Skip to content

AX Says · Nov 22, 2026AI bear market

AI stocks fall 10% in a week as Nasdaq enters bear market

The biggest AI story today is not a model launch. It is market sentiment: AI stocks fell 10% in a week, dragging Nasdaq into its first bear market since the pandemic. Anthropic fell 9.1% on its second trading day. Two other stories are worth watching: two teams suddenly claimed proofs of the Riemann hypothesis, with verification still underway. Apple and Signal also both began tightening AI agents’ permission to read screens. First, the market.

AI stocks drag Nasdaq into a bear market as Anthropic falls 9% on day two

The Nasdaq 100 fell 5.6% on Friday and 10.4% for the week. It is now 20% below its October high, officially entering a technical bear market. This was its worst week since the pandemic. Beyond doubts about returns on AI infrastructure spending, sentiment took hits from an SEC investigation into Nvidia’s accounting, political headwinds after the Democratic midterm elections, and pressure to keep interest rates high as the Iran war pushed up oil prices.

Anthropic fell 9.1% on its second trading day, giving back most of its 17% first-day gain. Nvidia fell 14% in one day after Reuters reported SEC scrutiny of its sales disclosures; the US semiconductor index fell 10.7%. Reuters says an investigation is not itself evidence of wrongdoing, and many investigations end without enforcement action. But neither the SEC nor Nvidia commented, so the market sold first.

Regulation is not Nvidia’s only problem. It just reported a sharp rise in revenue, but tech executives including Elon Musk are pushing for former CEO Jensen Huang to return. Nvidia says there are no plans for his return. Strong results and a leadership dispute now sit alongside its decision to stop selling H200 chips to China. Nvidia’s story is getting crowded.

Anthropic’s IPO governance structure also deserves attention. Its seven founders hold, through Founder LLC, one share carrying 50.1% of the voting power. The Long-Term Benefit Trust can appoint and remove some directors but cannot directly intervene in day-to-day decisions. Shareholders have limited voting power, while compute spending and revenue growth still put pressure on safety investment.

It has been an ugly week. But a technical bear market is a statistical threshold. The real test is whether upcoming earnings show returns on AI spending.

Two Riemann hypothesis proofs appear in one day, but verification comes first

The math world erupted today: two claimed proofs of the Riemann hypothesis appeared within a day. Anthropic mathematician Levent Alpöge says he and Claude found proofs of the Riemann hypothesis and its generalized form. Hours later, Stanford made a draft written entirely by humans public. Alpöge wrote his argument in Lean code, split it into 14 data packages, and says it would exceed 250,000 pages in print. Another mathematician, Alessandro Masullo, also made an independent proof public at almost the same time.

The excitement is real, but neither claim has passed verification. Alpöge’s Lean code is still compiling, and machine verification could fail because of an error encoding the initial problem. Masullo’s approach is less transparent: he has not disclosed which Claude model he used or how much human guidance it received. I would discount both claims for now. The Riemann hypothesis is one of math’s toughest problems. People have claimed proofs every year for the past several years, and almost all those claims have fallen away. This time, though, Lean formal verification adds a check that a paper alone lacks.

Let verification run before we pop the champagne.

OpenAI gives GPT-6.2 templates as Claude Code moves tests before submission

Today, two major coding agents are setting rules for developers. OpenAI released prompt templates for GPT-6.2 Astra and Sol. They package writing, proactive execution and coding requirements into reusable modules. Answers should stand on their own. Within authorized bounds, agents should keep working from reversible assumptions and reuse existing code. They should add tests only when the user requests them, the repository requires them, or a specific regression needs to be caught. In plain terms: do not stop to ask at every uncertainty, and do not add tests just to create noise.

Meanwhile, the Claude Code team discussed the code-review bottleneck alongside a Harvard working paper. After enterprises adopted coding agents for 12 months, engineers wrote code 73.7% more efficiently per hour, but work items ultimately shipped rose only 13.7%. Review was the bottleneck. Claude Code’s approach moves end-to-end, property and differential tests before submission. It also reads REVIEW.md from the main branch, ties review rules to specific commits, and invokes reliability checklists when their conditions are met. Preventing changes under review from rewriting their own grading criteria is crucial.

Both are tackling the same problem: agents can do the work, but they are hard to control. As model capabilities advance, process controls must keep up.

Apple and Signal tighten AI screen access as permissions get more granular

System-level AI assistants are starting to read screens, and Apple and Signal both moved quickly today. The second beta of macOS 27.3 separates cross-app control permissions from Accessibility. It supports authorization by folder, by app and for each access. It also adds permission-risk warnings and a blue monitoring icon in the menu bar. The public release is expected in December. There is a catch: permissions granted before the upgrade will not be revoked automatically. Users must clear them manually.

Signal went further, launching screenshot restrictions on Android and iOS. President Meredith Whittaker says that when agents scan screens, they may capture and store sensitive data. That data may then go to a local database or remote server, beyond Signal’s control.

Trust in images and video is also being reconsidered. Generative AI makes real and fake video hard to tell apart. Some argue for skepticism by default, then checking provenance with C2PA signatures, Google SynthID watermarks and sensor signatures on some newer iPhones. Live video calls could also be verified with frame-by-frame signatures.

The more granular permissions become, the more their defaults matter. If users do not change them, open permissions stay open.

Together AI launches Qwen4 Max and reports a prompt leak the same day

Cloud inference platform Together AI had two stories today: a launch and an apology. The launch is Alibaba’s flagship Qwen4 Max, available on its Serverless inference platform with a 1M token context window and 30% off as of November 22, 2026. The same day, Together AI reported a security incident. An AI coding agent committed private Hugging Face credentials to the public GitHub repository xorl. An unauthorized party then downloaded stored prompt samples submitted to services including MiniMax M3, GLM-5.2 and Kimi K2.6 in June and July 2026, and used them to search for other service credentials.

The irony is that the leak began with an AI coding agent itself. Open-source models also had news today: Thinking Machines released Inkling 2 Mini with open MoE weights. It repeatedly runs some Transformer layers to reduce token consumption. The community also quantized it to 5.38 GB to run on Apple Silicon.

The faster model capabilities grow, the greater the risk of an agent accidentally leaking credentials. We need to be more careful about giving agents access to private repositories.

UN faces an autonomous weapons decision after 13 years of deadlock

International rules for autonomous weapons are stuck again. Countries failed to agree on a UN negotiating mandate for a legally binding treaty to regulate autonomous weapons. With Russia opposed, they instead accepted a weaker option that may carry no legal obligations. This is an old problem. More than 100 countries want new law, but the Convention on Certain Conventional Weapons requires unanimous agreement from its 128 parties. Every military AI power has a veto. The US and several countries developing AI weapons are unwilling to accept new legal constraints. Seven countries, including Austria, Belgium and Brazil, are still pushing to reach a binding treaty within a year. Critics are already impatient: after 13 years of discussion, there is still no binding framework.

Unanimity lets the slowest country set the global pace. Autonomous weapons are advancing far faster than treaty talks.

Small signals today

  • ClawDOS brings an AI coding agent to an IBM PC XT with 4.77 MHz and 640K RAM. Through the Claude Messages API, it reads and writes files, compiles code and runs DOS programs. That is some serious commitment to retro computing.
  • Pandex offers European legal search for AI agents. It indexes 2.8M official documents and links 37M citations. Its records are complete for 33 of 41 institutions, and it covers legislation and supreme court rulings from the EU and seven countries.
  • In a roundup of messaging-based AI agents, Instinct, Fambot and others take tasks by text or iMessage. They handle calendars, email and bookings, with varying prices and availability.
  • Shortical signed a likeness license directly with actor Aki Avni to make fully AI-generated short dramas using his digital double. He receives industry-standard compensation as a partner, but its public likeness-licensing portal has not yet produced a deal.
  • Slopstrap 4.2 is out. Its CSS and JS components make websites look like they were hastily generated by a large language model. Gradient text, glowing cards and animated borders are all ready for you.

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks

Product launches & updates

  1. Apple splits AI agent permissions in macOS 27.3 beta, adds per-app authorizationComputing Life · Share · Yage