|
🧮 OpenAI claims to have solved one of math’s “Millennium Problems”
LINK
|
- OpenAI said its internal AI model proved the Navier-Stokes equations can "blow up," a $1 million Millennium Prize Problem, but faces accusations it copied the approach from two outside mathematicians who got there first.
- The company said the proof took 88 hours and about 10,000 coordinating AI agents working in parallel, checked with Lean, software that verifies each step, producing a 100-page proof nobody outside OpenAI has reviewed.
- NYU's Tristan Buckmaster says he told an OpenAI mathematician about his year-long project with Anthropic's Levent Alpöge; days later, OpenAI's Sébastien Bubeck said an internal model had a matching proof, which OpenAI denies stealing.
|
🤖 Meta launches a personal AI agent
LINK
|
- Meta has launched Muse, a personal AI agent people can message to handle digital chores like sending emails, booking travel, or selling a car, all running inside a secure cloud setup the company says protects privacy from the start.
- Muse is out today on iOS and Android through a dedicated app, the Muse.ai website, and WhatsApp, with AI glasses support coming; it's free to try, but heavy use needs one of Meta's paid AI plans.
- The agent can buy things using Stripe's Link tool, which issues single-use card numbers so real financial details stay private, and it isolates each user's web activity in a virtual machine watched by a system Meta calls Sentinel.
|
🇺🇸 US accuses Chinese firms of copying AI models
LINK
|
- The U.S. government has accused six Chinese AI companies, including DeepSeek, Moonshot AI and Alibaba, of maliciously copying technology from American firms to train their own products more quickly and cheaply.
- Officials said the Chinese firms used a method called distillation, which trains smaller models using output from larger ones, to tap variants of models from Anthropic, OpenAI, Google and SpaceX, "likely with Chinese government awareness."
- The accusations landed just before Chinese President Xi Jinping's planned visit to the U.S. in late September, with officials warning the copying could boost China's military and cyberattack strength against the U.S. and its allies.
|
☠️ Anthropic researcher quits over AI fears
LINK
|
- Jacob Coxon, an AI researcher who worked at both OpenAI and Anthropic, quit Anthropic over ethics, warning that AI could kill countless people, possibly everyone, within the next decade if development keeps racing ahead.
- Coxon spent three years on pretraining research and accused both companies of "racing straight to self-improving superintelligence," saying executives privately share the fear that AI could wipe out humanity even as they build it.
- Anthropic team lead Evan Hubinger backed Coxon, putting the chance AI kills all humans within ten years above 10%, and admitting the company has no clear plan yet to keep a superintelligence aligned with human safety.
|
🪰 Google mapped a fruit fly's brain
LINK
|
- Google and the Howard Hughes Medical Institute's Janelia Research Campus published a complete wiring map of an adult male fruit fly's brain on September 3, capping a decade-long project to trace every neural connection.
- The map, called a connectome, charts more than 166,000 neurons and 125 million synaptic links, making it the largest brain map by neuron count and a resource for studying how nervous systems perceive and react.
- To build it, scientists sliced the fly's nervous system into millions of thin sections, imaged them with an electron microscope, and used AI to turn the 2D pictures into verified 3D reconstructions of the neurons.
|
🎨 ChatGPT image generation gets faster and sharper
LINK
|
- OpenAI has launched ChatGPT Images 2.5, an update to its image-generation tool that produces pictures up to 50% faster than before while following instructions more closely across several rounds of edits.
- The update keeps details stable during edits, holding onto facial features, subjects and backgrounds the user didn't ask to change, and it adds Sketch, a feature that turns finger or stylus drawings into reference points for generated images.
- It comes in two versions: GPT-Image-2.5 Flare, the default that makes images two to four times faster with better transparent backgrounds, and Sunburst, which is slower but sharper for product photos and marketing work.
|
|
