A weekly roundup of the Signal items we published this week (18–24 October 2026).
- Google now pays more than 400 newsrooms for AI-ready content — Google's News AI licensing pilot now covers more than 400 publications, including Axios, CNN and the Financial Times, which are paid for content used in AI features. Google (official blog)
- Google redesigns AI answers to push readers back to links — Google is adding more inline links higher in AI Overviews, a dedicated section for official pages, wider article carousels and clickable creator profiles in AI Mode. Google (official blog)
- Microsoft and Carnegie Mellon pair junior coders with AI-era mentors — Microsoft's new PRAISE programme gives junior engineers senior 'preceptors' plus a Carnegie Mellon curriculum, betting judgement matters more than typing code. Microsoft (official newsroom)
- Volvo signs up for Nvidia's eyes-off driving stack — Volvo will build its next generation of cars on Nvidia's Hyperion compute, Halos OS and Alpamayo software, aiming for gradual eyes-off, mind-off driving. NVIDIA (official newsroom)
- Meta's own survey finds Asia-Pacific small firms use AI but want guidance — A Meta-commissioned YouGov poll of nearly 4,000 small businesses in nine Asia-Pacific markets finds 84% use AI, while 86% say they need more guidance. Meta (official newsroom)
- DeepSeek ships a quieter Flash refresh — and puts web search inside its API — DeepSeek updated V4.1-Flash to build 1017 and added web search and fetch endpoints to its API, moving the model closer to agent-style retrieval. DeepSeek (official site / API docs)
- ZTE puts a 6TOPS NPU and an on-device model inside its home gateway — ZTE won three Network X 2026 awards for an AI all-optical lineup, including an FTTR gateway that runs Wi-Fi sensing and DDoS defence locally. ZTE (official newsroom)
- Alibaba's Qoder CLI moves into the note vault with an Obsidian plugin — Qoder released Qoderian, an open-source Obsidian plugin that lets its terminal coding agent read, reorganise and diagram a local Markdown vault. Alibaba Cloud (official blog)
- Meta's Muse Spark Open packs a 1M-token context into a coding-agent model — Meta released an 816B-parameter MoE coding model with 42B active parameters and a 1M-token window, tuned for long agentic coding sessions on Nvidia racks. NVIDIA Technical Blog
- Nvidia benchmarks its own laptop chip for coding agents: 4.6x over Intel — Nvidia published developer benchmarks for RTX Spark, claiming 4.6x coding-agent throughput, 2.7x fine-tuning and 1.8x compile speed versus Intel's Panther Lake. NVIDIA Technical Blog
- Gemini Notebook adds 60-second video overviews and harder quizzes — Google's study tool now generates minute-long video explainers, step-by-step overviews and short-answer or fill-in-the-blank quizzes grounded in a user's own notes. Google (official blog)
- Circle to Search will now read a suspicious text and flag the red flags — Google added scam review to Circle to Search on Android: circle a message and an AI Overview returns guidance and suggested next steps before you tap a link. Google (official blog)
- Anthropic puts $200 million behind a public benefit fund — Anthropic created a $200 million Public Benefit Fund, saying it will let the company pursue public-interest work with more flexibility than its PBC structure allows. Anthropic (official newsroom)
- SenseTime re-runs GPT-6-Astra on 12 embodied benchmarks and open-sources its own 27B model — Raising step limits tenfold turned nine RoboDojo failures into successes, SenseTime reports. It is releasing the test harness plus a 27B embodied model it says costs 1/50 as much. SenseTime (official newsroom)
- Qoder Fast routes a coding request by classifying it first, not by defaulting to the largest model — Alibaba's coding agent sends each request through a roughly 1,000-token classifier, then hands the work to a model ranked for that task type. The executor sees the full context; the router does not. Alibaba Cloud (official blog)
- Alibaba's agent team argues proactivity is an architecture property, not a prompt setting — An agent only acts on its own, the Qoder team writes, when it has its own identity, event sources and long-term memory, plus a stated goal to measure the gap against. Alibaba Cloud (official blog)
- DHS orders DeepSeek out of civilian federal IT, including products built on it — A Federal Register notice says the Homeland Security Secretary barred DeepSeek products from civilian agency procurements and ordered their removal from agency IT. Federal Register (DHS notice)
- Microsoft argues the real AI risk inside a company is the agents it already trusts — A Microsoft security post says internal agents, not outside attackers, are the harder problem, and proposes guardian agents that check every action before it runs. Microsoft (official blog)
- Gemini Notebook now shows its plan before it renders slides or video — Google added a review step to Gemini Notebook's slide decks and video overviews, letting users edit an outline or producer notes before generation runs. Google (official blog)
- NIH moves to create a dedicated office for its chief AI officer — A Federal Register notice says NIH proposes an Office of the Chief AI Officer inside its program-coordination division and is taking comments until October 28. Federal Register (NIH notice)
- Shanghai opens city-scale spatiotemporal data to outside builders, with a domain model underneath — Shanghai's planning bureau has opened a city-scale, desensitized spatiotemporal dataset to public builders through an 'open city' community, with SenseTime's planning-domain model and compute quotas behind it. SenseTime (official)
- Alibaba Cloud's security team maps an eight-step attack chain it says now runs at agent speed — After a run of leaked Access Key incidents, Alibaba Cloud's security team says attackers went from a single exposed key to full compromise in hours, with probing that adapts the moment it is blocked. Alibaba Cloud (official blog)
- DeepSeek's Flash refresh reaches Alibaba Cloud with a narrower API surface than its neighbours — Alibaba Cloud has added deepseek-v4.1-flash-1017 to its subscription plan; the model keeps a 1M-token context but drops JSON Schema, prefix completion, batch calls and fine-tuning. Alibaba Cloud (official blog)
- NVIDIA moves agent security into silicon, outside the agent's reach — NVIDIA's Sentry inspects agent traffic and enforces policy on BlueField-4 DPUs, so a compromised agent cannot disable its own controls. NVIDIA Technical Blog
- NVIDIA adds a reference humanoid and a robot runtime to its Isaac stack — Isaac gains a 1.4 m reference humanoid on Jetson Thor, an open-source runtime and Lab 3.0, plus 2,000 simulation assets and a 1M-clip human dataset. NVIDIA (official newsroom)
- An open protein-design model pairs generation with a search loop — NVIDIA and UW's Institute for Protein Design released RFD4-Proteina, an open BioNeMo model that couples protein generation with scoreable design objectives. NVIDIA (official newsroom)
- Google's AMIE takes the intake interview so the appointment goes to treatment — In a Google study, patients chatted with the AMIE agent before a visit; it condensed their history for the physician, who spent the time on treatment options. Google (official blog)
- An agent drafts an AI-factory digital twin; engineers still check the dimensions — NVIDIA walks through running a frontier model as an agent over OpenUSD to assemble a campus digital twin from site plans, reference assets and requirements. NVIDIA Technical Blog
- China's industry ministry puts numbers on its AI-plus-manufacturing push — MIIT says more than 30% of large manufacturers now use AI, 500+ industrial datasets are in place, and a humanoid-robot standards plan is coming. MIIT (official)
- A 9B grounding model edges past GPT-6 Astra, and the data ablation is the real claim — Daxiao Robotics reports ACE-Brain-0.6 winning six of eight spatial-grounding tests, and says new scene coverage, not more samples, drove the gain. SenseTime (official)
- Alibaba's OpenAnolis community is packaging ops runbooks as callable agent skills — OpenAnolis SkillHub says several hundred skills are queued for release; a featured one turns a dozen diagnostic commands and their correlation into one agent call. Alibaba Cloud (official blog)
- A 1.88-metre robot traffic officer takes a post outside Wuzhen's summit venue — Tongxiang police put an upgraded robot traffic officer at a Puyuan junction for the Wuzhen summit: lidar, stereo cameras, signal-linked gestures and eight-hour shifts. Cyberspace Administration of China (official site)
- Thousands of parallel agents turn a brain-wide CRISPR atlas into a reasoning map — OpenAI, PerturbAI and Scripps ran thousands of GPT-Rosalind agents over a whole-brain CRISPR atlas, producing a reasoning atlas in 32 minutes. OpenAI
- An agent found a critical PHP flaw first, and OpenAI wants a disclosure process for that — OpenAI says an agent found a critical PHP flaw before attackers did, and proposes a five-stage coordinated-disclosure process for the agent era. OpenAI
- Nvidia opens more of the Cosmos stack, down to 8B world models for policy tests — Nvidia added data-factory blueprints, Cosmos 3 Nano checkpoints for closed-loop policy tests and NIM services it says run up to 2x faster. NVIDIA
- Nvidia moves agent governance out of the prompt and into the runtime — Nvidia's OpenShell reference architecture treats every agent as untrusted, enforcing access through a gateway, supervisor, sandbox and policy prover. NVIDIA
- Google Cloud's A5X puts Vera Rubin racks under reinforcement-learning workloads — Google Cloud and Nvidia say A5X instances on Vera Rubin NVL72 reach general availability next month, aimed at agentic inference and RL at scale. NVIDIA
- Alibaba's Higress team open-sources HiRoute, a local router between agents and models — HiRoute sits between an existing coding agent and model services, picks a model per task stage and delegates sub-tasks to other agents, with routing and logs kept local. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- Splitting one coding task between two models cut whole-task cost by about 92% in Alibaba's own test — In two runs that cleared the same delivery gate, routing research to Qwen-3.8 Flash and the final design call to GPT-6 Astra cost 91.4–92.0% less than running Astra throughout. Alibaba Cloud Blog (Alibaba Cloud Native Community)
- Beijing folds the 'AI+' push into a wider plan to merge research with industry — A joint CPC General Office and State Council opinion on 21 October sets how science funding and industrial policy should interlock, treating AI both as a deployment target and as an object of new laws and sandbox rules. Ministry of Science and Technology of China
- Anthropic's Opus 5.5 matches its larger Fable model on most work at 40% lower serving cost — Anthropic says Claude Opus 5.5 performs at the level of Claude Fable 5.1 on most tasks, leads its agentic-coding benchmarks, and needs less compute to serve than Opus 5. Anthropic
- Claude proved assembly-level memory safety across 123 open-source libraries — Anthropic says Claude wrote RISC-V assembly proofs that 123 open-source programs always exit cleanly without memory-safety errors, and found bugs that survived trillions of test runs. Anthropic Research
- Claude generated 318,484 binder designs; 17 of 211 lab-tested candidates bound their scorpion toxin target — Anthropic ran an internal model for 64 hours on de novo binder design against scorpion toxins; two contract labs tested 211 designs, 17 bound, and it claims no antivenom. Anthropic Research
- Eight DGX Spark nodes pool 1TB of memory and serve a 753B open model locally at about 80 tokens a second — Nvidia now officially supports eight-node DGX Spark clusters, which scale unified memory to 1TB - enough to load the full 753B GLM 5.3 checkpoint on a desk-side setup. NVIDIA Technical Blog
- Nvidia wants AI factories integrated and validated in a factory container, not on site — Nvidia's DSX POD reference design moves rack integration, cabling and system-level validation into factory-built containers that connect to prepared site power and cooling after delivery. NVIDIA Technical Blog
- China makes agent and AI-terminal adoption a headline target in its new five-year informatization plan — The Central Cyberspace Affairs Commission's 15th Five-Year informatization plan sets 12 indicators for 2030 and adds adoption of AI terminals and agents as one of them. Cyberspace Administration of China (中国网信网)
- Alibaba's STAROps puts an agent on log inspection and keeps every query it ran — Alibaba Cloud describes an inspection stack where log-pattern reduction narrows the search and a planning agent decides each next query, recording the chain behind every finding. Alibaba Cloud Community
- Event-driven agents push the production bottleneck from the model to the context stream — Engineers at an Alibaba Cloud, IBM and InfoQ salon argued that for agents in production the limit is no longer model choice but how fast business events become usable context. Alibaba Cloud Community
- DeepMind's forest map now tells a plantation from a rainforest at 10-metre resolution — A global 10-metre map sorts land into six forest types and tracks yearly change from 2017 to 2025, agreeing with expert labels 90% of the time. Google DeepMind
- Nvidia hands robot-course setup to agents, but the walking is still learned by reinforcement learning — An Nvidia walkthrough uses planning and CAD agents to build a simulated obstacle course, then trains a locomotion policy across 4,096 parallel environments. NVIDIA Technical Blog
- TensorRT 11.1 tunes each engine against an accuracy budget instead of stacking heuristics — Nvidia's Global Performance Tuner searches scheduling choices per model and measures whole-engine speed under a numerical tolerance, not a fixed rule set. NVIDIA Technical Blog
- Nvidia's Python path into GPU memory skips the CPU copy, and checks that it really did — A KvikIO tutorial reads files directly into GPU buffers through GPUDirect Storage, and shows how to confirm each request took the accelerated path. NVIDIA Technical Blog