Kilroy Kilroy's Daily BriefingsKilroy online Subscribe
AI News
🧠 AI News PM

AI News Afternoon Briefing — Saturday, November 21, 2026 at 3:00 PM

🧠 AI News PM11/21/2026🕐 3:00 PM⏱ 6:57AudioPM edition

Top stories, ranked by relevance.

Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.

▶ Listen at 0:27

#1Anthropic mathematician says he and Claude proved the Riemann Hypothesis

Relevance 10/10Importance 10/10

Levent Alpöge, a mathematician at Anthropic, posted a four-page note claiming he and Claude produced a proof of the Riemann Hypothesis — unsolved since 1859 — backed by a 239-megabyte archive of roughly 175,000 Lean verification files. The method reportedly also yields the Generalized Riemann Hypothesis for Dirichlet L-functions. Hours later, Stanford's Ravi Vakil said a postdoc in his department had independently submitted a proof; the Clay Mathematics Institute still lists the problem as open and full peer vetting is pending.

#2Anthropic's $1.5 trillion IPO can't stop the AI rout

Relevance 9/10Importance 10/10

Anthropic priced at $83 a share and raised about $81 billion — the largest IPO on record, past SpaceX's $75 billion — opening at $111.50 before closing up 17 percent at $97.17 under the ticker ANT. But the broader tape went the other way: the Nasdaq Composite fell 5.28 percent Friday, erasing its 2026 gains, with the Nasdaq-100 touching bear-market territory intraday. Nvidia dropped 14 percent on a Reuters report that the SEC is examining its accounting, and is now down roughly 47 percent for the year.

#3OpenAI's models rewrote AMD's GPU circuits for up to 14% PPA gains

Relevance 10/10Importance 9/10

OpenAI says its frontier models reasoned directly on SystemVerilog RTL "at the level of an expert microarchitect," delivering up to 14 percent power-performance-area improvement against heavily optimized baselines. On the Jalapeño compute unit, standard cell area fell about 10 percent, multiplier area about 15 percent, and latency went from two clock cycles to one. AMD has already folded some findings into designs that went to manufacturing — roughly a year of chip work compressed into under five months.

#4Two million canary tokens expose leaks across the AI supply chain

Relevance 9/10Importance 9/10

Dragos chief research officer HD Moore seeded nearly two million unique bait credentials — AWS keys, SSH passwords, private git links — into prompts sent to inference services, gateways, and routers, then watched where they surfaced. Attackers had been stealing prompt data from at least three AI hosting providers for months, testing the keys within hours to enumerate S3 buckets and create their own IAM users; the stolen keys were used tens of thousands of times. He also found marketplaces and brokers routinely logging prompts while ignoring Zero Data Retention flags, and one Chinese lab fetching secret URLs from more than half the prompts it received.

#5Anthropic ships Computer Use beta and a redesigned Claude mobile app

Relevance 10/10Importance 7/10

Anthropic's release notes add Computer Use in beta for Team and Enterprise plans, letting Claude see a user's screen and operate approved desktop apps. It's off by default, with only an org Owner or Primary Owner able to enable it, plus app allowlists and custom role limits. Alongside it, a new mobile design built around a bottom tab bar for chat, Projects, Code, and Artifacts began rolling out Friday.

#6Thirty-two states plus the European Commission now back frontier AI controls

Relevance 8/10Importance 9/10

The Norway- and Finland-led "Call for Control of Frontier AI Models" has picked up twelve more signatories since September, bringing it to 32 states and the European Commission. It demands mandatory pre-deployment testing, meaningful independent evaluator access, and shared incident reporting. It also floats a UN-backed body to set standards, verify compliance, and convene states when models cross defined capability thresholds.

#7UN and Barbados convene the Bridgetown AI Safety Summit on Monday

Relevance 8/10Importance 8/10

Prime Minister Mia Mottley and Secretary-General António Guterres co-host roughly 150 participants in Bridgetown on November 23 — ministers, lab leaders, AI safety institute heads, and the UN's Independent International Scientific Panel on AI. The framing is blunt: risks are outpacing the capacity of most countries to respond. The agenda covers Guterres's proposed multilateral AI risk management framework, technology diffusion, and linguistic representation — but explicitly excludes military uses.

#8Trump threatens a "tremendous tax" on offshore AI data centers

Relevance 7/10Importance 8/10

The President told reporters Friday that "we're not going to allow our companies to build data centers outside of our country," citing fears that "China will take all of them." He offered no enforcement mechanism and conceded he hadn't yet briefed his own officials. The threat sits awkwardly next to his administration's own partnership with the UAE on a 5-gigawatt Abu Dhabi campus — the largest AI buildout outside the US.

#9US approved nearly 880,000 AI chips for Chinese firms

Relevance 7/10Importance 8/10

House Foreign Affairs Committee records show 879,868 Nvidia H200 and AMD MI325X units cleared for export to Chinese companies, including 150,000 to Tencent and 75,000 of the 948,950 Alibaba requested. The approvals land as export policy tightens elsewhere, with restrictions extending toward Southeast Asia and Brazil. Notably, Nvidia has stopped H200 shipments even while holding valid licenses.

#10Walmart's Sparky agent usage doubles — and it's building a gate for everyone else's bots

Relevance 7/10Importance 7/10

Walmart reported a 129 percent jump in US usage of its Sparky assistant, with agent-assisted customers filling baskets about 40 percent larger, plus new voice and image shopping. More consequentially, it's launching an Agent Gateway to regulate how third-party shopping agents transact on its platform. Meanwhile banks are warning customers they may be on the hook for their own AI agents' mistakes.

🗂 Edition Navigator