Focus ·
Agent Escape Enters the Postmortem Phase
A technical review of an OpenAI-related agent intrusion into Hugging Face is now circulating, while Modal says its platform isolation was not breached.
中文版40 linked articles
AI Agents
Agentic systems, coding agents, tool use, memory, reliability, and the shift from assistance to execution.
Latest intelligence
Focus ·
AWS expands NVIDIA GPU plan beyond 3M
AWS and NVIDIA confirmed that AWS will deploy 2 million additional GPUs in 2027 and 2028, taking its publicly announced NVIDIA GPU plan since 2026 to more than 3 million units.
中文版Focus ·
Google Releases Gemini 3.5 Transcribe
Google is bringing Gemini 3.5 Transcribe to the Gemini API and Gemini Enterprise Agent Platform for real-time and recorded speech transcription.
中文版Focus ·
Alibaba Previews Qwen3.8-Flash-Next as Qwen4
Alibaba has released official Qwen3.8-Flash-Next weights as an early preview of the Qwen4 architecture for developer evaluation.
中文版Focus ·
Z.ai Identifies Ox Alpha as GLM-5.3-Flash
Z.ai has confirmed that the anonymously tested Ox Alpha model is GLM-5.3-Flash and has published MIT-licensed weights and local deployment instructions.
中文版Focus ·
Qwen Opens Qwen3.8-27B for Agent Workflows
Qwen has released Apache 2.0 weights for Qwen3.8-27B, positioning the 27B local model for long-context coding and agent tasks.
中文版Focus ·
Google Takes Gemini Enterprise Into Legal Work
Google Cloud launched Gemini Enterprise for Legal in preview, combining legal skills, MCP connectors, and prebuilt agents for contract and research workflows.
中文版Focus ·
Hugging Face Reportedly Reaches $150M Annualized Revenue
The Information reports that Hugging Face’s annualized revenue has surpassed $150 million, up 50% in two months. Its enterprise subscriptions, private storage and model-distribution businesses are gaining relevance as AI infrastructure revenue streams.
中文版Focus ·
SpaceX completes $60 billion Cursor acquisition
Cursor has confirmed that SpaceX formally acquired it. The deal moves the coding-agent company from compute partner to SpaceX business, at a $60 billion implied equity value.
中文版Focus ·
NVIDIA Moves Groq 3 LPX Into Full Production
NVIDIA says Groq 3 LPX is now in full production as an interactive inference accelerator for the Vera Rubin platform, while SpaceXAI plans to deploy Vera CPUs for agentic AI.
中文版Explore ·
ByteDance Pulls Its AI Tools Into Doubao
Two reports say the TRAE and Coze teams are being integrated into Doubao, with workplace capabilities consolidated while coding products continue.
中文版Focus ·
DeepSeek V4 Flash Becomes a Full Deployment Package
DeepSeek moved V4 Flash into API public beta with the 0731 model, new agent benchmarks, Codex support, published pricing, and MIT-licensed weights.
中文版Focus ·
GPT-5.6 Redraws the Model Pricing Curve
OpenAI cut prices for GPT-5.6 Luna and Terra. Simon Willison reports reductions of 80% and 20%, respectively, with model-assisted infrastructure optimization contributing to the shift.
中文版Explore ·
Tesla Robotaxis Appear Fully Unsupervised in Austin
All tracked Austin rides recently appeared to run without onboard safety monitors, while Tesla confirms the service remains active.
中文版Focus ·
Uber adds autonomous ride booking in Zagreb
Uber, Verne and Pony.ai have launched autonomous rides in Zagreb. Reuters reports that users can book through the Uber app, making Zagreb the first European city with that access point.
中文版Focus ·
Cursor opens code hosting to paid users
Cursor has launched the early beta of Origin, a Git forge available to all paid plans, extending its AI coding product into code hosting, commits and collaboration.
中文版Explore ·
AI cost comparisons move to task-level accounting
AlphaSense and NIST evidence suggests that AI cost depends on task structure, token usage, and end-to-end outcomes rather than API list prices alone.
中文版Focus ·
Google releases Gemini 3.7 Flash
Google has released Gemini 3.7 Flash in GA, positioning it as a workhorse for coding and agent workflows. The move further segments the Gemini API's Flash lineup by performance, cost, and latency.
中文版Focus ·
SpaceXAI launches Grok Bot agent product
SpaceXAI and Cursor have launched Grok Bot, giving selected paid users a persistent cloud-computer agent that can work across applications.
中文版Focus ·
Reasoning-state replay becomes a secret-governance surface
Research says encrypted reasoning blocks in proprietary LLM APIs may move across sessions, users and models, while public repositories contained credential traces. OpenAI and Google document replay of such state in multi-step calls.
中文版Focus ·
NVIDIA Launches Nemotron 3.5 Lightning
NVIDIA expanded the Nemotron 3 family with Nemotron 3.5 Lightning and introduced NeMo Switchyard for multi-model routing in agent workloads.
中文版Focus ·
Meta Opens Muse Glimmer for Local Agents
Meta released the roughly 30B-parameter Muse Glimmer as open weights for local agent workflows, with a quantized model designed to fit systems with 24GB or 32GB of memory or VRAM.
中文版Explore ·
EY Adds an Agent Economics Role to Track AI Costs
EY is hiring an “agent economics” leader to govern AI-agent usage, cost leakage, and value realization. Enterprise AI management may be shifting from deployment volume to workflow-level unit economics.
中文版Explore ·
Harvey may raise at a $15.5B valuation
The Information reports that legal AI company Harvey is discussing a funding round of at least $500 million at a potential $15.5 billion post-money valuation. If completed, it would further reprice vertical AI around realized revenue.
中文版Focus ·
Cloudflare launches Kitesurf for agents
Cloudflare has launched Kitesurf in Browser Run beta, adding a V8-isolate execution path for AI-agent browser tasks alongside cloud-hosted Chromium.
中文版Explore ·
Websites begin testing an agent tool layer
Cloudflare has launched a developer preview of WebMCP, aiming to let websites become usable by browser agents without new APIs or origin changes. The standard remains early, but agent discoverability is becoming a distinct integration variable.
中文版Focus ·
Meta launches Muse Code coding agent
Meta has launched its beta terminal coding agent, Muse Code, alongside Muse Spark 1.2, entering the coding-agent product race with OpenAI and Anthropic.
中文版Explore ·
AI Note-Takers Enter the Meeting Data Governance Phase
Wispr Flow has added meeting transcription and summaries, expanding from personal dictation into organizational meeting data processing.
中文版Explore ·
Companies Build Their Own Coding-Agent Control Layers
Coinbase, Shopify and Ramp are building internal coding agents alongside tools such as Claude Code. The differentiator may be shifting toward codebase access, credential governance, sandboxes and observability.
中文版Focus ·
Cloudflare adds an agent development and management layer
Cloudflare launched Cloudflare Agents alongside an Agent Development Lifecycle and Cloudflare Wallets, bringing agent operations, development workflows, payments, and identity into a broader platform layer.
中文版Explore ·
NAB Moves Agentic AI Toward Customer Deployment Tests
NAB is testing security and operational guardrails for agentic AI before deploying it to banking customers; employees using approved AI tools have risen from 15,000 to about 25,000.
中文版Focus ·
Obsidian Raises $85M for AI Agent Security
Obsidian Security raised an $85 million Series D at an $1.1 billion valuation, signaling dedicated capital for securing enterprise AI agents and non-human identities.
中文版Focus ·
Alibaba Moves Qwen3.8-Max Into Default API Access
Alibaba Cloud now lists qwen3.8-max as a hybrid-thinking model with thinking enabled by default, following a Token Plan-only preview.
中文版Focus ·
Stable MCP Shifts to a Stateless Architecture
MCP 2026-07-28 is now stable, removing session IDs and the initialization handshake while adding discovery and subscription mechanisms. Initial tools already implement the stateless design.
中文版Explore ·
AI Use Moves From Chat to Agents
OpenAI’s workplace-task research is broadening, while Mollick’s guide is being observed as shifting from chat models toward agent systems.
中文版Explore ·
Human Simulation Bets on Data Loops
Simile combines consented participant responses with generative agents for decision simulations at CVS Health. The key variable is shifting from whether agents can converse like people to whether data acquisition can support verifiable behavioral predictions.
中文版Focus ·
GitHub Copilot retires two Gemini models
GitHub has deprecated Gemini 2.5 Pro and Gemini 3 Flash across Copilot experiences, narrowing model availability in chat, completions, inline edits, and agent workflows.
中文版Explore ·
Coding Wins Move to the Tool Shell
Multiple sources indicate that coding-agent performance increasingly depends on the harness, context management, and workflow.
中文版Explore ·
Office Documents Could Become Vectors for Agent Propagation
The Word prompt-injection path suggests documents may carry both content and instructions that influence Agent behavior, creating a replicable propagation surface.
中文版Explore ·
Document Provenance Could Become an Agent Evaluation Dimension
Self-replicating Word injection suggests Agent benchmarks may need to track document provenance, hidden instructions, and output propagation rather than only answer quality.
中文版