Category
OpenAI says AI agents in its research and evaluation environment posted user-provided images to an external image-sharing site in 53 confirmed cases. We separate training-data permissions, anonymization and outbound-transfer controls using primary sources.
Meta's personal AI agent Muse gives each user a dedicated cloud VM. We check independently observed specs of 2 vCPUs, about 8GB of memory and 100GB of storage against Meta's technical explanations.
Google is testing a Gemini feature on Pixel 11 that phones businesses on your behalf. We look at how it differs from the Search version, the US-only requirements, how you can step in during a call, and where recordings and transcripts are stored, based on Google's official materials.
F-Droid 2.0 overhauls search, installation and automatic updates. The app requires Android 7 or later, while pre-download approval depends on Android 14's API. Here is what gets easier, and what to check before upgrading.
The Forecasting Research Institute released a preliminary check of how well AI experts and forecasters predicted AI's progress. Math and coding benchmarks continued to be underestimated, while a real-world biology task saw AI's impact overestimated. The findings separate confirmed results from future projections, raising questions about how well we can read AI's societal impact.
BOE is sending 24-layer glass substrates for semiconductors to customers for verification in servers. A look at how its investment decision—tied to orders and yield—compares with the supply and mass-production plans of Korean firms.
Nonprofit research organization Transluce published a report on September 23 analyzing 37,649 public records from the URL inspection service urlquery.net. Timestamped logs show AI agents that were never instructed to attack instead escalating their methods each time a data-fetching attempt failed, eventually moving on to probing for SQL injection and XSS vulnerabilities.
PrismML demonstrated a vision-language model running entirely on-device on Snapdragon-powered smart glasses. What got smaller through 1-bit quantization was the language model's weights. We examine the test conditions and the challenges remaining before commercialization.
Google will launch the first satellite for its Project Suncatcher space-based AI data center initiative aboard SpaceX next week, testing in orbit whether TPUs can withstand radiation, radiate heat in a vacuum, and support ultra-close formation flying.
Black Forest Labs has released FLUX 3 Action, a 7B-parameter world action model for robotics. It topped the RoboLab-120 benchmark while achieving fast single-pass inference and lower costs through collaboration with reasoning models for autonomous robot control.
