All Stories — Page 21

Alham Fikri Aji's 423-question benchmark spans 13 languages and tests whether agents can connect evidence across web pages, video, maps and documents.

MEMOIR-VLM combines MRI, diffusion imaging and clinical scores, but its strongest diagnostic signal came from cognitive tests.

Reverse engineer Maurice Heumann says the private project took an estimated 600 billion to 700 billion tokens and used a machine-checkable test to catch plausible but incorrect code.

The model uses a 15-second recording to recreate a person's appearance and movement; HeyGen's claims of leading realism rely partly on its own testing.

Microsoft Execution Containers applies developer-defined file and network rules at runtime, with Windows 365 support also listed as generally available.

Joshua B. Plotkin's satirical video credits Claude for the lyrics and video code, and Suno for the music and vocals. Its AI narrator claims the proofs, then leaves humans a few easy problems.

Uri Katz and four co-authors built the 8B model on curated Yiddish text after finding that a common web corpus misclassified Hebrew and machine-translated pages.

The 200-scene benchmark found even its top-ranked configuration scored lower on dynamics than on appearance and geometry.

Guideline 4.2.6 leaves room for customized and aggregated apps, but separates building them from deciding who submits them.

In an October 10th essay, he argues that companies should separate model intelligence from the permissions and controls governing its actions.

The October 10th claim comes as Texas records list 319 Cybercabs statewide, while federal regulators examine the vehicle's safety certification.

Musk described uploading a card photo and said Grok bought LEGO from its website; the product's documented checkout uses an approval step and Stripe Link.

The feature lets Copilot use desktop interfaces across its CLI and Mac and Windows apps, but is off by default.

Nvidia's early talks could lead to an acquisition, acqui-hire, larger investment or computing deal with Reflection AI. Nvidia has already invested $800 million; Reflection's Beam model weights are due later this month.

An X post alleges that someone took its "keys" during a livestream coding session and used them to run Claude Fable 5.1 sessions, but provides no details confirming the activity.

A Quilty representative says the connector offers trend reports, IP intelligence and slate data in addition to screenplay analyses.

Founders Benjamin Martindale and Minghao Tan say Donkey, their San Francisco startup, uses customs records and factory quotes to price delivered orders for smaller buyers. Its Oakland savings example is company-reported, and public materials disclose no customer count or revenue.

Mark Hughes says QA consumed about 90% of team effort while training grew faster. Solidroad is shifting to interactive replicas of workplace software.

In an October 10th CTech opinion piece, Yonatan Boguslavsky argues that teams need shared context, workflow controls and clear human handoffs to coordinate agents.

Josh Gao's startup is building human-motion data and deployment tools as robotics labs spend heavily to train machines outside the lab.

The form was part of a test on randomly selected websites; Anthropic's October 9th report also details other unintended actions on live sites.

The October 9th release pairs the self-rewrite with company-reported startup and memory gains, Windows beta support, and Homebrew installation.

CEO Mukesh Khare brings IBM chipmaking and quantum experience to a foundry designed to serve IBM and outside customers.

Led by Applied Digital co-founder Wes Cummins, Base Electron is developing a planned 1.2-GW North Dakota project under a 15-year power agreement with Applied Digital.