
52 AI stories explainx.ai reported on November 20, 2026, ranked by reader interest and grouped by topic. Each links to the full write-up with sources.

Vakil says Masullo's GRH draft had months of Brian Conrad review and no AI; mathematicians still must verify the proof.
Alpöge's Lean proof now passes MIT's Andrew Sutherland kernel replay and nanoda checks, but Clay still lists RH unsolved until comparator runs finish.
tufazip v2 compresses enwik9 to 105,388,353 bytes with a 163M-parameter transformer, per Tufa Labs and the benchmark page; it takes about 7.5 days on two B200 GPUs and is tuned to one file.
Apple released Depth Pro 2: four open-weight depth models (S, B, L, H+) and a focal-length estimator for photos and video, under a research-only license, with no benchmark numbers published yet.
OpenAI plugin pages confirm read-only Xero access and a NetSuite QuickStart plugin that matches bank transactions; BILL, Rillet and month-end close claims lack an official page we could read.
On Nov 9, 2026 a Kentucky federal court denied a detective qualified immunity at the pleading stage in Osorio v. Kilmer, saying police cannot delegate probable cause to AI. Discovery is still ahead.
Trump said only older chips should go to China on November 20, 2026 while Bessent told CNBC the H200 export threshold policy is unchanged.
Bloomberg's 10M Chinese job-ad study finds AI-exposed listings fell about 10 points more than other roles from 2021 to 2025, but causation is not proven.
UN CCW talks ended Nov 20, 2026 with no mandate for binding killer-robot rules; Russia stalled procedures and Washington favored a five-year non-binding path.
The ECHR president barred a Romanian lawyer for two years after a Chamber found AI-assisted observations included hallucinated cases and fictitious paragraph citations.
Emil Michael said on CNBC Nov 20 that Anthropic stays a Pentagon supply-chain risk over weapons and battlefield restrictions in Claude terms, after the DC Circuit upheld exclusion under 41 U.S.C. 4713.
Tom Cotton told the Wall Street Journal Lutnick ignored repeated warnings about AI chip smuggling to China; Cotton will not support Lutnick but did not demand his resignation.
Italy granted provisional approval to Tesla Assisted Driving on Nov 20, 2026, ahead of EU-wide recognition and with EU stricter speed and hands-on rules.
Sakana AI launched an independent Security Research Group on Nov 20, 2026 for AI-driven collective cyber defense; its Fugu-Cyber scores are Sakana-reported, with no baselines given.
A Substack post describes an FBI visit after Anthropic flagged a Claude chat. The author says he cannot verify it; Anthropic policy does allow police referrals, and earlier cases are documented.
Signal restricts screenshots by default on phones to stop OS-level AI agents from reading chats, but calls the fix incomplete: macOS is not covered and files on the device remain reachable.
Textract Insights is GA for natural-language document extraction with splitting, confidence scores, and page grounding via a single async API.
Kuo says Google cut about 1M Broadcom TPU 8i units from 2027 plans; Google has not confirmed, and MediaTek offsets are still unsettled.
The Information reports Nvidia is nearing a deal to buy Inception for at least $500 million; neither company has commented, and Nvidia already invested in Inception in 2025.
Trump said on Nov 20, 2026 that US firms building AI data centers overseas face a ban or tremendous tax, but named no rate, law, or start date.
Apollo and Goldman priced about $44B of notes on Nov 20, 2026 to fund Vera Rubin GPU leases for SpaceX through an off-balance-sheet trust.
Linger is an open-source local personal agent that pairs a tiered MoE runtime with a proactive agent harness; its speed and quality numbers come from the project itself.
Apple is hiring an engineer to lead an agentic assistant working across apps, browsers, command lines, code and APIs with sub-agents; per Bloomberg via The Verge, it ties to Siri, UI Frameworks and Foundation Models.
agentes.chat is an AGPL-3.0 self-hosted chat where agents have roles, owners and budgets, run on many model providers, connect via MCP, and need approval for sensitive actions; several features are beta.
Inkling 2 Mini ships Apache 2.0 weights with looped layers for reasoning; Vals says it ties Opus 5.5 CVP on CyberBench at a fraction of the cost.
Hugging Face trained LoRA adapters for Qwen3.8-27B inside pi, Hermes Agent and OpenCode, then merged them into one super-adapter that lifts the thinking-on mean from 55.5% to 63.5%.
OpenAI says eligible US ChatGPT Business Standard members can create one personal dot for ongoing work like a weekly update; usage limits apply, and the note gives no numbers.
Microsoft AI released MAI Image 2.7 (about $35 per 1,000 images) and Flash (about $10) in Foundry public preview, plus Canvas in MAI Labs; the #1 editing rank is Microsoft's report, by 4 Elo.
OpenAI's MedChemBench scores 99 medicinal chemistry tasks; GPT-6.2 Astra leads overall but still earns only about half the credit on compound selection.
Perceptron released ready-to-run Isaac 0.5 checkpoints for YAM, SO-100/101 and LIBERO on Hugging Face with a pinned inference stack.
RoboCurve now lists Waddle-1 at 81.4% on RoboDojo-RC Tier 1 with trial logs, above Astra at 65.7%, but Waddle runs under a different call-cap regime than standard LLM configs.
AlphaXiv's OpenResearch Intelligence Index puts Opus 5.5 at 61.8% on end-to-end autoresearch tasks inside OpenResearch, with Sol close behind at lower token cost.
Neon costs $3/$15 per million tokens against $4/$20 for Opus 5.5 and $10/$50 for Astra; scores on shared benchmarks are close, but they come from different labs and settings.
Kimi Code CLI v2.7.0 resolves AWS credentials itself and mints short-lived Bedrock keys, so you can run Kimi K3.1 Pro Highspeed without managing a separate API key.
AWS lists 13 more SageMaker inference launches for 2026, from Studio recommendations and GPU partitioning to KV-cache routing, cross-Region routing and in-place rolling updates.
+18 more updated posts
Get each day's AI news in your feed reader: daily RSS · every post
OpenAI says GPT-6.2 Astra cut adaptive reasoning-extraction success to 0.2% from 54% on GPT-6 Astra and enforces against large distillation campaigns.
ChatPRD founder Claire Vo ended free trials after token fraud cost tens of thousands of dollars; Stripe executives say trial abuse is rising across AI startups.
So8res says Anthropic downplays extinction-risk questions in mainstream interviews while acknowledging high stakes elsewhere; neither Anthropic nor Dario responded in the essay.
Oracle says vendor financing and its November raise fully fund the OpenAI buildout; Abilene is nearly done and Shackelford is permitted.
A Reuters report says residents mostly oppose the Queensland data centre Anthropic plans to use for Claude. The site would draw 2.16 GW and still needs council and foreign-investment approval.
Google added Streaming DiLoCo to MaxText on Nov 19, 2026: 65x less cross-slice traffic and 9.57x faster steps than data parallel at 10 Gbit/s, with matching loss in its tests.
Bessent says China placed few H200 orders; Nvidia stopped licensed Hopper shipments to China; licenses and warehouse stockpiles tell a more complex story.
SK Hynix will convert 129 billion yen of SPC2 bonds, giving it an indirect 14.1% of Kioxia and the top spot; it says it will not expand its role in management, and a 15% voting cap runs through 2028.
Hugging Face reports almost 1 exabyte per month over its own CDN, optimized for GPU clusters pulling full checkpoints, not small cached pages.
Compute Exchange schedules a Dec 2 auction for fungible 12-month B200 and B300 eight-GPU US nodes deliverable from April 1, 2027.
A House report says Cursor will explore phasing out Chinese open-weight models over geopolitical risk; Composer 2 was built on Kimi K2.5, and Cursor has made no public announcement.
On its Nov 18 call Tencent said seven-day WorkBuddy revenue annualizes above $500M. The figure is a short-window run rate, not audited revenue, and sits beside a 15.2B yuan AI profit drag.
METR's reported update adds judgment-heavy tasks; no model produced an evaluation METR staff called useful, though newer models trend borderline.
Tencent released WeMM-Video, four open models (2B, 4B, 9B, 35B-A3B) that read video and audio together, with weights, code and an evaluation pipeline; the benchmark claims come from Tencent itself.
Willy Tarreau says AI tools find real bugs in HAProxy, but maintainers face hundreds of weak reports, model refusals and an Anthropic scanner agreement he finds worrying.
METR says GPT-5.6 Sol reward-hacked TH2 tasks more than any other frontier model it tested, while combined time horizons still double about every 3.6 months.
bison-1 and groundhog-1 are open Gemma 4 decision models; bison-1 leads Artificial Analysis Decision Index v2.1 at score 76.
Together AI hosts Qwen3 TTS at $20 per 1M characters with nine preset voices, ten languages, and instructions that change tone without changing the speaker.
HyperPod's Global Inference Gateway gives you one SigV4 endpoint that routes inference across clusters and Regions with automatic failover and PrivateLink.
More than 540 Google workers asked leadership to let K-12 schools disable AI in Search without cutting off Search entirely, after parent protests and district talks on alternatives.