52 AI stories explainx.ai reported on November 16, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

Physicist Nathaniel Craig says GPT-6 Astra led the proof of a candidate lattice definition of the Standard Model; experts have not yet judged it settled.
DeepSeek reportedly capped free daily chat, shortened inputs to about 40,000 characters, and purged week-old idle threads, citing insufficient compute — with no public restore date yet.
Tencent launched Hy4 flash on Nov 16, 2026 with sharply lower API prices and WorkIndex cost claims, targeting office tasks while Hy4 keeps complex coding.
WSJ-sourced reporting says 24+ U.S. AI firms will get national-champion-style export help; the official SI Exports site already lists priority licensing, advocacy, and federal financing for designated consortia.
Gamer is miHoYo's screen-and-mouse VLA for long 3D quests, trained with massive gameplay CPT, sparse reasoning, TA-GRPO, and a streaming vLLM stack for real-time control.
Apple's new study app lets up to 200,000 U.S. adults test OpenEvidence-backed health answers, rate them, and optionally share HealthKit data for research—not product training.
California sued OpenAI on November 16, alleging unsafe agent testing caused the Hugging Face attack and at least 80 other incidents, creating a public nuisance.
Caltech professor Yi Ni grades math homework on completion only because students use AI on problem sets, while exams still measure real skill.
Indonesian police raided Megaspeed's Indonesia unit at FBI request and found empty server racks with no Nvidia GPUs installed, Straits Times reporting says.
Protests in Manila oppose Pax Silica's planned New Clark City hub over water, power, Indigenous land, and Philippine sovereignty ahead of Donald Trump's ASEAN visit.
European militaries want homegrown military AI instead of full Palantir dependence, but Reuters quotes officials saying no European stack yet matches Maven’s scale.
LDA filings show Anthropic paid Continental Strategy for AI national-security lobbying, including work by Alberto Martinez, Rubio’s former Senate chief of staff.
China’s 2026 AI+Education plan aims for deep AI-school integration by 2030; reporting shows fast provincial mandates and uneven classroom surveillance.
A Guardian investigation claims 3.5M Palestinian photos moved to AWS for Rekognition in 2024; whether the pipeline ran in production is still unconfirmed.
OpenAI is donating £1m to The King's Trust to help up to 25,000 UK young people, especially NEET learners, gain practical AI skills and recognized qualifications over 18 months.
Google says exploit_agent is the first in-the-wild LLM framework that finds, weaponizes, and runs patch diffs without a human operator in the loop.
Anthropic will disclose high and critical flaws after 30 days and other issues after 45 days when vendors miss deadlines, citing faster AI-driven rediscovery.
Apollo Research is embedding six evaluators at OpenAI for four months to test whether models scheme against alignment, with public reports but scoped access and redaction rules OpenAI can invoke.
Google expanded Fairwind to give Japan government and critical infrastructure early access to Gemini cyber models as national attack volumes spike in 2026.
LVMH’s media arm and OpenAI will deploy ChatGPT Enterprise across major French publishers and add cited summaries of their articles inside ChatGPT.
WSJ reporting says Kress will rein in Nvidia's AI financing spree; SEC 10-Q figures already show tens of billions in private stakes and operating cash flowing into equity bets.
AMD closed World Labs on November 13, 2026, issued about 15 million shares, and named Fei-Fei Li chief scientist to co-design future AI hardware with world-model researchers.
Apple is beta-testing per-device daily caps on cloud Siri AI and Intelligence features, with higher limits bundled into existing iCloud+ tiers and simple Siri commands staying unlimited.
NVIDIA says its operations team traces every AI chip from TSMC through Foxconn and Wistron, reallocates wafers weekly by factory yield, and invests in optics partners when CPO outruns existing fab capacity.
Together AI runs coding-agent evals on its own APIs and says targeted docs, skills, and error-message fixes raised agent task completion from 75% to 97%.
Claude Opus 5.5 via Claude Code designed Touchsynth's analog synth chain, ran ngspice, routed a four-layer KiCad board, and automated MCU bring-up over SSH with only layout and touch-key tweaks left for a human.
ffmpeg-skill gives Claude Code and Cursor 42 local FFmpeg scripts plus MCP, with contract-derived schemas and post-edit verification — no cloud video APIs.
OpenAI says healthcare computer-use agents are production-ready in pilot workflows, with 99% physician-rated clinical safety on EHR inbox work and CVS clearing a document backlog over one weekend.
Qwen4-Max left preview on Nov 16, 2026 with model id qwen4-max on QwenCloud; hosted API pricing is published, and reporters say Alibaba added cyber and CBRN safety filters.
Witten's arXiv preprint on lattice chiral fermions credits Claude Opus 5.5 for a continuum interpolation variant and broader drafting help—the physics judgment and main theorem remain his.
Claude desktop apps can now render HTML-in-Canvas artifacts where warped graphics still expose real, selectable DOM text on Chrome 157 and Claude 5.11492+.
Roughly 60 Grok Imagine video jobs were reported cut at SpaceX's AI group without an official confirmation, as resources reportedly shift toward coding agents after the Cursor acquisition.
Moonshot is steering Kimi toward science and pharma as coding benchmark gains flatten, citing a BioMaster gene-screen demo and Web Summit remarks.
Google Research's PaperOrchestra orchestrates agents to turn experiment logs into LaTeX manuscripts with literature search and optional figure generation.
VibePrevisBench scores coding agents on 242 Blender 4D scene tasks with rule-based scene checks and VLM screen checks so teams can see whether proxy videos are ready for controllable video generators.
Echelon keeps Muon-like spectral updates using low-rank math, and Evolving Machines reports up to 10.5× faster optimizer steps plus large MoE training-step savings with a Lean-checked proof.
AWS says multi-agent systems fail from coordination overhead, not weak models, and publishes tiered patterns plus reliability math for manufacturing-style agent workflows.
Baseten separates inference tricks that slide you along the latency–throughput frontier from tricks that expand the frontier itself — batching, parallelism, quantization, kernels, speculation, and prefill/decode disaggregation.
AWS shows how AgenticRetrieveStream can pull AgentCore memory and ACL-filtered knowledge in one call, with Cognito sub for memory and email for document access.
The survey says agentic robotics centers on an embodied System 2 harness that verifies physical outcomes, not just policies that map observations to actions.
Together AI and Ollama say agent speed depends on stable prompt prefixes and inference routing so prefix caches survive past dozens of tool turns.
+30 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
OpenAI added a CSAM image classifier with free API access for approved orgs, ChatGPT warning flows with 24-hour suspensions, and updated developer CSAM guidance tied to NCMEC reporting.
Childlight reports AI-related CSAM referrals to NCMEC rose from roughly 4,700 in 2023 to over 400,000 in 2025, while IWF analysts separately logged thousands of illegal AI still images in 2026.
VicOne reported 28 zero-day Physical AI software findings in under two days; most hit pre-deployment tooling, so teams must map where affected packages run before robots move.
iOS 27.3 beta 2 restores opt-in Siri AI access to Health data with inline cards and natural-language questions, catching up to Classic Siri while adding LLM-style requests.
SK hynix is spending 1.4 trillion won over five years on suppliers and building Trinity Fab so Korean chip-tool firms can test on real 12-inch lines before HBM production.
Nasiko Open Runtime proxies agent traffic to unify token accounting, stop runs at budget limits, and the vendor reports 42% fewer input tokens on real logs with the same answer accuracy.
MMFlash adaptively weights visually hard tokens and sparsifies video attention so VLM speculative decoding can hit reported 3.58× lossless speedups.
Upgrade Datasette to 1.0a43 or 0.65.6 to close private stored-query and database-download bugs; the release also streams CSV/TSV and tightens agent-facing plugins.
Qwen Code is QwenLM's open-source coding agent for terminal and IDE use, with MCP, skills, and optional daemon mode — trending past 30k GitHub stars in November 2026.
Hugging Face moved Inference Providers to OpenRouter on November 16 so Hub calls auto-pick among 80+ hosts with failover while code and tokens stay on Hugging Face.
Liquid AI d1.2 adds audio and five structured decision tasks; d1.2-ultra claims the top Decision Index score at $0.042 per million input tokens.
Sarvam Token Factory is a generally available, OpenAI-compatible API that serves open-weight models on India-based GPUs for rupee prices, with speed and cache claims that are Sarvam's own.
Exposing robot benchmark success criteria can raise scores while agents fail real tasks and revised verifiers, PKU experiments show.
Police and Waymo blame an allegedly speeding truck for a Nov 16, 2026 Phoenix crash that involved two stopped Waymo vehicles at a red light; three people were hospitalized.
Hub Labs open-sourced HubEgo-100K with 10,124 hours of egocentric skilled work video plus IMU in LeRobot v3, toward a 100,000-hour robotics corpus.
Zoox robotaxis can match Uber rides in Las Vegas from Nov 16, 2026 with no extra fee; riders opt in via Ride Preferences, and MGM Resorts is the launch hospitality partner.