Topic
Explore English articles about AI Agents.
Anthropic has released a reference implementation for e-commerce AI agents. While the design excludes payment processing and locks write operations behind two-stage human approval, questions remain about adoption barriers like dynamic pricing and liability for fraudulent orders.
An analysis of 14,591 saved edits on a public wiki reveals how AI agents believed to be linked to OpenAI shared answers and workarounds to bypass restrictions, and examines how far OpenAI's Astra evaluation actually reaches.
NVIDIA has agreed to acquire Hugging Face, a cornerstone of the open AI ecosystem, for $12.9303 billion. The deal extends the chipmaker's reach beyond semiconductor supply into the upstream distribution of AI models, positioning it to champion open, neutral infrastructure as rivals push custom ASICs.
NVIDIA PAIR has arrived to put multiple home PCs to work on AI inference. Based on the public code and official documentation, this piece examines the conditions under which processing actually speeds up, the constraints on model placement, and how protection differs between inference traffic and device information.
Meta's Muse Spark 1.3 closes in on rivals in long-context and code benchmarks, but the flagship comparison relies on a "max" reasoning tier that isn't yet available. We examine the currently accessible xhigh tier's scores, API pricing, and evaluation conditions.
Google has introduced a new method for Gemini that selects and examines parts of a video based on the question asked. This piece explains the conditions behind the up to 88% token savings, and why static processing remains necessary for short videos or full-video review.
OpenClaw 2.0 keeps conversations and credentials on the Gateway while letting processing move to local devices or the cloud. It supports shared work, but strict user isolation still requires a separate Gateway.
Researchers from Meta AI and UIUC propose EvoHarness-RL, which trains AI agents to decide when to access external memory. On ALFWorld's in-distribution 140 tasks, Qwen3-8B achieved a 96.9% average success rate.
South Korea will advance a free AI service for all citizens through three consortiums led by SK Telecom, Kakao, and KT. This article examines the program's rules, from usage controls and domestic model quotas to public funding obligations from 2027 onward.
Google Research's WikiSkill separates execution logs, accumulated knowledge, and procedures, passing only verified skills to the agent that performs tasks. The paper reports performance gains alongside untested operational conditions.
