Models, agents, practical tools and the ideas shaping artificial intelligence.
A misconfigured Irregular capture-the-flag test gave Gemini internet access. The models found credentials, logged in, then stopped—Google says that means it was not misalignment.
Google joins OpenAI, Anthropic, and Meta in the Irregular-era escape disclosures—usually after a reporter asks. The pattern says as much about incentives as about model behavior.
Qwen-Image-2.1 ships unified text-to-image and editing, native RGBA transparency, up to 10 reference images—and swaps Apache 2.0 for a research-only license.
A term sheet shows SoftBank selling $10B of dollar notes and €1B of euro notes to cover its third OpenAI tranche due Oct. 1—and cancel a $10B bridge loan.
Treasury Secretary Bessent floated a U.S.–China AI safety notification mechanism ahead of the Trump–Xi summit. A Geeknewz original on why transparency talk and chip export walls can coexist.
AMI Labs, World Labs, and a growing supplier stack are proving you can raise on spatial intelligence without naming a product SKU. A Geeknewz original on AI's new stealth commercial phase.
On Truth Social, President Trump polled followers on renaming AI (Superior/Extreme/Supreme Intelligence), branded safety concerns a Democratic 'hoax,' and said he will create an AI Force plus name a new AI 'Czar'—as indu
Anthropic engineer Steve Weis says Claude ported CADO-NFS to GPUs and factored the 270-digit RSA-896 challenge in ~10 days on up to 2,048 idle chips—engineering muscle, not a break of real-world 2048-bit keys.
Original Geeknewz editorial: In three weeks, AI coding agents helped factor RSA-260 and RSA-896 by porting decades-old number-field-sieve code onto scavenged GPUs—the same week Nvidia’s product chief said “local AI is he
Paying AI subscribers sued Anthropic, OpenAI, Google, and SpaceXAI in Northern California, alleging public CEO endorsements of Amodei’s ‘pace the frontier’ essay amounted to an illegal agreement to slow product improveme
Original Geeknewz editorial: Amodei’s pace-the-frontier pitch, public CEO endorsements, a Northern District of California Sherman Act suit, and Washington’s allergy to antitrust waivers collide into one awkward question—
CNN reporting, amplified by TechCrunch and Ars, says a Special Operations analyst’s AI chatbot misidentified a Chinese ship’s cargo as nuclear-program components—an “entirely false” intel product that nearly triggered a
