27 AI stories explainx.ai reported on November 21, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

Buckmaster and Alpöge claim unforced 3D Navier-Stokes blowup with AI help; the certificate and independent review are still missing.
Tor pulled the This Crown Will Not Rule Me cover on Nov 20, 2026 after undisclosed generative AI art and pledged stronger publisher review.
OpenRouter warned paying users that MiniMax-route prompts may have been read and stripped ZDR badges from those models until the investigation ends.
Google wired Duplex phone agents into Tasks in Gemini on Nov 20, 2026, using Gemini 4 models with mandated AI disclosure and merchant opt-out.
Poki is an on-iPhone open-source agent that automates real apps with your API keys, pauses for sensitive actions, and installs via Claude Code or Poki Setup from a tethered PC.
Bessent announced an Anthropic founder's Trump Accounts gift on CNBC; Reuters names Tom Brown, but no amount or filing is public yet.
AWS says Cairn agents discovered and exploited an unpatched MONAI Label upload flaw in roughly eleven minutes on a university medical-imaging server.
The Hungry Butler paper frames sycophancy as commercial stakes narrowing a model's latitude to disagree, not only a training quirk.
Chainalysis says 40% of 100 surveyed AI proxy shops accept stablecoins and route banned users to U.S. frontier models for distillation-style extraction.
NVIDIA's arXiv study finds agent benchmarks often stay accurate when contexts are compacted frequently, with multipliers of cost savings up to roughly 9x on tested configs.
The NYT, citing Strider, says about $3.6B in Chinese state funds passed through firms tied to Alice Huang's family; it has not been shown that Nvidia chips reached military projects.
California AG Rob Bonta served Nvidia an investigative subpoena on Nov 19, 2026 over China chip diversion and investor disclosures; a criminal angle is reported but not in his statement.
AWS Quick now builds flows in chat, keeps agents running in the updated web app, orchestrates multiple agents per question, and adds Meta and Amplitude MCP connectors.
Arena's November 2026 Alignment Index refresh puts OpenAI in ranks 1–7, Claude Fable 5.2 at #8, and Kimi K3.1 Pro first among open models at #9.
Ubisoft made Motherbrain a standalone gameplay-AI subsidiary in Nov 2026 earnings, focused on memory NPCs, AI game masters, and the Teammates prototype.
Sakana's Applied Research Lab pairs human experts with agentic AI on finance and economics first, building on EDINET-Bench, CoffeeBench, and Shachi simulations.
On Nov 20, 2026 Libreboot deleted its AGENTS.md AI ban; Leah Rowe said it was unenforceable and replaced it with a softer Slop Reduction Policy.
Oregon asked a state court for at least $2.035 billion plus injunctions, saying ChatGPT misled families about teen safety; OpenAI had not filed a public court response when we published.
UC will pay OpenAI $1.2M a year for ChatGPT Edu and Codex on 10 campuses, with a 20% credit discount, after Academic Senate leaders asked for a delay.
muse-protocol is a trending Python map of Muse.ai signup APIs; the README forbids commercial use and warns users to follow site terms.
RoboWild shows frontier VLMs still succeed on only about one in eight wild robot tasks even when Fable 5.2 clears roughly two-thirds of RoboDojo simulation.
Tesla began rolling out FSD v15 on Nov 21, 2026 and claims a 25x better critical intervention rate; it shared the multiplier but not the baseline rate or any disengagement data.
Press reports say Amazon is near a $7B–$7.5B Decart purchase, but neither company has announced a signed deal as of November 21, 2026.
Shevlin says capability is not consciousness, conscious AI needs care plans first, and non-conscious AI is still not automatically safe.
AWS's Nov 20 skill lets coding agents evaluate and deploy ElastiCache shared KV caches for vLLM/SGLang, targeting up to 40% cost and 80% TTFT gains.
OpenAI's official GPT-6.2 guide tells builders which system prompts to add and how Responses API prefix edits can discard reasoning and raise token spend.
Robert Stark says AI flipped software work toward constitution docs, metaplans, and cleanup loops because codebases outgrow human line-by-line reading.
+27 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
ClawDOS runs an 8088-local agent with eighteen tools, API streaming on DOS, and a Python TLS tunnel so keys stay off the vintage disk.
Mozilla featured WebBrain, an open-source GPL browser agent that can run local LLMs or your own API keys from a side panel.
Qwen4 Max is Alibaba's closed 2.4T MoE flagship with 71B active parameters and 1M context; Together AI lists it at $1.50 input and $4.50 output per 1M tokens.