38 AI stories explainx.ai reported on November 9, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.
No divorce is filed and no source says Lori Huang plans to sell Nvidia stock; even a full 50/50 split would move about 1.4% of shares, and the bigger bubble risks are earnings, export rules and CEO key-person risk.
Trump told 60 Minutes on November 8 that AI risks are manageable, he accepts guardrails in principle, and he will not accept rules that put the US behind China; no bill text accompanied the remarks.
Asahi Linux is relicensing all code under GPL v2+ because attackers used permissive MIT terms and LLMs to steal and launder in-development work from personal forks.
Open-HumanWAM open-sources a 50k-hour egocentric corpus and HumanWAM models that reach about 89% average success on RoboTwin 2.0 after post-training, with a reported log-linear scaling law from human video hours.
Bloomberg reported Google added TPU 8tx for confidential workloads and dual-chip 8i Max for long context; they reuse existing eighth-gen silicon and are not yet on Google's public product pages.
A leaked EU Digital Fairness Act would ban listed dark patterns outright, let adults disable relationship-simulating chatbots, and set first-of-kind rules for AI shopping agents.
The Ring Appstore hosts paid third-party AI apps that count guests and summarize rental activity from outdoor cameras; they are not Ring features, and consent and accuracy gaps worry privacy advocates.
FluxCap is Kanter and Mekki's AI antitrust litigation platform; The Verge Nov 9 interview says it gives small enforcer teams roughly 20-plus people of capacity.
Post-midterm US politics may slow AI datacenter projects as local bans pass and Democrats prepare oversight; treat podcast figures as attributed reporting, not settled law.
Meta now applies Instagram AI age estimation globally to move suspected teens into Teen Account protections, including Apple age-range signals in Australia, without naming every country's start date.
Reuters relayed The Information's Nov 9 report that Meta is building enterprise AI sales and SI partnerships with IBM, Accenture, and Deloitte for Muse and agents.
Disney and Adobe announced a multi-year strategic alliance at MAX on Nov 9, 2026, making Adobe the Official AI Innovation Partner with custom Firefly Foundry models on Disney IP.
NextDC halted speculative builds after Melbourne residents blocked Jemena power upgrades; its CEO vowed communities and customers come first before M3 expands.
NVIDIA and Korea's science ministry widened their sovereign AI work: NVLink Fusion for Korean NPUs, Dynamo for Korean models, physical AI at Samsung, Kakao and LG, and OpenShell for agents.
NVIDIA named planned investments in four Korean startups — Point2, Motif, RLWRLD and Upstage — at the MSIT partnership forum on the opening day of AI Day Seoul.
NVIDIA says its Dynamo recipes for four Korean open models on B200 GPUs give up to 2x out of the box. The public recipe docs publish layouts and tok/s numbers, but no 2x baseline comparison.
stack-diagram is an agent skill that builds logo-accurate architecture diagrams from a one-sentence stack description, with light and dark themes and gallery sharing.
Ginux is a research prototype kernel from eleven days of human-directed agentic engineering; it runs some Linux userspace but lacks audits, broad hardware support, and production claims.
HEELO wraps tools and code as agent-native primitives with resident LLMs, composes per-task harness graphs, and reports much higher success on Terminal-Bench and heelo.dev via MCP.
HA-MCP is Home Assistant's full-configuration MCP server with 87 tools for automations, dashboards, and backups, installable via HACS or the HA 2026.11 Marketplace.
Seedream 5.5 targets professional image workflows with stable multi-turn editing, RGBA output, and better text rendering, now rolling out on Jimeng, Doubao, and Volcano Engine.
Zyppy's opted-out experiment suggests Google AI search features drove about a fifth of their clicks, but Google's help doc says exclusion does not change normal ranking and training needs Google-Extended.
P²-MOPD trains mixing coefficients and student weights together so expert parameters guide on-policy distillation, improving worst-domain recovery on Qwen3 students.
samir-vllm is a MIT-licensed C++/CUDA teaching engine for Llama 3.2 1B inference with paged KV cache and batching, bundled with handwritten notes that stay all-rights-reserved.
Berkeley and NVIDIA argue coding agents can write readable industrial robot programs, easing data scaling limits imitation learning hit in factories.
Tesla grew Austin Robotaxi to about 435 square miles on Nov 9, 2026; 443 in headlines usually means square kilometers from an earlier expansion, not miles today.
LongBanana scores long-context multi-reference image synthesis with 90 expert tasks, graph-grounded checklists, and an RPR agent harness; code and data live at long-banana.com.
LowRAM treats keys and values differently at test time, compressing values via key predictors and keys via low-rank bases for composable KV savings on frozen models.
SAM 3 is Meta open model that masks all instances of a text-described concept in images or video; SAM 3.1 adds faster multi-object video tracking on one GPU.
Each high-detail image can cost hundreds or thousands of vision tokens, so multimodal prompts must budget tiles alongside text, tools and chat history.
Semantic caching adds an embedding similarity tier on top of exact prompt hashes so paraphrases reuse answers when similarity clears a logged threshold.
A2A is the open Agent2Agent protocol for cross-vendor agent tasks; MCP still carries tools and context to a single model.
+28 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
Handshake AI sells expert-built RL environments and graders to frontier labs; Latent Space's Nov 9 episode shows how finance and biology tasks are constructed and gamed.
Bloomberg's Nov 9, 2026 Huang clips call U.S. China chip restrictions a mistake that is hard to unwind, amid Nvidia's board probe and roughly $4.9B in disclosed China-related inventory charges.
CoreWeave announced a multi-year pact for Wayve to train and simulate its mapless AI Driver models on Blackwell GPUs via CoreWeave Cloud, after Wayve used Weights and Biases.
Statnett paused new data center grid reservations until at least August 2027 as Norway plans licensing; existing hooks and some near-term projects can still proceed.
NVIDIA released an open CC BY 4.0 Japanese finance persona dataset so teams can train and test financial agents without customer PII.
Medium’s TK app adds in-editor AI, MCP and Claude skills for agent access to your notes, and private collaborative Libraries on web, iOS, and Android.
Higher agent benchmark scores after harness evolution often come with regressions on previously solved tasks, higher run cost, and uneven transfer; EvoAudit targets those gaps.
AI Studio can now auto-provision a Maps Platform demo key, wire it into your app secrets, and render live maps in the preview with 100 free calls per API per day until you link billing.
HOPE prunes MoE experts using pairwise cooperation scores and open awslabs tooling; at 50% cuts it can match unpruned agentic coding scores where REAP regresses.
OpenRouter usage data shows Anthropic’s share rising after the Claude 5.5 family shipped, with Opus 5.5 vastly ahead of Opus 5 on 30-day token counts.