPU, free Claude Team
An agent that turns an Airbnb listing into a UGC video, an open-source accelerator designed by AI that runs Qwen on an FPGA card, and a free year of Claude Team for startups.
Mistral Large 4: a 1T open-weight model built in Europe
Mistral's Le Chonk is a 1 trillion parameter, 49B active multimodal model. The API preview is live now and weights are promised by the end of October.
Two-year Khanmigo trial: AI tutors work only if kids use them
A randomized trial in 18 Tennessee middle schools found Khanmigo raised math scores about as much as plain Khan Academy, because most students barely talked to the tutor.
AI datacenters are killing the $100 smartphone
Memory makers are shipping chips to AI datacenters instead of phones, so cheap phones are getting pricier or disappearing, with 800 million people exposed.
Utah lets an AI prescribe acne meds with no doctor in the loop
Utah approved a one-year pilot where Nolla Health's app scans your face, scores your acne and prescribes from eight approved treatments, with human review phased out over time.
Websites start locking out AI agents
Consumer agents like Meta's Muse keep hitting walls: Amazon now blocks Muse outright, and Wikimedia says OpenAI bots may be linked to a May outage.
No stories in this category for this edition.