Daily AI Digest

2026-11-20

Source:橘鸦 AI 早报 · 22 items

2026-11-20
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

ChatGPT dots opens to Business Standard users and select Plus users

要闻

ChatGPT dots opens to Business Standard users and select Plus users

OpenAI employee Tibo announced that ChatGPT’s dots has entered testing with a small number of Plus users and is being made available to all Business Standard users. He also said that multiple rounds of improvements had yielded positive feedback on its proactive behavior.

OpenAI employee Tibo announced that testing of ChatGPT’s dots had begun with a small number of Plus users, with access for all ChatGPT Business Standard users starting on the day of his post. The source does not provide the specific publication date. Availability differs between the two plans, with Plus access currently limited to a small number of users. According to Tibo, dots has undergone multiple rounds of improvements and has recently received considerable positive feedback on its proactive behavior. He said he hopes to continue collecting user feedback as testing expands. Tibo’s X profile lists work related to Codex and ChatGPT, and related posts and test screenshots have also appeared in the linux.do community.

Read original →

ChatGPT health features expand to desktop, Android, and dots

要闻

ChatGPT health features expand to desktop, Android, and dots

OpenAI is gradually expanding ChatGPT’s Health feature to desktop apps, Android, dots and Sites for Free, Go, Plus and Pro users in the United States, with additional plan restrictions for Sites and dots.

OpenAI announced the expansion of ChatGPT’s Health feature to desktop apps, Android, dots and Sites, with the update rolling out gradually. The feature is available to Free, Go, Plus and Pro users in the United States, but Sites is restricted to Plus and Pro, while dots is currently limited to Pro. According to OpenAI, Health supports connecting health data through Android Health Connect so users can analyze sleep and exercise patterns. When meetings conflict with workouts, a dot proactively suggests rescheduling, and users can also create sites to track their marathon training progress.

Read original →

Antigravity desktop app adds SSH connectivity

要闻

Antigravity desktop app adds SSH connectivity

The Antigravity desktop app has added SSH connections to remote Linux hosts in separate windows, with automatic remote server setup. Projects and Agents run on the host, and Agents can keep working after the window closes or the laptop goes to sleep.

The Antigravity desktop app now supports SSH connections to remote Linux hosts; the source does not specify a launch date. According to its documentation, the app automatically sets up a remote server over OpenSSH so that projects, Agents, Git worktrees, terminals, and customizations run directly on the remote host. Closing the connection window does not stop the remote server or Agents, and Agents continue working when the laptop goes to sleep. Each host opens in a separate window, with the connection identified in the title bar and sidebar.

Connections start from the app sidebar, where users can select a recent host, read hosts from ~/.ssh/config, or enter a host address. An alternative configuration file can also be specified. The app automatically uses OpenSSH keys, ssh-agent, and ProxyJump settings. After the remote window opens, sign-in completes through an OAuth flow in the local browser, with credentials saved on the remote host. Supported options include personal Google Accounts, Google Workspace, Pay As You Go, and Gemini Enterprise. Users can switch to a local window without disconnecting. Uninstalling the remote service requires running Stop SSH Server… and then deleting ~/.antigravity-hub-server on the host.

Read original →

Gemini app for Mac adds computer use

要闻

Gemini app for Mac adds computer use

Google has begun rolling out computer use for the Gemini app on Mac in English in the U.S. According to Google, Gemini can carry out described tasks across local files, desktop apps, and the web in the background, while users can restrict app access and pause execution at any time.

Google has started rolling out computer use in English for the Gemini app on Mac in the U.S.; the source does not specify a launch date. According to Google, users can describe multi-step workflows in text and have Gemini handle menu navigation, clicks, typing, and form entries across local files, desktop apps, and the web. Execution takes place in an isolated background window with a live preview in the chat, leaving the main screen and cursor available to the user. Users can choose which apps Gemini may access, monitor actions live, pause at any time, and approve key actions before execution. The app is available to download at gemini.google/mac.

Read original →

Gemini API previews on-demand tool loading

开发生态

Gemini API previews on-demand tool loading

Google has introduced Tool Retrieval in preview in the Gemini Interactions API, letting Agents load tool parameter schemas on demand rather than send an entire tool library each time. The feature supports Gemini 3.6 and later models and carries no separate charge.

Google has launched Tool Retrieval in preview in the Gemini Interactions API, supporting Gemini 3.6 and later models, including all Gemini 4 models. Developers enable it by adding tool_search to a request and setting defer_loading: true on tools whose loading should be deferred. The model can still see those tools’ names and descriptions, but their parameter schemas stay out of the initial context. When a tool is needed, the model requests it by name, the API loads its full schema, and the model generates a function_call. These steps happen within the same request and are recorded in interaction.steps. Google says the approach is intended to reduce input tokens and latency while maintaining tool-selection accuracy as tool libraries expand.

By default, Gemini performs retrieval on the server, so developers do not need to build or host an index. Frequently used tools can also be loaded upfront while the remaining tools are deferred. The feature works with prefix caching and supports setting defer_loading: true on an mcp_server tool for remote MCP servers. Developers maintaining their own tool registry, embedding index, or user-permission filter can set "execution": "client" on tool_search, but each search adds one round trip. Tool Retrieval has no separate charge, and input and output tokens are billed at the selected model’s standard rates. tool_search adds a small, fixed block of input tokens, while loaded schemas continue to be billed as input throughout the rest of the conversation.

Read original →

Qoder ends daily 100 Credits claims for free users

开发生态

Qoder ends daily 100 Credits claims for free users

Qoder CN will end free users’ eligibility to claim 100 general-purpose Credits daily from November 19, 2026, at 10:00. Trial users, individual paid subscribers, and membership card holders remain eligible, while previously claimed balances retain their original expiration dates.

Effective November 19, 2026, at 10:00, Qoder CN will restrict its daily 100 general-purpose Credits program to trial users, individual subscribers to the Professional, Advanced, and Flagship plans, and membership card holders, excluding Experience (free) users. According to the official explanation, Credits claimed before the change will not be revoked and will remain usable for 30 days from their respective claim dates. Enterprise subscribers, including those on the Team, Enterprise Standard, and Enterprise Dedicated plans, are not eligible. Eligibility is determined by the account’s subscription or membership card status at the time of claiming. Users become ineligible starting with the next claim round if their trial ends without a paid subscription, their individual paid subscription expires or is canceled, or their membership card expires.

Each claim window runs from 10:00 on one day until before 10:00 the following day. Each account must actively claim its allocation and may claim 100 general-purpose Credits once per round, exclusively through Qoder CN’s desktop app, mobile app, or IDE. Missed claims cannot be recovered, and unclaimed allocations do not accumulate. According to the official explanation, rewards are recorded as Add-on Credits, are for personal use only, and can be used across all Qoder CN products without changing subscription tiers or membership card benefits. Each reward has its own 30-day validity period and can be retained across days, but subsequent claims do not extend earlier rewards’ expiration dates. Credits expiring earliest are consumed first; when expiration dates match, plan Credits are consumed before resource-pack Credits.

Read original →

ChatGPT rolls out a new mobile interface, starting with select Pro users

产品应用

ChatGPT rolls out a new mobile interface, starting with select Pro users

According to an OpenAI employee, ChatGPT is rolling out a new mobile interface to some Pro users who have a dot or are eligible to create one; dot will have a dedicated tab, and users can also pin Codex, Projects, and Images to the tab bar.

ChatGPT is rolling out a new mobile interface, initially to some Pro users who have a dot or are eligible to create one, according to an OpenAI employee. The changes focus on access to features: dot gets a dedicated tab, while users can pin features such as Codex, Projects, and Images to the tab bar. According to the employee, these changes are intended to help users find frequently used features and discover other ChatGPT capabilities. Some users will see the new interface by default, but users can turn it on or off in settings at any time.

Read original →

Google AI Studio launches a periodic table-style model selector

产品应用

Google AI Studio launches a periodic table-style model selector

Google AI Studio has launched Model Selection, displaying options including Gemini 4 in a periodic-table-style grid. Team member Ammaar Reshi confirmed that the interface idea came from a joke video previously posted by user seteiro.

Google AI Studio team member Ammaar Reshi confirmed in a post that the Model Selection feature is officially live; the source does not provide a specific launch date. The feature arranges available models, including Gemini 4, in a periodic-table-style grid. The interface idea originated in a joke video by user seteiro depicting what Google’s model selection screen might look like a few months later. Reshi subsequently quoted that post and said the team had implemented and released the idea.

Read original →

Grok Bot community directory receives over 1,200 submissions in its first day

产品应用

Grok Bot community directory receives over 1,200 submissions in its first day

Grok Bot Directory, introduced by SpaceXSI staff member Eric Zakariasson, has launched. According to its author, it received more than 1200 bot submissions in its first day, and projects were added to Grok Bot more than 7000 times.

SpaceXSI staff member Eric Zakariasson introduced the community directory Grok Bot Directory in a post, reporting more than 1200 bot submissions in its first day and more than 7000 instances of projects being added to Grok Bot. The directory provides browsing, installation and submission options, allowing users to find bots and apps that others have built for Grok Bot and to submit their own creations. Its listings also include projects submitted to the official marketplace but not yet approved for publication. Addressing the review process in a reply, Zakariasson confirmed that listed projects undergo review before receiving approval to enter the marketplace.

Read original →

Google to bring Genie 4 to Gemini API and AI Studio

模型发布

Google to bring Genie 4 to Gemini API and AI Studio

Google has made Genie 4 available to developers through the Gemini API and Google AI Studio. According to the company, the model generates continuous worlds at 720p and 24fps with synchronized native audio and supports real-time interaction.

Google announced that Genie 4 is now available in the Gemini API and Google AI Studio, with API access provided through the Interactions API. Developers can supply a text prompt, a reference image, or both as input. According to Google, the generated continuous worlds support 720p visuals at 24fps and synchronized native audio, including ambient sound, sound effects, and character dialogue. Users can combine real-time keyboard controls with live text prompts to explore environments and change weather, lighting, objects, or events. The company says Gemini reasoning maintains consistency and coherence in the world state, and all outputs include imperceptible SynthID watermarking. Google AI Studio also provides an interactive canvas and starter apps.

Read original →

火山引擎 launches API access to 豆包's real-time voice and video calling model

模型发布

火山引擎 launches API access to 豆包's real-time voice and video calling model

Volcano Engine has launched API access to Doubao’s SeedRealtime 1.0 real-time audio-video calling model, which handles sound, visuals, temporal sequencing and expression in a single model. The company says audio-video conversational timing problems are halved compared with traditional cascaded systems.

Volcano Engine announced that the API service for Doubao’s SeedRealtime 1.0 real-time audio-video calling model is now officially available. The service targets use cases including voice assistants, smart hardware, companion applications and real-time customer support. The model uses an end-to-end architecture to process sound, visuals, temporal sequencing and expression together, supporting real-time interaction while seeing, listening and speaking. It can also choose when to initiate speech as it continuously interprets changes in a scene. According to the company, compared with traditional cascaded systems that connect modules such as ASR, VLM and TTS, the model reduces accumulated latency and information loss, halves audio-video conversational timing problems and increases the probability of a complete, smooth exchange within a single conversation.

Read original →

GPT-Live-1.1 launches on API with a new standby option

模型发布

GPT-Live-1.1 launches on API with a new standby option

OpenAI has made GPT-Live-1.1 available through the API, introducing a lower standby rate for WebRTC while inference is paused. According to OpenAI, the model can listen while speaking and improves transcription accuracy and accent and voice consistency across languages.

OpenAI has made GPT-Live-1.1 available through the API under the model identifier gpt-live-1.1; the source does not specify a launch date. For billing, WebRTC sessions qualify for a lower standby rate while inference is paused, with specific prices listed on the GPT-Live pricing page. Account-specific rate limits are available under the organization's limits. On capabilities, OpenAI says the model can continue listening while speaking and, compared with GPT-Live-1, transcribes speech more accurately and maintains a more consistent accent and voice across languages. Developers can describe in natural language when tasks should be handed to their backend for reasoning or tool use. The documentation also provides instructions for starting voice sessions, writing prompts, and configuring task delegation.

Read original →

Cloudflare releases two Clairic information extraction models

模型发布

Cloudflare has released two open-weight extraction models, 27B Clairic and 9B Clairic-flash, on Workers AI. Both cost $0.10 per million input tokens during the promotional period.

Cloudflare now offers 27B Clairic and 9B Clairic-flash on Workers AI, with full weights for both models available on Hugging Face. The update moves the preview’s 9B model to @cf/cloudflare/clairic-flash, while the existing @cf/cloudflare/clairic endpoint now serves the 27B model. Existing applications will receive the new extraction capabilities as the original endpoint is updated. Both models are billed only for input tokens, with no separate charge for output tokens. During the promotion, both cost $0.10 per million input tokens. When the promotion ends on December 1, 2026, 27B Clairic returns to $0.24, while 9B Clairic-flash remains at $0.10.

Read original →

Ideogram 4.6 launches with a focus on native 4K editing

模型发布

Ideogram 4.6 launches with a focus on native 4K editing

Ideogram has released image model 4.6 with native 4K output, making it available through its app, API, and partner platforms. The company reports a score of 1892 on Rapidata’s multi-step image editing benchmark, with API pricing starting at $0.008 per image.

Ideogram image model 4.6 has been released and is now available through the Ideogram app, API, and partner platforms Leonardo, Runway, and fal, with native 4K output and API pricing starting at $0.008 per image. Editing capabilities are the focus of this update: according to data published by the company, 4.6 ranked first on Rapidata’s multi-step image editing benchmark with 1892 points, compared with 1866 for the previous model, 4.5. Scores exceeded those of 4.5 across all five editing tasks, including character consistency, precise recoloring, and masked editing, with precise recoloring rising from 46% to 83%. The company also says the new model improves realism, texture detail, and color accuracy while reducing pixel drift outside the edited area.

Read original →

OpenAI details shifts in AI R&D: agent use grows, but human oversight remains essential

技术与洞察

OpenAI said in a research progress report that coding agents are changing its AI R&D workflows, with researchers’ use of agentic AI tools continuing to grow exponentially. Success rates for completing delegated tasks autonomously are rising overall, but outputs still require human review.

OpenAI released a research progress report describing how coding agents are accelerating and changing its internal AI R&D workflows. The report did not provide a publication date. According to the company, researchers’ use of agentic AI tools continues to grow exponentially, and agents’ success rates in completing delegated tasks without human intervention are rising overall, although performance varies considerably across tasks. Researchers said agents can help write code and run experiments, but their outputs still require human review. Separately, OpenAI’s internal measurements showed that Safety & Alignment recorded the largest increase in research investment among the research areas identified.

Read original →

OpenAI reveals how a Codex agent automatically fixed a performance regression

技术与洞察

OpenAI reveals how a Codex agent automatically fixed a performance regression

OpenAI disclosed Performance Factory, an internal system powered by Codex that automatically investigates and fixes ChatGPT performance regressions. The team said it tracked over 1700 investigations in just over three months, with approximately 15 deployed fixes directly using Codex code.

OpenAI disclosed Performance Factory, an internal system powered by Codex for automatically investigating and fixing ChatGPT performance regressions. According to the company, a performance alert triggers an investigation task, and an Agent identifies the cause and proposes a fix. Engineers review the code before merging and deploying it, after which another Agent verifies whether performance has recovered. The performance team said the system tracked over 1700 investigations in just over three months and identified root causes in more than 600 cases. Among deployed fixes, approximately 50 were influenced by Agent suggestions, and approximately 15 directly used code submitted by Codex. The company said proposing reliable fixes remains harder than identifying causes, and improving fix quality remains a priority.

Read original →

OpenAI unveils frontier AI risk management measures

技术与洞察

OpenAI unveils frontier AI risk management measures

OpenAI has announced frontier AI risk management measures covering safety reviews across three stages: research, internal deployment, and public deployment. According to the company, the Control Room in its research infrastructure can halt a task’s subsequent model outputs within 35 seconds of receiving a stop instruction.

OpenAI has announced frontier AI risk management measures requiring models to pass safety reviews at every stage of research, internal deployment, and public deployment. To establish the basis for those reviews, the company has introduced processes including safety cases, requiring research tasks to submit safety arguments and supporting evidence, as well as more extensive evidence of behavioral alignment during training. For task execution control, OpenAI says the Control Room in its research infrastructure works with monitoring systems to stop tasks when potential problems are detected; after receiving a stop instruction, the system can halt a task’s subsequent model outputs within 35 seconds.

Read original →

Biggest IPO in history: Anthropic closes debut session up 17%

行业动态

Anthropic listed on Nasdaq, raising $81.3 billion through the issuance of 980 million shares and surpassing SpaceX’s $75 billion to become the largest IPO on record; its shares closed their first trading day at $97.17, up 17.07% from the offering price.

Anthropic began trading on Nasdaq under the ticker ANT on November 19, local time; the source does not specify the year. The IPO was priced at $83 per share, above the previous range of $70 to $78 per share. The company issued 980 million shares and raised $81.3 billion, exceeding SpaceX’s $75 billion and setting a record for IPO proceeds. Shares opened at $111.50, approximately 34.3% above the offering price, and finished at $97.17, a gain of 17.07% over the offering price. Based on the closing price, its market capitalization was approximately $1.27 trillion to $1.3 trillion on a common-stock basis and nearly $1.8 trillion on a fully diluted basis.

Read original →

DeepSeek reportedly secures 100 billion yuan in funding from 31 institutions

行业动态

DeepSeek reportedly secures 100 billion yuan in funding from 31 institutions

DeepSeek has completed its second external funding round, according to Caixin, with new investors holding about 17.27% through Hangzhou Chengyan, which added 31 limited partners. The related registration changes were completed on November 17, 2026, and Liang Wenfeng retains control of the platform.

DeepSeek has completed its second external funding round, according to Caixin, and its operating entity, Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., completed its corporate registration changes on November 17, 2026. Registration records show that the company’s registered capital increased by 21.1%, from RMB 16.4475 million to RMB 19.9187 million. Its new shareholder, Hangzhou Chengyan Enterprise Management Consulting Partnership, holds about 17.27% and serves as the equity-holding platform for the round’s new external investors.

Hangzhou Chengyan also completed registration changes on November 17, 2026, increasing its registered capital from RMB 6,000 to approximately RMB 3.439 billion and adding 31 limited partners. Hangzhou Chengce Business Consulting Co., Ltd. remains its executive partner. DeepSeek founder Liang Wenfeng is no longer a direct partner in Hangzhou Chengyan, but he continues to hold 99% of Hangzhou Chengce Business Consulting Co., Ltd., retaining control over Hangzhou Chengyan.

Read original →

腾讯 President Liu Chiping says 元宝 fell short of expectations, prompting sharp investment cuts

行业动态

Tencent disclosed at its earnings briefing on November 18, 2026, that Yuanbao had fallen short of expectations and that it had cut investment while shifting its focus to model capabilities. Its other AI assistant, WeChat’s Xiaowei, reached 30 million daily active users and more than 30 touchpoints within WeChat.

Tencent President Martin Lau said at the third-quarter earnings briefing on November 18, 2026, that Tencent had cut investment in Yuanbao after the AI assistant fell short of the company’s expectations. Its focus would shift from pursuing user scale alone to improving model capabilities. At the same briefing, Tencent reported that WeChat’s Xiaowei AI assistant had 30 million daily active users and more than 30 touchpoints within WeChat. Tencent’s earnings report said Xiaowei had expanded its phased testing coverage and use cases while deepening integration with Mini Programs and WeChat’s shop functionality. Lau said Xiaowei uses a small model because of the cost considerations involved in serving all users. According to him, even with access extended to everyone, its quarterly cost would remain below Yuanbao’s cost at the beginning of 2026.

Tencent’s revenue for the third quarter of 2026 was RMB 207.336 billion, up 8% year on year, while Non-IFRS operating profit was RMB 68.073 billion, down 6%. Excluding contributions from new AI products, operating profit on the same basis was RMB 83.3 billion, up 8%, and the operating margin was 40%, unchanged from a year earlier. Without that exclusion, the margin was 33%, compared with 38% a year earlier. The earnings report identified the main new AI products as Hy (Hunyuan), Yuanbao, CodeBuddy, WorkBuddy and Xiaowei. On Xiaowei’s monetization, Lau identified advertising and transaction matching as two direct avenues, with service matching between Agents as a longer-term direction.

Read original →

阿里's Wu Yongming says 千问 is already training its next-generation model on 平头哥 chips

行业动态

Alibaba Group CEO Eddie Wu disclosed that Qwen is already training its next-generation model on T-Head chips and that Bailian is providing inference services using domestically produced chips. Alibaba Cloud separately said its next-generation AI chip, Zhenwu V900, is scheduled for mass production in the first quarter of 2027.

Alibaba Group CEO Eddie Wu disclosed during the analyst earnings call for the second quarter of fiscal year 2027 that Qwen has begun training its next-generation model on T-Head chips and that Bailian is already providing model inference services using domestically produced chips. He said coordination between in-house chips and models would continue to improve the cost per unit of computing capacity. On future supply, he projected that 2027 and 2028 would be the two years of China's most acute computing capacity shortages. After that, he said, Alibaba Cloud's chip supply would primarily consist of domestically produced T-Head chips, alongside purchases of other domestically produced chips for internal inference. Separately, Alibaba Cloud said mass production of its next-generation AI chip, Zhenwu V900, is planned for the first quarter of 2027.

Read original →

OpenCode says its own GPUs handle 44% of tokens

行业动态

OpenCode says its own GPUs handle 44% of tokens

OpenCode said its own GPUs handled 44% of tokens on the day of its disclosure, totaling 16 trillion, with 10 times as much hardware due to come online over the next few months. A team member also said its existing 1,000 GPUs were all on short-term contracts.

In a disclosure without a specified date, OpenCode said its own GPUs served 44% of that day’s tokens, totaling 16 trillion, and that 10 times as much hardware would come online over the next few months. On its existing hardware, team member Dax Raad had previously reported holdings of 1,000 GPUs, comprising 40% GB300 and 60% B300, all on short-term contracts. According to Raad, the GPUs are distributed across the United States and other international locations, excluding China; capacity constraints mean many are located in India or Malaysia. Raad also recalled that the success and price increases of DeepSeek 4.1 had broken the team’s cost structure. At that point, the team had no hardware or leases; 3 months later, he said, it could run DeepSeek and GLM at the same quality as their developer labs. That quality claim is Raad’s own account.

Read original →