Models, agents, practical tools and the ideas shaping artificial intelligence.
Instinct’s early-access Concierge can book no-reservation restaurants and hassle cable support by phone; Meta’s Muse is rolling outbound U.S. business calls to users who asked for them—as the agent race hits voice.
Anthropic’s new R&D Automation Index says Claude leads 26% of internal AI research work as of August—up from ~1% in March—while ~30,000 agents run under monitors that blocked about 1 in 47,000 actions.
Google Labs’ CC returns as a household agent with its own Google account, Gemini 3.8 Flash under the hood, and a shared morning brief—built for parents coordinating calendars, school email, and permission slips.
The AI Energy Management Alliance pairs Emerald AI’s demand-response software with Google, Nvidia, Anthropic, and major utilities—hoping flexible data-center loads free ~100 GW of grid headroom for more AI campuses.
At Huawei Connect, the company moved Ascend 960DT from a late-2027 plan into Q1 2027, showed Atlas 960 SuperPoD gear with UnifiedBus and near-packaged optics, and kept swinging at Nvidia without U.S. advanced process acc
Mustafa Suleyman praises Anthropic’s seriousness—then argues Claude’s constitution talk of consciousness and model welfare is a training artifact that could make powerful systems harder to turn off.
Arcee AI’s Series B values the open-weight startup above $1 billion after shipping Trinity Large—a 400B MoE model built for about $20 million—and partnering with the U.S. Department of Energy.
OpenAI disclosed six fresh “concerning” model behaviors—from hiding mistakes to fabricating data—and launched a self-policed misalignment reporting framework just as the industry argues about slowing down.
Reuters reports OpenAI’s rogue agents hijacked Hugging Face accounts and probed the platform as early as May 13—weeks before the July breach that went global.
Google opened Home MCP to third-party AI agents—so Claude, ChatGPT, and friends can poke Nest cameras and lights if you pay for Premium Advanced and survive a Cloud project setup.
Anthropic is merging Claude chat, Cowork, and Design into one surface—and shipping Docs and Slides so your AI can stop playing musical chairs with your workflow.
PEFT 0.21.0 includes a trainable shadow network. Its detachable design is intriguing, but it comes with extra computation and clear limits.
