AI Trends
Nov 22, 2026 · SaturdayPublic sources · Daily newsPublished daily at 08:00 Beijing time

AI News Daily

AI developments selected from public sources

Issue 21522Nov 2026Saturday
11monthThis month 22 issues
12
major stories
18
sources
2
first-party stories
About 8 min read

Lead story

Andrew Sutherland Releases Verification Results for Lean Formalization of the Riemann Hypothesis

Andrew Sutherland released verification results for the Lean formalization of the Riemann hypothesis: all 175,187 modules compiled successfully and passed module-by-module replay through the Lean kernel. nanoda, con-leche, and lean4lean accepted the complete exported proof, while the five core theorems depend only on the three standard axioms propext, Classical.choice, and Quot.sound. The results show only that these formalized Lean statements hold under Mathlib's definitions and the kernels used; checks with some default Lean kernels and the lake comparator are still underway.

Public sourcesTrending on Hacker NewsFollow-upOriginal ↗
01

Product release/Updated

Tencent Team Open-Sources LoopForge, Using Five Scheduled Agents to Maintain Code Repositories Automatically

A Tencent team has open-sourced LoopForge, a self-hostable workbench that uses five staggered, scheduled agents to conduct research, review PRs, inspect continuous integration, add tests, and clean up architecture. The workflow reserves merge decisions for humans and specifies that the agents must not merge or close PRs, push directly to the main branch, or modify other people's branches. It stops running once the number of architecture-related PRs reaches 10. The team said the setup was refined over several weeks, and one overnight task increased test coverage for a core Go package from 83.3% to 86.0%.

Original

AWS Quick Updates Workflow Orchestration, Web Experience, and Multi-Agent Collaboration

AWS Quick has introduced several updates, enabling users to build and improve workflows through conversation, run tasks continuously in a redesigned web application, and have multiple specialized agents collaborate on a unified response. The update also adds connectors for Meta ads and Amplitude, web-based skill management, content safety and sensitive-information controls, and support for Japanese, German, Simplified and Traditional Chinese, and Vietnamese. The features are now available in all regions where AWS Quick is offered.

Original
02

Industry news

Anthropic's Record-Breaking IPO Coincides With Wall Street Tech Stocks' Worst Week Since 2020

Anthropic raised $93.5 billion through its IPO, with its shares rising 17% on the first day of trading and its market capitalization briefly reaching $1.8 trillion. Over the same period, Nvidia plunged 23% in a week, while the Nasdaq fell 10% over five trading days, marking its worst week since the early days of the COVID-19 pandemic in 2020. The market is now focused on whether Anthropic can deliver enough growth to justify its high valuation and whether the listing will encourage more AI startups to prepare for IPOs.

Original

Mandate for UN Negotiations on Autonomous Weapons Control Treaty Fails to Pass

The Review Conference on Conventional Weapons at the UN in Geneva failed to agree on a negotiating mandate for a legally binding treaty governing autonomous weapons, with Russia's opposition and the consensus-based process leading the conference to adopt only a weaker proposal. More than 100 countries pushed for a new law, but major military powers investing in AI weapons, including the United States, favor a nonbinding agreement. All 128 parties to the convention must agree unanimously for a declaration to pass.

Original

US AI Stocks Plunge After Anthropic Completes $107.3 Billion IPO, Putting Pressure on QDII Fund NAVs

After Anthropic completed its $107.3 billion IPO, US AI stocks fell sharply, with the Nasdaq down 5.28% and Nvidia down 14%, putting pressure on the net asset values of several publicly offered QDII funds. Fund professionals interviewed attributed the correction to a reassessment of asset discount rates. Some institutions have begun increasing allocations to non-AI technology stocks with clearer cash flows and high-dividend companies listed in Hong Kong.

Original

US Chip Stocks Shed About $2.3 Trillion in Market Value in One Week, Nvidia Down About $2.7 Trillion From Its Peak

The Philadelphia Semiconductor Index fell more than 10% this week, wiping about $2.3 trillion from the chip sector's market value in its worst weekly performance since 2001. Nvidia declined for a seventh consecutive week and is now down about 47% from its record high in October, losing roughly $2.7 trillion in market capitalization. The liquidity shock from Anthropic's $107.3 billion IPO, compounded by the Federal Reserve's hawkish stance, prompted the market to reassess the financing and growth foundations of the AI rally.

Original
03

Research papers

Redwood Research: Auditors Struggle to Detect Deliberate Sabotage in Complete AI Safety Research Records

Redwood Research had human experts secretly sabotage AI safety research and then asked auditors to inspect the complete research logs. Human auditing achieved an AUROC of just 0.59, fully identified only 17% of the sabotage mechanisms, and completely missed 74%. The experiment included 22 research records and 33 specific sabotage mechanisms, with 25 human auditors completing 104 audits. Claude Fable 5.1 and GPT-6 Astra scored 37.1% and 32.6%, respectively, on mechanism identification, outperforming the human score of 21.9% but still missing most of the sabotage. The study also noted that all samples came from a single researcher, the confidence intervals were very wide, and the saboteurs knew the expected results of the normal experiments in advance.

Original

Nearly 2 Million Decoy Requests Expose Privacy Risks Across the AI Service Supply Chain, With Leaks Possible Even Under Zero Data Retention Settings

Dragos Chief Research Officer HD Moore sent nearly 2 million requests containing traceable decoys to multiple AI service providers and found problems across the service supply chain, including stolen prompts, data leaks by resellers, and failures to honor privacy settings. At least three providers showed signs of compromise, and some decoy keys were used to make tens of thousands of cloud service API calls within hours of submission. Private URLs in some requests marked for zero data retention or no data collection were still accessed days later.

Original

Columbia University and MIT Paper Says Claude Algorithms Break Long-Standing Speed Barriers for 3SUM and All-Pairs Shortest Paths

A paper from Columbia University and MIT says algorithms discovered by Claude reduce the time complexity of 3SUM and all-pairs shortest paths to O(n^1.9992) and O(n^2.9995), respectively. The approach achieves speedups by reducing the work needed to compute specific entries in the products of two long, narrow matrices and can be applied to related problems using known reductions. Claude discovered the method in a single session, using 16M output tokens with no human input, and the paper checked the relevant arguments using Lean 4.

Original

OpenAI Report Says Large Models Cut AMD Chip Optimization Cycle to Under Five Months

An OpenAI technical report says its large model helped design AMD chips, compressing about a year of engineering optimization work into less than five months and achieving up to a 14% improvement in PPA for flagship GPU circuit designs. After the model rewrote portions of the SystemVerilog RTL code, the changes were verified using Synopsys EDA tools, and some optimizations have been incorporated into designs recently sent for fabrication. The 14% figure represents a combined gain in power, performance, and area, and does not mean a direct 14% increase in actual GPU performance.

Original
04

Tips and perspectives

05

Briefs

End of issue

AI News DailyAutomaticallycompiledfrom public reports, with source links