2026-11-22 · ~8 min read · Past edition

10 picked from 42 candidates · ordered by significance

Today's Insight

This week AI ran ahead of its checks: in Anthropic's stock, a math proof and security patching

Anthropic's stock fell 8.5% on its second trading day, November 20, closing at $88.91 after dipping to within five cents of its $83 IPO price, AI Times reported. AI stocks fell together that day, with Nvidia down about 14%. On The Verge's podcast, its senior AI reporter Hayden Field said what Anthropic sells to large customers is its safety brand, and the listing puts that brand against pressure to make a profit.

Continue reading (2) Show less

Anthropic mathematician Levent Alpöge announced that he and Claude had proved the Riemann hypothesis, but mathematicians have not yet confirmed the proof is right. Hours later Stanford's Alessandro Masullo rushed out an all-human competing draft. In both the stock and the proof, a big claim arrived before the checks did.

Wired reports that a flood of AI-found is forcing companies to push patches without time to validate them, and researchers say that pressure is starting to play a role in outages. The number found this year passed 100,000, about four times the 2022 total. The encrypted messenger Signal blocked screenshots so AI agents cannot read the screen, and the UN gave countries two years just to propose rules on autonomous weapons.

Signal to watch Over the next few days, watch whether Anthropic's stock falls below its $83 IPO price and whether reviewers report that the proof states the Riemann hypothesis correctly.

Can Anthropic afford to put safety first?

Summary

The Verge's senior AI reporter Hayden Field told host Nilay Patel that Anthropic's IPO sets its safety-first company structure directly against the pressure of shareholders. Anthropic listed on the Nasdaq at $83 a share and climbed as high as $111 during the day. It also carries roughly $900 billion in compute bills over the coming years, so it has to keep growing revenue to pay them. Field said she heard that OpenAI, the maker of ChatGPT, once badly wanted to list before Anthropic but stepped back after a year of setbacks, and can now ride on the precedent Anthropic set.

Why it matters

Why It Matters

What Anthropic sells to large customers is, in the end, its safety brand, and the listing puts that brand against the demand to turn a profit. The real test is whether the safety promise holds in a quarter when revenue slows.

Story so far Anthropic's IPO and what its filing revealed

Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft

Summary

Within hours of Anthropic announcing a Claude proof of the Riemann hypothesis, Stanford mathematician Alessandro Masullo rushed out an all-human draft proof of the generalized Riemann hypothesis. Anthropic mathematician Levent Alpöge said on November 20 that he and Claude had proved the Riemann hypothesis and its extension, the generalized Riemann hypothesis. The proof is written in , a programming language that lets a computer check a math proof line by line, comes in 14 data packages, and would run past 250,000 pages if printed. Masullo had planned to polish his draft for a year before releasing it, and Ravi Vakil, a Stanford professor and president of the American Mathematical Society, announced it. According to AI Times, the Lean code runs to about 17 million lines.

Why it matters

Why It Matters

Neither proof has been accepted by mathematicians. Lean code that runs without errors shows the code is well formed, not that it states the Riemann hypothesis correctly. So what settles it is whether the statement was translated into code correctly, not how big the proof is.

Anthropic falls 8.5% on its second day of trading, nearing its IPO price

앤트로픽, 상장 이틀째 8.5% 급락...공모가까지 근접

Summary

Anthropic's stock fell 8.5% to close at $88.91 on its second trading day, November 20, and at one point dropped to $83.05, just five cents above the $83 IPO price. On the first day it opened as high as $111.50, which briefly put the company's value above $2 trillion, and it closed at $97.17, 17.1% over the IPO price. Co-lead underwriter Morgan Stanley used all 147 million shares of the , which lifted the total raised to $93.5 billion, but it also means no buying power is left to prop the price up if it keeps falling. AI stocks dropped together that day, with Nvidia down about 14% and AMD down about 13%.

Why it matters

Why It Matters

Retail investors traded about $772 million on the first day, the most for any IPO first day since records began in 2012, yet their actual net buying was only about $60 million. The excitement was bigger than the buying. Attention is shifting to OpenAI, which has hinted at a 2027 listing, and Anthropic's trading from here is the nearest guide to when OpenAI lists and at what price.

Story so far Anthropic's IPO and what its filing revealed

Nvidia researchers argue huge AI context windows are a dead end

Summary

Nvidia researchers argue that endlessly enlarging the , the amount an AI model reads at once, is usually overkill and expensive. In several tests, summarizing periodically and looking up only the relevant parts performed about as well as feeding in the full text. The Decoder noted that GPT-4, released in March 2023, handled 8,192 tokens, while current models handle about one million, roughly 120 times more and enough for all of War and Peace.

Why it matters

Why It Matters

If summaries and targeted lookup can match pasting in a whole long document, users can save on token costs. The tasks and models behind that result still need checking.

OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions

Summary

OpenAI published prompt templates and API changes for GPT-6.2 Astra and Sol. Astra is meant for the hardest tasks, and Sol handles complex work at lower cost. Astra answers more briefly and directly than earlier models, so instructions like "be concise" need a second look, and its habit of suggesting a next step and stopping to ask is to be fixed with a prompt telling it to make a reversible assumption, flag it, and keep going. In the API, temperature and top_p are no longer supported, and editing the conversation history or switching models now discards earlier reasoning. Up to 272,000 tokens, Astra 6.2 costs $8 per million input tokens and $40 per million output tokens.

Why it matters

Why It Matters

Carrying old prompts over to the new model can make answers too short or make it stop after suggesting a next step. Teams with a pile of existing prompts should strip out the instructions written for the previous model first.

What to do now

If you call GPT-6.2 through the API, remove temperature and top_p from your requests and move anything set to none or minimal reasoning effort to low.

The Outages Are Just Getting Started

Summary

Last week's outages hit Outlook, Apple, a half dozen US banks, United Airlines and even the NSA's classified network, but there is no evidence AI attacks caused them, and each had a routine explanation, Wired reports. Microsoft's 365 outage came from a flawed infrastructure update made worse by other network failures, and officials said the NSA outage came from botched patching after October's critical PHP updates. Still, researchers say AI-found are pouring in and attackers exploit them faster, which forces companies to push patches without time to validate them. This week the number of vulnerabilities found in 2026 passed 100,000, about four times the roughly 25,000 recorded in 2022, the last year before LLMs went mainstream.

Why it matters

Why It Matters

Wired found no AI attack behind these outages; the pressure shows up instead in how fast patches have to ship. Not every new bug matters, though: VulnCheck counts only about 1,500 exploited vulnerabilities this year, 1.5% of the total, so the number of bugs is not the same as the size of the risk. Matthew Prince, CEO of Cloudflare, predicts it will feel like the sky is falling for two years.

25000건 2022 100000건 2026 +300%
This year's vulnerability count is four times that of 2022, the last year before LLMs went mainstream.

Signal restricts mobile screenshots as it confronts ‘existential threat’ of AI surveillance

Summary

Signal, the messenger known for encrypted chats, announced on Friday in a blog post that it is restricting screenshots in its Android and iOS apps. The Windows app already blocked screenshots by default. Signal president Meredith Whittaker called operating systems with built-in AI agents that can scan the screen an existential threat to Signal's ability to protect privacy on its own. Users can turn the restriction off in settings, and Whittaker admitted the step is incomplete because it cannot block the other ways agents can reach app data.

Why it matters

Why It Matters

A messenger can only fight OS-level AI agents from inside its own app. An agent scanning any one participant's screen exposes everyone in the chat, and Whittaker herself called the screenshot limit incomplete.

What to do now

If you use Signal, update the app and check in settings that the screenshot restriction is turned on.

A tale of two Nvidias

Summary

TechCrunch's Equity podcast framed Nvidia's week as two stories: soaring revenue, yet a falling stock. Interim CEO Colette Kress announced that Nvidia will stop selling H200 chips in China, and the Trump administration is tightening rules to also close the remote-access route that let Chinese firms use chips abroad. Nvidia's critics estimate China is about 20% of its business once indirect sales through Malaysia and Southeast Asia are counted. Meanwhile Elon Musk and others are circulating petitions for the already departed Jensen Huang to return, and Nvidia has denied that is possible.

Why it matters

Why It Matters

If much of Nvidia's China business really runs through indirect routes that do not show up in its financial statements, closing the remote-access route reaches business the reported numbers do not show. The board is trying to look in control while the investigation into who knew what about chips allegedly sold to China is still unresolved.

Story so far Jensen Huang's exit and the question of Nvidia chips in China

The UN fails to agree on rules regulating ‘killer robots’

Summary

UN diplomats in Geneva spent a largely fruitless week on rules for autonomous weapons, or killer robots, and agreed only to start discussing options. The International Committee of the Red Cross urged binding rules within a year, backed by the UN secretary-general and Pope Leo, but the compromise gives countries two years to propose regulations. According to NPR as cited by The Verge, Russia delayed proceedings on day one by objecting to the agenda, and the US proposed a five-year window to discuss nonbinding, optional guidelines. Some of the eventual rules may stay optional.

Why it matters

Why It Matters

The Red Cross asked for binding rules within a year; countries gave themselves two years just to propose them, and some may stay optional. Semi-autonomous weapons and drones are already responsible for civilian deaths in Ukraine, Gaza and Sudan.

Google adds GIF generation to the Gemini app, following OpenAI

구글, 제미나이 앱에 GIF 생성 기능 출시…오픈AI 이어 도입

Summary

On November 20, Google said on its official X account that the Gemini app can now turn a typed scene or action into a looping GIF of about three seconds. It is made by the video model Gemini Omni 1.2 Flash, and the result can be copied elsewhere or downloaded as a .gif file. It is rolling out in eight countries, including the US, Canada, the UK, Japan, South Korea, Australia, Germany and France, to free users as well as paid subscribers. According to AI Times, OpenAI added GIF generation to ChatGPT first, on November 4 with a new image model, and Google followed about 16 days later.

Why it matters

Why It Matters

South Korea is among the first countries, so users there can start making chat GIFs right away. A developer API for GIF generation has not been released, so building it into a product has to wait.

What to do now

Type a scene or action in one sentence in the Gemini app to get a GIF. If you do not see it yet, it is rolling out in stages, so check again in a few days.

Terms in this briefing

Past Briefings

Monthly Reports

Weekly Recaps