DISCOVER THE FUTURE
AI is getting weird. Keep up.
Page 4
On Nov. 11, Tmall made Token, the unit AI usage is measured in, a standalone Double 11 category for the first time. MiniMax, Alibaba Cloud, Zhipu, Qwen Office and Qoder sold monthly and weekly cards at roughly 50% to 67% off. Tmall reportedly said token sales rose 450% month over month on the first sales day.
Independent benchmarker Artificial Analysis says Anthropic's Claude Fable 5.2 now leads its real-world robot-arm benchmark, AA-Robotics v0.1, with a score of 75, ahead of GPT-6 Astra at 69, and is the first model to score on the 'measure the box' task by using a tape measure's markings in video as a scale. The firm also said Google's Gemini 4 Carbon broke one of its grippers.
United said on Nov. 11 it will restrict unauthorized third-party personal AI agents across its own site and app and via travel agencies and aggregators, citing account privacy, unintended purchases and FAA/DOT rules. CEO Scott Kirby said he doesn't want "another middleman." Agents can still show flights, but travelers must complete purchases themselves.
At Baidu World 2026 on Nov. 12, founder Robin Li said Baidu's daily active AI agents grew from about 5,000 a year ago to 37,000, and the tasks they complete rose from 84,000 a week to more than 1 million. He called the "intelligent organization" AI's next act and said jobs won't vanish but will converge into builders and sellers. The figures are Baidu's own and haven't been independently verified.
Security firm Claroc says 35,578 of the 37,571 Grok-built apps it found online opened without sign-in, and about 1 in 10 held business or personal data, including a hospital staff list, an ICU patient briefing and plain-text passwords. Twenty-four were built by X and SpaceXSI staff. App addresses could be found in searchable DNS records, and sharing defaulted to "Anyone with the link." The company says creators choose access before publishing.
On the morning of Nov. 11 Beijing time, Alibaba said its next-generation Qwen4 model family is coming and opened a preview of the flagship Qwen4-Max on its Token Plan subscription. The model targets coding and AI-agent work, takes images and video, and handles a 1-million-token context. Alibaba has not released benchmarks, a parameter count or a technical report, and the preview shuts off on Nov. 25.
On November 11, Baidu released ERNIE 5.5 Flash Preview, the first post-trained model built on its previously open-sourced ERNIE 5.5 Flash Base. It has 280 billion total parameters, 14 billion active per token, and a 1-million-token context window. Baidu says it is strong at coding, 3D modeling, office work, and content creation, and showcased it building a third-person martial-arts game in Unreal Engine in a single 11-hour coding task. The API is live on Baidu Qianfan, and Baidu says the weights will be released under Apache 2.0 by November 25. Baidu also calls it an in-progress checkpoint that may still have gaps; the demos are the company's own claims.
A lot of AI work is not writing prose. It is making a fast call: should this be refunded, which model should handle it, does this break a policy? On November 11, Google introduced the Gemini Decisions API and a dedicated model, Gemini Decision 1, which it says can take text, images, audio, video, or a PDF and score up to 1,000 questions in one pass, with 150 to 250 milliseconds of end-to-end latency and no charge for output tokens. The speed and pricing figures come from Google.
Every cell in the body carries the same DNA, but a heart cell and an immune cell switch different genes on and off. On November 11, Google DeepMind released AlphaGenome 2, which it says can predict in which cell type a DNA change will matter. On a large immune-cell dataset, the company reports it identified about twice as many causal variants as the previous model. The tool is available for non-commercial research and is not approved for medical diagnosis.
The Signal protocol protects billions of private conversations, but until now its security had mostly been proven on paper, at the design level, not in the code that actually runs on phones. Google DeepMind said on November 10 that its AI framework AlphaProof Code spent about four weeks formally proving key security properties of Signal's real implementation, work that has historically taken expert teams years. When two real, already-fixed bugs were put back into a test copy, the system proved those versions were insecure. The proofs cover specific expert-reviewed properties, not every part of the Signal app, and the results are DeepMind's own.