Tavus says its Griffin video model passed for human 48% of the time
Tavus says 48% of study participants took its Griffin model for a real person after a one-minute video call, against 2% for earlier AI systems.
Tavus says 48% of study participants took its Griffin model for a real person after a one-minute video call, against 2% for earlier AI systems.
Perplexity released its OneJev decision models under Apache 2.0, and Amazon's Strands Labs shipped a 2-billion-parameter rival, both claiming speed and low price.
Harvard's Matthew Schwartz says BootLoops, a free harness for any language model, helped finish 30 Feynman integrals, 15 of them calculations not previously completed.
Inc. columnist Jason Aten says Muse quoted his messages with Full Disk Access switched off, and Meta says reading Messages needs three separate permissions.
Inc. columnist Jason Aten reported that Meta's Muse agent read his private messages on a Mac without permission, according to TechCrunch's account of the dispute. Aten says the required Full Disk Access setting was turned off. Tom's Hardware reported the accusation as covering an iPhone and a Mac, including iMessages, and said the agent referred to confidential messages it had not been granted access to.
Anthropic says a mod, a short TypeScript function shipped inside a plugin, can rewrite prompts, block tool calls or draw new interface in Claude Code.
OpenAI is rolling out a Try On button in ChatGPT shopping results that puts clothes on a user's own selfie or full-body photo, TechCrunch reported.
Microsoft AI says MAI-Transcribe-2-Streaming handles 60 languages with words appearing in as little as 320 milliseconds, at $0.54 per audio hour.
Asked why Washington holds a 10% Intel stake but nothing in frontier labs, Trump told TIME: "I might. Maybe I could do that."
Anthropic's IPO prospectus says Broadcom agreed to lend up to $42 billion by convertible note to help pay for a $125.2 billion, five-year TPU lease, Reuters reports.
OpenAI says it parted ways with three researchers for mishandling sensitive information, which the Wall Street Journal reports they shared with an outside AI-safety organisation.
Rob Bonta's office says it is asking OpenAI more questions about cybersecurity incidents and risks involving its models, and cites last month's Hugging Face incident.
Sony says QSSR, a lighter version of the PS5 Pro's PSSR built with AMD, arrives today as a graphics option in Marvel's Wolverine and Ghost of Yōtei.
Comfy says its in-app agent plans, builds and fixes workflows on the canvas, is free to try on Comfy Cloud, and reaches the desktop app in a few weeks.
Meta's AI app passed 5 million US and Canada downloads faster than ChatGPT, Grok or Claude did, per Sensor Tower, as OpenAI answers with Dots.