Daily AI Digest

2026-11-22

Source:橘鸦 AI 早报 · 12 items

2026-11-22
2026-11-22 2026-11-21 2026-11-20 2026-11-19 2026-11-18 2026-11-17 2026-11-16 2026-11-15 2026-11-14 2026-11-13 2026-11-12 2026-11-11

Antigravity resets Gemini quotas for all users

开发生态

Antigravity resets Gemini quotas for all users

Google Antigravity head Varun Mohan said the platform has reset Gemini quotas for all users and completed its previously announced caching improvements. According to Mohan, users can now complete 22% more work within their quotas than before.

Google Antigravity has reset Gemini quotas for all users and completed its previously announced caching improvements, according to an update shared on X by its head, Varun Mohan. He said the caching changes allow users to complete 22% more work within their quotas than before; the figure describes the change in work possible within a quota, not an increase in the quota itself. Mohan also closed the same post by inviting users to build with Neon over the weekend.

Read original →

Google Antigravity updates /btw to let users branch into side tasks without losing progress

开发生态

Google Antigravity updates /btw to let users branch into side tasks without losing progress

Google Antigravity has expanded its /btw command so users can open a branch outside the main task, continue exploring it, and promote it to the main thread while retaining their current progress. The official announcement includes a demo video and invites users to try it in the CLI.

Google Antigravity has updated its /btw command to support opening branches outside the main task and promoting them to the main thread; the source does not specify a release date. According to the official announcement, the update goes beyond asking questions while the main task runs: users can continue exploring an issue in a branch without losing their current progress when switching threads. The official announcement includes a demo video and invites users to try the feature in the CLI. Sharing the announcement, Antigravity head Varun Mohan described it as an experience improvement he frequently uses and said it helps him understand progress during long-running tasks. He also said more features for checking progress would follow.

Read original →

Blogger claims Cursor will soon be renamed Grok Build

开发生态

Blogger claims Cursor will soon be renamed Grok Build

Blogger Puck says Cursor will be renamed Grok Build and Grok Build CLI 2.0 has entered the alpha-unified channel. Cursor has not officially confirmed the rename, and its connection to a combined subscription remains Puck’s speculation.

Blogger Puck says AI coding tool Cursor will soon be renamed Grok Build, but provided no date for the change, and Cursor has not officially confirmed it. He also speculates that the rename will be tied to a unified subscription plan covering both Grok and Cursor. According to Puck, Grok Build CLI 2.0 has already entered the alpha-unified channel, uses Cursor’s agent engine, and lets users sign in with either a Grok or Cursor account. He says the related announcement also promises faster desktop and web apps, Projects supporting collaboration among multiple agents, and models from multiple providers integrated into a single app.

Read original →

OpenCode announces 60% price cuts for GLM 5.3 and DeepSeek Flash

开发生态

OpenCode announces 60% price cuts for GLM 5.3 and DeepSeek Flash

OpenCode announced a 60% cut in usage-based pricing for GLM 5.3 and DeepSeek V4.1 Flash, covering API calls and usage beyond Go subscription allowances. According to the company, these are new official prices, not a limited-time offer.

OpenCode announced a 60% price reduction for GLM 5.3 and DeepSeek V4.1 Flash under its pay by token model; the source does not specify an effective date. The change covers direct API usage as well as token-based charges incurred after a Go subscription allowance is exhausted. Go subscription allowances already include a discount. OpenCode’s dax said the change establishes new official pricing rather than a limited-time promotion, and that both models are hosted on its own GPUs. He also said the new prices are 60% lower than those of Fireworks and Baseten, with higher throughput than either provider, and that the service is not exclusively for coding agents.

Read original →

OpenAI voice team member announces post-call transcripts in Dots

产品应用

OpenAI voice team member announces post-call transcripts in Dots

OpenAI voice team member Atty Eleti announced that Dots now supports viewing transcripts after calls end and shared related screenshots. The feature provides a text record of a completed call; the source does not specify a launch date.

OpenAI voice team member Atty Eleti announced in a post that Dots users can now view a transcript after a call ends. According to Eleti’s explanation, the transcript available to users is a text record of that call, accessible after the call has ended. Eleti also attached related screenshots to the post to illustrate the feature. The source provides no specific launch date and does not specify supported versions, pricing, or availability scope.

Read original →

Hidden periodic table toggle discovered in Gemini's web version

产品应用

Hidden periodic table toggle discovered in Gemini's web version

X user Lumina demonstrated a hidden switching interface in Gemini’s web version: holding down the model selector reveals a periodic-table layout listing 4 Gemini 4 model names. Which account tiers can access the interface remains unclear.

In a demonstration, X user Lumina opened a switching interface called The Periodic Table of Gemini by holding down the model selector in Gemini’s web version. The interface arranges its options as a periodic table, listing the model names Gemini 4 Hydrogen, Argon, Carbon, and Neon, with Nano Banana occupying the position corresponding to potassium. Regarding account visibility, a user reported that the interface did not appear on their Plus account and, on that basis, speculated that it might be available only to Pro and Ultra users; that proposed access restriction remains user speculation.

Read original →

OpenAI lets businesses outside the US claim their business profiles and publish agents in ChatGPT

产品应用

OpenAI lets businesses outside the US claim their business profiles and publish agents in ChatGPT

OpenAI now lets eligible businesses across regions where ChatGPT is available claim their business, update information and access visibility insights, extending beyond the US. Business Agents, however, remain limited to eligible US ecommerce and online service businesses.

OpenAI currently allows eligible businesses in regions where ChatGPT is offered to claim their business, while Business Agents remain available only to eligible ecommerce and online service businesses in the US. According to the company, claiming is initially available to businesses with a website, and eligibility depends on industry, location, and verification and integrity checks. Only owners or authorized representatives may manage a business profile. After claiming, businesses can update information in Business Hub, and their profile displays a blue check. They can also see how often they are mentioned in ChatGPT and the conversation topics in which their business appears.

Businesses do not need a Shopify store to claim their business, and those with an OpenAI account executive can work through that representative. Retailers can upload a product feed containing product details, pricing and availability; connecting Shopify automatically shares this information through Shopify’s catalog. Eligible merchants can configure the information and guidelines their Business Agent uses, test it and publish it to answer product and policy questions. OpenAI says ChatGPT may use business-provided information in its answers, and businesses may appear in relevant answers even without an Agent. Claiming does not guarantee higher placement or sales, and ChatGPT continues to make independent recommendations and comparisons.

Read original →

微信 tests desktop AI Agent “做个小工具” for creating apps with a single prompt

产品应用

微信 tests desktop AI Agent “做个小工具” for creating apps with a single prompt

WeChat is testing the desktop AI Agent “做个小工具,” with macOS and Windows Beta downloads and QR-code signup available. According to the report, the connected service WeChat Xiaowei has more than 30 million daily active users during testing.

WeChat is testing the desktop AI Agent “做个小工具,” a standalone client built from WeChat Xiaowei’s “小工具” functionality, with macOS and Windows Beta versions available and QR-code signup open on its official website. According to tests by several bloggers, a cloud-based subagent handles development tasks and generates code in the cloud without leaving files on the local machine, then deploys the result directly. Finished tools can be sent through WeChat Xiaowei for use on PCs and phones. The client uses WeChat’s internally developed WeLM model, with one tester identifying the version as welm-v4.6-flash. The report also states that WeChat Xiaowei has more than 30 million daily active users during testing and that these tools require no registration, regulatory filing, certification, submission for review, or review.

Read original →

Lean and Mathlib communities certify proof of the Riemann hypothesis as correct

技术与洞察

Lean and Mathlib communities certify proof of the Riemann hypothesis as correct

The Lean and Mathlib communities said in a joint statement that they had verified Levent Alpöge’s proof of the Riemann Hypothesis, released on November 20, 2026. The proof is provided as a Lean library containing 17 million lines of code.

Levent Alpöge released a proof of the Riemann Hypothesis on November 20, 2026, as a Lean library comprising 17 million lines of code; the Lean and Mathlib communities subsequently said in a joint statement that they had confirmed the correctness of both the theorem statement and the proof. The statement was endorsed by the Lean FRO and the Mathlib Initiative. According to the statement, the codebase depends directly only on Mathlib, and the two communities conducted verification work during the day preceding the statement, confirming that all four criteria listed in an article linked in the statement were satisfied. The statement also thanked community members who provided resources for the verification.

Read original →

Novita AI confirms some MiniMax requests were forwarded to the official API

行业动态

Novita AI confirms some MiniMax requests were forwarded to the official API

Novita AI confirmed that, from May 28 to November 11, some MiniMax requests submitted directly or routed through OpenRouter were sent to MiniMax’s official API, potentially allowing third parties to access or retain their contents.

Novita AI disclosed in a privacy notice that some MiniMax requests were forwarded to MiniMax’s official API between May 28 and November 11, including requests submitted directly through Novita and those received through OpenRouter; the notice did not specify the year. According to the company, it suspended the affected models after receiving a report from an independent security researcher. It attributed the issue to a routing configuration error that applied a forwarding path intended for specific customers more broadly. The contents of those requests may have been accessed or retained by third parties. Novita said it does not retain customer prompts or generated outputs or use them for model training, and is checking data retention and deletion with MiniMax. The platform advised revoking or rotating confidential information such as API keys if it was included in requests.

Read original →

台积电 reportedly suspends orders from Chinese companies for advanced processes at 16nm and below

行业动态

TSMC has suspended chip orders from Chinese companies for processes of 16nm and below, according to AppleInsider. Resuming orders requires U.S. permission or exemption eligibility and order approval; customers can no longer rely on TSMC to assess exemptions itself.

TSMC has notified customers that it is suspending chip orders from Chinese companies for advanced processes of 16nm and below, according to AppleInsider. Under tightened U.S. export controls, affected customers must obtain U.S. permission in advance or qualify for an exemption, and orders can resume only after approval. The previous arrangement, in which TSMC assessed whether chip designs qualified for exemptions, has been replaced by advance applications from customers. The current U.S. list of approved companies includes Apple, AMD and Nvidia, but no Chinese companies. The report states that, as a result, Chinese companies can obtain only chips made using older processes such as 28nm or 40nm and cannot use the HBM high-bandwidth memory used in AI data centers. TSMC said it would comply with the new rules.

Read original →

腾讯混元 says next-generation Hy models will scale up significantly, with Hy4-mini already on the roadmap

前瞻与传闻

腾讯混元 says next-generation Hy models will scale up significantly, with Hy4-mini already on the roadmap

Tencent’s Hunyuan team said in a Reddit AMA that the next Hy model will have more parameters and continued investment in large-scale reinforcement learning, with Hy4-mini and Flash-Lite both planned. The team also explained why Hy4 uses MLA.

Tencent’s Hunyuan team outlined its expansion plans for the next Hy model during a Reddit AMA, confirming that Hy4-mini is on the development roadmap and Flash-Lite is also in the planning stage. The source provides neither a specific date for the event nor release dates for the models. According to the team, future work will increase the model’s parameter count and continue investment in large-scale reinforcement learning.

For Hy4, the flagship model whose weights have already been released, the team described how it selected its attention mechanism. According to its explanation, development involved repeated comparisons between GQA and MLA, with MLA ultimately chosen because DSA reduced its computational intensity during decoding. The source also states that Tencent had released Hy4’s weights and launched Hy4 flash before this disclosure, but gives no specific dates for either release.

Read original →