AI Trends
Nov 21, 2026 · FridayPublic sources · Daily newsPublished daily at 08:00 Beijing time

AI News Daily

AI developments selected from public sources

Issue 21421Nov 2026Friday
11monthThis month 21 issues
12
major stories
28
sources
4
first-party stories
About 7 min read

Lead story

Thinking Machines Lab Releases Recurrent Transformer Model Inkling 2 Mini

Thinking Machines Lab has released Inkling 2 Mini, an open-weight recurrent Transformer model whose middle layers run three times for each token and which can run locally on a laptop. The MoE model has 26.4B total parameters, 3.4B activated parameters, 5.4B activated weights per token, and a memory footprint of about 20GB; it scored 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2. The model is available on Tinker for inference and fine-tuning, while the full weights and a variable-loop version are now available on Hugging Face.

First-party sourcesThinking Machines Lab: Official BlogAlso 4 other sources reported thisOriginal ↗
01

Model releases / updates

Tencent Hunyuan Releases Hy4 Model Weights and System Card

Tencent Hunyuan has released the weights for Hy4, its native multimodal flagship MoE model, which has 770B total parameters, 49B activated parameters, and a 1M context window. The company also released the Hy4 System Card, including a comprehensive safety evaluation report, along with links to GitHub, Weights, GGUF, and the License.

Original
02

Product release/Updated

ChatGPT Launches Space Collaborative Creation Workspaces for Plus Users

ChatGPT has launched Space for Plus users, enabling multiple people to collaborate on interactive pages, slide decks, and websites while centrally managing images, files, and uploaded content. Users can write, conduct research, generate images, and create interactive charts, while connected tools can automatically update to-do lists, project trackers, and travel plans. Space now supports mobile editing and offers improved responsiveness and reliability.

Original
03

Industry news

Anthropic Completes Upsized $93.5B IPO and $26B Private Placement

On November 20, 2026, Anthropic completed an IPO of 1.127 billion Class A shares at $83 per share, raising approximately $93.5B in gross proceeds. At the same price, the company also privately placed 313,253,007 non-voting Class N shares with AMD, Broadcom, Google, Micron Technology, Microsoft, NVIDIA, and TSMC Partners, in a transaction worth approximately $26.0B. Together, the two transactions generated approximately $119.5B in gross proceeds. Its Class A shares closed at $88.91 on November 20.

Original

Anthropic's Nasdaq Debut Fails to Lift AI Stocks as Asian Markets Plunge

Anthropic's Nasdaq listing failed to lift the AI sector. On November 20, the MSCI Asia Pacific ex Japan Index fell 3.3%, extending its weekly decline to 8.5%. Markets in Japan, South Korea, and Taiwan dropped 5%, 5.2%, and 6.2%, respectively, that day. Anthropic's Class A shares opened at $111.50 and closed at $97.17. Reuters linked the sell-off to pressures including the unwinding of crowded trades, insufficient liquidity, a backlash against data centers, and concerns about industry debt.

Original

Anthropic IPO Proceeds Rise to $119.5 Billion as Shares Fall 8.5% on Second Trading Day

Anthropic shares closed their second trading day at $88.91 per share, down 8.5% from their first-day closing price, giving the company a market capitalization of approximately $1.6 trillion. Following the exercise of the overallotment option, the company and co-founders participating in the offering raised a total of $119.5 billion, comprising $93.5 billion from the IPO and $26.0 billion from the concurrent private placement.

Original

Anthropic Completes Largest IPO in History, Raising $81.3B

Anthropic listed on Nasdaq under the ticker ANT, raising $81.3B in what the article calls the largest IPO on record. The stock closed its first day at $97.17, up 17% from its offering price, valuing the company at more than $1T. CEO Dario Amodei said the company would control the pace of its development based on the safety of its models and would not yield to commercial or political pressure.

Original

Nvidia Stock Plunges 12%, Marking Its Longest Weekly Losing Streak Since the 2008 Financial Crisis

Nvidia shares fell 12% on Friday, declining for a seventh consecutive day and heading for a seventh straight weekly loss, their longest weekly losing streak since 2008. During that period, the stock fell a cumulative 44%, wiping $2.5 trillion off its market capitalization. Investors are concerned about leadership following co-founder Jensen Huang’s departure, the sustainability of AI infrastructure development, and chip competition from Google and AMD.

Original

SpaceX Sells $32B in Debt to Purchase Nvidia Rubin GPUs

SpaceX is selling $32B in debt to purchase about $47.3B worth of Nvidia Rubin GPU equipment, with chip deliveries scheduled to begin in January. Saturn Funding Trust, managed by a unit of Apollo Global Management, will purchase the chips and lease them to SpaceX for about five years. This structure keeps the related borrowing off SpaceX’s balance sheet.

Original

Hundreds of Google Employees Demand Withdrawal of Policy Restricting Schools From Disabling Gemini

Hundreds of Google employees have signed a petition calling on the company to withdraw a policy that would bar schools that disable the Gemini AI tool from using Google Search starting next year. School districts in New York, Los Angeles, Chicago, and elsewhere are discussing alternative search engines. Google management has confirmed receiving the petition but has not yet responded to its specific demands.

Original
04

Research papers

OpenAI Details Defenses Against Adversarial Knowledge Distillation and Risks of Reasoning Extraction

OpenAI has unveiled defenses against adversarial knowledge distillation that use adversarial training, classifiers, API safeguards, and coordinated-activity investigations to limit the extraction of model capabilities and protected reasoning. In adaptive black-box evaluations, the average attack success rate for extracting reasoning from earlier requests fell from 54.0% with GPT‑6 Astra to 0.2% with GPT‑6.2 Astra, while the success rate for extracting reasoning from the current request dropped from 22.9% to 0.9%. OpenAI also reported discovering multiple unauthorized distillation operations this year, taking action against thousands to tens of thousands of accounts each week.

Original
05

Tips and perspectives

Did Claude Really Prove the Riemann Hypothesis? Terence Tao and Other Mathematicians Respond

Anthropic mathematician Levent Alpöge claims that he and Claude collaborated to find a method for proving the Riemann Hypothesis and the Generalized Riemann Hypothesis, and released about 17 million lines of Lean formal proof code. Mathematicians including Terence Tao are following the development, but the mathematical community is still verifying the code and proof approach. Under the Clay Mathematics Institute’s rules, the result must also be formally published, withstand at least two years of scrutiny, and gain broad acceptance.

Original
06

Briefs

End of issue

AI News DailyAutomaticallycompiledfrom public reports, with source links