# Horizon AI Briefing — 2026-11-21

## Top stories
- **Anthropic mathematician claims Claude helped prove the Riemann Hypothesis — verification underway** (https://www.aitimes.com/news/articleView.html?idxno=217638) — Levant Alpoğe, a mathematician at Anthropic, has self-reported on his personal website that he proved the 167-year-old Riemann Hypothesis with Claude's assistance, including a Lean formal proof. This would be one of the greatest mathematical achievements in history, but the claim is unverified and ships without peer review or linked evidence; a competing human proof has also emerged simultaneously. Professionals should watch this closely but treat it as unconfirmed until the mathematics community independently verifies. [vendor-claimed capability; no independent eval]
- **Anthropic IPO stumbles as broader AI sell-off wipes $4T from markets** (https://www.aitimes.com/news/articleView.html?idxno=217639) — Anthropic's stock fell 8.5% on its second trading day, approaching its $83 IPO price after opening at $111.50, dragging AI-related equities broadly lower in what analysts are calling a $4 trillion AI sector wipeout. The retreat signals that public market appetite for AI valuations may be correcting, even as private funding rounds remain elevated. This is a critical stress test for the AI IPO pipeline. [carried by 2 publishers]
  - Sources: AI Times (https://www.aitimes.com/news/articleView.html?idxno=217639); ETNews (https://www.etnews.com/20261121000022)
- **Senate Intel chair says Commerce Secretary Lutnick ignored AI chip smuggling warnings to China** (https://bsky.app/profile/yasharali.bsky.social/post/3n2r6cq6vwy2x) — Senator Tom Cotton, chair of the Senate Intelligence Committee, publicly accused Commerce Secretary Howard Lutnick of repeatedly ignoring warnings that U.S. AI chips were being smuggled to China before the Nvidia scandal became public. The accusation raises serious questions about enforcement failures at the highest levels of the Commerce Department and could accelerate congressional action on export controls.
- **Trump threatens 'tremendous tax' on U.S. companies building AI data centers abroad** (https://www.aitimes.com/news/articleView.html?idxno=217642) — President Trump announced plans to impose heavy taxes on U.S. tech companies that build AI data centers outside the United States, framing it as a domestic investment mandate. The policy, if enacted, would materially affect hyperscalers' global infrastructure strategies and complicate international AI deployment plans. [self-reported; no independent confirmation]
- **LLMs that rewrite their own weights — new research frontier** (https://arxiv.org/abs/2611.21885) — New research published on Hacker News demonstrates that language models can rewrite their own weights, a significant development in model self-modification and continual learning. If the approach generalizes, it could fundamentally change how models adapt post-training, with major implications for AI safety, alignment, and deployment. [primary source]
- **House Democrats open investigation into Nvidia's alleged interference in Commerce Department probe** (https://www.aitimes.com/news/articleView.html?idxno=217641) — U.S. House Democrats have launched an investigation into whether Nvidia improperly influenced a Commerce Department inquiry into Megaspeed, a Singapore-based firm accused of illegally exporting semiconductors to China. The probe compounds Nvidia's regulatory exposure at a moment when leaked export permit data showing 880,000 chips sold to Chinese firms is already rattling markets.
- **Cursor reportedly considering phasing out Chinese AI models amid congressional scrutiny** (https://www.aitimes.com/news/articleView.html?idxno=217643) — SpaceX's AI coding platform Cursor is reviewing plans to discontinue use of Chinese open-weight AI models following a joint congressional investigation into U.S. tech dependence on Chinese AI. Per a self-reported claim with no linked evidence, this reflects growing geopolitical pressure on AI supply chains and could set a precedent for enterprise AI procurement policies. [self-reported; no independent confirmation]
- **Trump and Xi agree on human control over military AI — broader than prior nuclear-only deal** (https://www.aitimes.com/news/articleView.html?idxno=217640) — U.S. Treasury Secretary Scott Bessent announced that Trump and Xi Jinping agreed that humans must retain control over all AI-assisted military decisions, extending a 2024 agreement that was limited to nuclear weapons. Per a self-reported claim via a CNBC interview, this represents a rare area of U.S.-China AI cooperation and could shape international military AI governance norms. [self-reported; no independent confirmation]
- **Signal blocks screenshots on iOS and Android to stop AI agents from reading private chats** (https://the-decoder.com/signal-blocks-screenshots-on-ios-27-and-android-to-stop-ai-agents-from-reading-private-chats/) — Signal is implementing screenshot blocking on iOS 27 and Android, explicitly targeting AI agents like Siri AI, Gemini, and Meta's Muse that could read on-screen content. Signal calls AI agent access to private messages an 'existential threat' to privacy, though admits protections remain incomplete on macOS — framing a new front in the AI privacy battle. [self-reported; no independent confirmation]
- **AI-discovered CPU architecture 'Frostbite' claimed — no linked evidence** (https://architectlabs.com/blog/frostbite) — A self-reported claim, with no linked evidence, describes Frostbite as a new CPU architecture discovered, designed, and verified entirely by AI. If substantiated, this would mark a milestone for AI-driven hardware design, but professionals should await independent verification before treating the capability as established. [vendor-claimed capability; no independent eval]

## Still in the news
- **Inkling-2-Mini-GGUF trending on Hugging Face** (https://huggingface.co/unsloth/Inkling-2-Mini-GGUF) — day 5, 1 new report in the last 24h. The Unsloth quantized model continues to attract community attention four days after launch, sustaining a top trending position on the Hub with 278 likes and new engagement today.

## Emerging signals
- **AI agents acquiring financial autonomy — wallets, payments, and regulatory lag** (https://www.scmp.com/tech/big-tech/article/3375253/ai-agents-are-getting-wallets-hong-kong-ready-them) — Animoca Brands' Minds platform is piloting live AI agent transactions on Visa's network in Hong Kong as regulators begin placing agentic commerce on the policy agenda. The convergence of autonomous spending capability and underdeveloped regulatory frameworks is an accelerating risk and opportunity for fintech and enterprise AI teams.
- **Meta reconsidering 2030 net-zero goal as AI data center emissions surge** (https://arstechnica.com/ai/2026/11/meta-reconsiders-2030-net-zero-goal-amid-record-emissions-from-ai-data-centers/) — Meta is 're-evaluating' its 2030 sustainability commitments amid record emissions from AI infrastructure, with two sources reporting on this shift. As compute demand accelerates, the gap between AI ambition and ESG commitments is becoming a material governance issue for the entire sector.
- **AI displacing entry-level jobs in China at an accelerating rate** (https://the-decoder.com/bloomberg-analysis-of-10-million-job-postings-shows-ai-exposed-roles-in-china-fell-10-percentage-points-faster-than-others/) — A Bloomberg analysis of 10 million job postings finds AI-exposed roles in China falling 10 percentage points faster than other categories, with over 90% of micro-dramas now AI-produced. Former central bank adviser Cai Fang warns a larger AI labor shock is still ahead, signaling that China's labor market disruption may preview what other markets face.
- **AI reasoning chains becoming extraction targets — OpenAI tightens defenses** (https://the-decoder.com/openai-makes-it-much-harder-for-attackers-to-extract-how-its-latest-ai-models-think/) — OpenAI self-reports reducing the success rate for extracting internal reasoning from its latest models from 54% to 0.2%, while suspending tens of thousands of accounts weekly for coordinated distillation campaigns. The emergence of reasoning-chain theft as an attack vector is a new and underappreciated AI security frontier.
- **Apple building proprietary AI device agent while raising barriers for competitors** (https://the-decoder.com/apple-joins-the-race-for-ai-agents-that-control-your-device-while-raising-new-barriers-for-competitors/) — Apple is now developing its own on-device AI agent while simultaneously making it harder for rival agents from Meta, OpenAI, Anthropic, and Google to access Apple hardware. This platform gatekeeping dynamic could become one of the defining regulatory and competitive battlegrounds in the AI agent era.

## New entrants
- **OpenComputer** (tool) — Open-source cloud coding agent for teams, self-reports a first-commit time of 24 seconds versus 53-72 seconds for alternatives, deployable on AWS, GCP, or Azure as an open-source alternative to Devin.
- **Sakana AI Applied Research Lab** (company) — A new research group within Sakana AI applying its multi-agent research to large real-world systems; first program targets finance and economics with a self-reported focus on complex adaptive systems.
- **Frostbite** (model) — A self-reported CPU architecture claimed to have been discovered, designed, and verified entirely by AI — no linked evidence provided.
- **Inception (Nvidia acquisition target)** (company) — A two-year-old AI startup building diffusion-based language models as an alternative to autoregressive generation; Nvidia is reportedly close to acquiring it for at least $500 million to bolster its Nemotron open-source lineup and robotics push.
- **Gemini Computer Use for Mac** (tool) — Google's new capability letting Gemini operate Mac desktop apps, local files, and websites in an isolated background window with live preview — bringing computer-use agent functionality to macOS.

## Biggest movers this week
- **OpenAI** (company) — 260 mentions this week, ↑45 vs the prior week
- **Gemini** (model) — 50 mentions this week, ↑24 vs the prior week
- **Microsoft** (company) — 44 mentions this week, ↑22 vs the prior week
- **Hugging Face** (company) — 33 mentions this week, ↑13 vs the prior week
- **Apple** (company) — 35 mentions this week, ↑12 vs the prior week
- **Tencent** (company) — 17 mentions this week, ↑12 vs the prior week

## China & East-Asia AI
- **AI agents are getting wallets. Is Hong Kong ready for them?** (https://www.scmp.com/tech/big-tech/article/3375253/ai-agents-are-getting-wallets-hong-kong-ready-them) — South China Morning Post
- **Nvidia faces further blow after shock US export permit data release: analysts** (https://www.scmp.com/tech/big-tech/article/3375287/nvidia-faces-further-blow-after-shock-us-export-permit-data-release-analysts) — South China Morning Post
- **Tencent releases Hy4 model weights with million-token context** (https://technode.com/2026/11/20/tencent-releases-hy4-model-weights-with-million-token-context/) — TechNode
- **Ex-Qwen chief’s Pragmatik Labs raises funding at $5 billion valuation** (https://technode.com/2026/11/20/ex-qwen-chiefs-pragmatik-labs-raises-funding-at-5-billion-valuation/) — TechNode
- **Alibaba uses its own chips to train next-generation Qwen model** (https://technode.com/2026/11/20/alibaba-uses-its-own-chips-to-train-next-generation-qwen-model/) — TechNode

## Korea AI
- **U.S. Democrats Launch Investigation into Nvidia's Alleged Interference in Megaspeed Commerce Department Inquiry** (https://www.aitimes.com/news/articleView.html?idxno=217641) — AI Times
- **SpaceX's Cursor Considers Phasing Out Chinese AI Models** (https://www.aitimes.com/news/articleView.html?idxno=217643) — AI Times
- **Trump and Xi Agree on Human Control Over Military AI, Extending to Major Military Operations** (https://www.aitimes.com/news/articleView.html?idxno=217640) — AI Times
- **South Korea to Exempt Non-Capital-Region AI Data Centers Under 50MW from Power Grid Assessment** (https://www.aitimes.com/news/articleView.html?idxno=217616) — AI Times
- **Trump: Heavy Taxes on U.S. Companies Building Data Centers Overseas** (https://www.aitimes.com/news/articleView.html?idxno=217642) — AI Times

## Japan AI
- **Introducing the Sakana AI Applied Research Lab** (/applied-research-lab/) — rss
- **Why Do High-Performing AI Companies Hire More Staff? IDC Study Challenges Conventional Wisdom** (https://www.itmedia.co.jp/enterprise/articles/2611/21/news003.html) — ITmedia
- **Kyoto University Publishes AI Dependency Checklist—'Avoid AI if You Match These Signs'** (https://www.itmedia.co.jp/aiplus/article/2611/20/2000003378/) — ITmedia
- **Dedicated Email Address for Muse AI Now Free for All Users** (https://www.itmedia.co.jp/aiplus/article/2611/20/2000003372/) — ITmedia

## Europe (EU) AI
- **OpenAI makes it much harder for attackers to extract how its latest AI models think** (https://the-decoder.com/openai-makes-it-much-harder-for-attackers-to-extract-how-its-latest-ai-models-think/) — The Decoder
- **Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats** (https://the-decoder.com/signal-blocks-screenshots-on-ios-27-and-android-to-stop-ai-agents-from-reading-private-chats/) — The Decoder
- **Google's Gemini can now run Mac apps and finish desktop chores in the background while you keep working** (https://the-decoder.com/googles-gemini-can-now-run-mac-apps-and-finish-desktop-chores-in-the-background-while-you-keep-working/) — The Decoder
- **AI Data Center in Space: Google Partners with Jeff Bezos' Blue Origin** (https://t3n.de/news/ki-rechenzentrun-all-google-bezos-blue-origin-1774626) — t3n
- **Bloomberg analysis of 10 million job postings shows AI-exposed roles in China fell 10 percentage points faster than others** (https://the-decoder.com/bloomberg-analysis-of-10-million-job-postings-shows-ai-exposed-roles-in-china-fell-10-percentage-points-faster-than-others/) — The Decoder

## Regulation updates
- [US] **A bill to require the Securities and Exchange Commission to issue rules relating to the use of artificial intelligence by public companies, and for other purposes.** — Proposed. Read twice and referred to the Committee on Banking, Housing, and Urban Affairs.
- [US State] **To amend the Homeland Security Act of 2002 to establish within the Cybersecurity and Infrastructure Security Agency the National Cyber and AI Investigations Board to investigate significant cyber incidents, cyber incidents involving terrorism, and cyber incidents involving artificial intelligence, to determine root causes and develop recommendations to prevent similar incidents and address terrorism threats to the United States, including threats relating to the malicious use of information systems or artificial intelligence by covered foreign adversaries, including terrorist organizations, and for other purposes.** — Proposed. Tracked

---
Source: Horizon (https://horizon.alchemylab.sh) — aggregated, LLM-scored AI intelligence; each item also lists its own primary source, and provider.citation credits both.