Anthropic open-sources FlashProve to verify which model generated an AI response
要闻
Anthropic has released FlashProve, an open-source prototype that uses zkSNARKs to verify whether an AI response came from the specified model. On the 400B-parameter Llama 4 Maverick, one H100 generated proofs at more than 100 tokens per second.
Anthropic has released FlashProve as an open-source software prototype for provable inference. The release provides code, Lean formalization, a technical description of the protocol, and documentation under the Apache 2.0 license. The system uses zkSNARKs to generate a short proof, allowing users to check whether a response was produced by the specified model without obtaining its weights or running the model again. The proof reveals no information about the model’s weights. According to Anthropic, on Llama 4 Maverick, a 400B-parameter mixture-of-experts model with 4-bit quantization, a single H100 GPU generated proofs at more than 100 tokens per second. Each 1,024-token request produced a proof of approximately 19 MB, while the GPU time was 70–95 times that required by vLLM to process prompts of the same length. The release contains five artifacts, and Anthropic describes it as intermediate progress intended to advance provable inference rather than a completed research agenda.
ChatGPT Side Chats no longer expire and can be saved indefinitely
开发生态
Brent Traut of OpenAI’s Codex team announced that ChatGPT Side Chats no longer have an expiration mechanism. They will not disappear after an app restart or update and will be stored like regular conversations, while users can still delete or archive them manually.
Brent Traut, a member of OpenAI’s Codex team, announced that ChatGPT Side Chats no longer disappear because of an app restart or update after expiring, and are now stored in the same way as regular conversations. Users can still delete or archive them manually, and Side Chats will also be archived or deleted together with their associated main conversation.
According to Traut, he found after joining OpenAI that Side Chats disappearing after expiration was one of the biggest user pain points. The issue has now been fixed. He also adjusted several usage details, with expected improvements to Side Chats startup speed, communication with the main conversation, tool calls, and the activity interface. Related changes for mobile and TUI are still in progress.
Stealth model Crouton Alpha launches on OpenCode and OpenRouter, free for one week
开发生态
According to OpenRouter’s page, Crouton Alpha, a stealth model developed and operated by an anonymous third-party provider, was released on November 3, 2026, with zero prices for prompts and completions.
Crouton Alpha was released on November 3, 2026, and OpenRouter’s page lists it as a stealth model developed and operated by an anonymous third-party provider. OpenRouter forwards requests but is not the model’s developer, owner, or provider. The model is hosted by a single provider and is not currently available through EU in-region routing.
The model is designed for scalable coding, agentic, and multimodal workflows. It supports tool calling, tool_choice, structured outputs through a JSON schema, and text, image, video, PDF and other file, and audio inputs, while returning text. Its context window is 1,048,576 tokens, and the page lists both prompt and completion token prices as zero. The provider may retain prompts and completions, but according to the stated policy, they are not used for training; all other use is governed by the Stealth Model Terms.
OpenAI team member says the model lineup for dots is being expanded
产品应用
Tibo, a member of the OpenAI team, said dots is expanding its available model lineup, currently including GPT-6 Astra and GPT-6.1 Terra, with more models to be added and assigned by task.
Tibo, a member of the OpenAI team, said dots is expanding the set of models it supports, with GPT-6 Astra and GPT-6.1 Terra now included and additional models planned for later. According to Tibo, Terra will improve speed and personality performance, allowing dots to serve more people. dots will select a suitable model for each task and use different models for different parts of the work. OpenAI previously described dots as being powered by GPT-6 Astra.
Doubao Work announced through its official WeChat account that it will provide individual users with free answers to everyday simple questions. The same conversation can continue into document or PPT creation, while later generation and in-depth research may use paid quota.
According to the announcement on its official WeChat account, Doubao Work has opened free access to everyday simple question answering for individual users. The company says that for small day-to-day questions, Doubao Work provides answers that are more accurate and reliable than Doubao’s chat mode. Users can start with a simple question and continue in the same conversation to write a document or create a PPT. Subsequent document generation, PPT creation, or in-depth research may consume paid quota.
MiniMax open-sources BYOB, allowing users to bring their own agent for shared interactions
产品应用
The MiniMax Code team has released BYOB Table Protocol v0.1 and a TypeScript reference implementation, enabling Agents owned by different people to collaborate through a shared ledger and seat system while requiring owner approval, execution by the owner’s Bot, and verification by another seat for affected resources.
The MiniMax Code team has released mcode-agent-table, which contains BYOB (Bring Your Own Bot) Table Protocol v0.1 and its TypeScript reference implementation. A Table Host running byob serve exposes HTTP + SSE endpoints, while Bots connect from their owners’ machines through an adapter, keeping owner keys off those machines. Table Core manages the shared ledger, seats, floor control and commit gateway. The gateway allows Bots to propose only; the affected owner approves, that owner’s Bot executes, and another seat verifies the ledger state. The project requires Node ≥ 22.19, mcode ≥ 0.6.3 on PATH and a MiniMax API key. Its example uses three MCode Agents with different owners, a judge and a deterministic checker. A decision affecting Jenny’s taskboard waits for Jenny’s signed approval, after which Jenny’s MCode edits the file and posts a receipt for the checker to verify. The project also includes an MCP Server, CLI, client library and the MiniMax Code Mini App in apps/rabbit-hole/; v0.1 currently runs on one Table Host.
ChatGPT Classic desktop app to be discontinued next January
产品应用
OpenAI has confirmed through an in-app notification that the ChatGPT Classic desktop app will be discontinued on January 30, 2027, with complete chat histories for current users transferred to the new ChatGPT app.
OpenAI has confirmed that the ChatGPT Classic desktop app will be discontinued on January 30, 2027. The notification was sent to users who are still using the older app, and according to the company, complete chat histories will move with their accounts to the new ChatGPT app. Clicking the in-app notice banner takes users to the download page for the new ChatGPT app.
Claude Team and Enterprise launch automatic plugin connection
产品应用
Claude has introduced automatic plugin connections for its Team and Enterprise plans, with official documentation covering setup and connection testing through JWT bearer grant, trusted issuer, and an MCP server.
Claude’s Team and Enterprise plans now provide automatic plugin connections. The official documentation lays out an enterprise-managed authorization workflow that moves from preparation to verification. It instructs users to fetch the complete documentation index and confirm support for the JWT bearer grant, then register a trusted issuer, understand the token request, add an MCP server as a custom connector, and open the managed authorization setup.
The documentation also requires registering the issuer in a test workspace and running a connection test to verify the connection. The page includes an entry for the documentation index and a “Was this page helpful?” feedback item, but it does not state the feature’s launch date, supported versions, or pricing.
Google launches Guided Learning for student accounts
产品应用
Google has launched Guided Learning in Gemini, using questions, step-by-step explanations, images, diagrams, videos and interactive quizzes to move learners from quick answers toward a deeper understanding of subjects.
Google has launched Guided Learning in Gemini. The feature uses probing and open-ended questions to engage learners, breaks problems into steps, and adjusts explanations to their needs. Its responses can include images, diagrams, videos and interactive quizzes to help learners understand and test their knowledge. It can be used to prepare for an exam about enzymes, draft a paper on bee populations and food systems, or explore photography.
Google developed Guided Learning with educators, pedagogical experts and students. In 2022, Google formed a cross-disciplinary team of AI experts, neuroscientists and cognitive scientists, then created LearnLM, a family of models tuned for learning, and brought those capabilities into Gemini 2.5 earlier this year. Educators can post a dedicated link directly in Google Classroom or share it with students. Users can try Guided Learning in the Gemini app and submit feedback.
WeChat’s vision team open-sources the WeVisOne-VL vision model
模型发布
Tencent’s WeChat Vision team has open-sourced WeVisOne-VL, a model that unifies detection, localization, counting, OCR, segmentation, and matting, while outputting None when a target is absent. It was trained on about 100 million images and evaluated across more than 30 benchmarks.
Tencent’s WeChat Vision team has open-sourced WeVisOne-VL. The vision-language model unifies object-level and pixel-level perception through structured next-token prediction. It can produce points, boxes, binary masks, and continuous alpha mattes at a user-specified granularity, and supports object detection, referring localization, counting, pointing, OCR, segmentation, and matting. WeVisOne-VL-4B uses 1,001 dedicated tokens (<0> through <1000>) to encode normalized positions: a box uses four tokens and a point uses two. When the queried target is absent, the model outputs None. According to the official description, the model was trained with a three-stage curriculum SFT followed by mixed-task RL on about 100M images, and led more than 30 benchmarks covering object-level and pixel-level tasks. Its training and feedback loop also addressed small, crowded objects and used model errors in subsequent data collection.
The model is built on Qwen3-VL without SAM or an external segmentation backbone. A 9.6M-parameter pixel decoder connects to the VLM through a special <OBJ> token and reuses Qwen3-VL’s own multi-level ViT features. Each <OBJ> hidden state is combined with a Fourier-encoded box prior to form an instance query, which is refined through six deformable cross-attention blocks. Dynamic kernels first predict masks, followed by a coarse-to-fine cascade from S/4 to S/2 and S/1 for full-resolution alpha mattes. The data pipeline uses a knowledge graph of categories and referring expressions for targeted acquisition; localization experts propose candidates, the VLM verifies them, SAM2 extends them to masks, and a dedicated annotator produces continuous alpha. Model errors are fed into the next round.
Claude Code product lead Cat Wu announces departure
行业动态
Cat Wu, Claude Code’s product lead, announced her departure from Anthropic, with Friday set to be her final day at the company. She looked back on the tool’s progression from an internal project to handling large migrations and new product development automatically.
Cat Wu, Claude Code’s product lead, announced that she is leaving Anthropic, making Friday her final day at the company. She said that two years ago, when Claude Code was still a small internal tool, she had it build an RL environment and run for dozens of hours without manually writing a single line of code, after which she joined the Claude Code team.
Wu said that when Claude Code first launched, most users manually approved permission prompts one by one and mainly used Claude to answer questions and write tests. Users now generally trust Claude’s judgment by default and approve automatically, while Claude can take on large migrations and new product development. Wu described helping turn Claude Code into an industry-defining product as a once-in-a-lifetime journey and thanked the teams and users behind Claude Code, Cowork, and Tag.
OpenAI hires Uber president as chief corporate affairs officer
行业动态
OpenAI announced that Jill Hazelbaker, Uber’s former president and chief corporate affairs officer, will join the company as chief corporate affairs officer. She will oversee global communications and corporate positioning while working to deepen public understanding of AI and trust in OpenAI’s mission.
OpenAI announced that Jill Hazelbaker will join the company as chief corporate affairs officer, with responsibility for global communications and corporate positioning. According to the company, she will work closely with business leaders across OpenAI to deepen public understanding of AI and strengthen trust in its mission.
Hazelbaker previously served as Uber’s president and chief corporate affairs officer. During her ten years at Uber, she led teams responsible for global communications, marketing, public policy, safety, and human resources.
OpenAI hires former U.S. Secretary of the Army as national security adviser
行业动态
OpenAI announced that former U.S. Secretary of the Army Dan Driscoll will serve as its newly created Senior Advisor for National Security Modernization, advising on the use of advanced AI for national and cyber security and helping advance related modernization efforts.
OpenAI announced that Dan Driscoll has taken the newly created role of Senior Advisor for National Security Modernization. His responsibilities include advising OpenAI on ways to use advanced AI to strengthen national security and cyber security, as well as helping the company’s internal teams understand the operational needs of national security institutions worldwide.
According to OpenAI, Driscoll will work with the company’s leadership and leaders from government, industry, academia, and allied institutions to advance modernization initiatives that support AI development aligned with democratic principles. The source does not provide an effective date for the appointment.
OpenAI plans a security sprint to increase bug bounties
行业动态
OpenAI plans a new internal security sprint with Cathedral Systems to use its own models to find system weaknesses and strengthen defenses. It is also raising bug bounties, with critical vulnerabilities paying more than four times the previous amount and top awards reaching $172,000.
OpenAI plans to conduct a new internal security sprint with Cathedral Systems, with researchers from both organizations testing the systems and searching for weaknesses. According to OpenAI, the effort will use its own models to assess system defenses, while the company also raises its bug bounty standards. Rewards for critical vulnerabilities will exceed four times the previous amount, and especially strong findings can receive up to $172,000 including the bounty. The new standards apply to reports currently awaiting bounty decisions. For vulnerabilities that received a bounty within the past 30 days and qualify for a higher reward under the new standards, OpenAI will pay the difference.
OpenAI says ChatGPT’s search volume is about 20% of Google’s
行业动态
Adam Fry, OpenAI’s head of consumer products, said in a Big Technology interview that ChatGPT triggers about 20% as many web-search queries as Google, while its weekly active users have reached 1.2 billion.
Adam Fry, OpenAI’s head of consumer products, said in an interview with Big Technology that the volume of web-search queries triggered by ChatGPT is about 20% of Google’s search-query volume. According to his explanation, the figure counts one query each time a user’s question triggers an online search, and the share is growing rapidly. ChatGPT’s weekly active users rose from 1 billion to 1.2 billion over two months.
Fry also disclosed that ChatGPT’s voice-input feature has 250 million weekly active users, including dictation and voice conversations. Users generate about 1.5 billion images each week. He said OpenAI plans to merge ChatGPT’s Chat and Work modes so that a single interface can automatically determine the user’s intent.
fal raises $1.65 billion, with Bloomberg reporting a $16 billion valuation
行业动态
Generative media company fal announced a combined $1.65 billion Series E and Series F financing. Bessemer Venture Partners led Series F, NVIDIA continued to participate, and Bloomberg reported that fal was valued at $16 billion.
fal announced that it closed its Series E earlier this year and subsequently closed and disclosed its Series F financing. The two rounds total $1.65 billion. Sequoia Capital led Series E, while Bessemer Venture Partners led Series F, with continued investment from NVIDIA. According to Bloomberg, fal’s valuation is $16 billion.
According to fal, the new funding will support work across the fal ecosystem in 2027. This includes expanding the inference capacity required to run image, video, and audio models, as well as pursuing new research and new forms of creation. The company said entertainment studios, advertising agencies, e-commerce brands, game developers, and marketing teams are integrating models, including fal’s own, into the pipelines that produce their final work.
Shanghai court sentences man to three years for stealing compute resources from a large AI model
行业动态
A Shanghai court sentenced Zong Moumou to three years in prison, suspended for three years, and fined him RMB 40,000 for making more than 290,000 unauthorized AI-platform API calls and earning over RMB 34,000 by reselling AI generation services.
On November 2, 2026, the Xuhui District People’s Procuratorate in Shanghai announced that Zong Moumou had been convicted of illegally obtaining computer information system data and sentenced to three years in prison, suspended for three years, plus a RMB 40,000 fine. Starting in December 2025, he used technical means to obtain a non-public interface belonging to an AI model aggregation platform, built a relay website, and sold low-cost image and video generation services. The services consumed upstream model compute paid for by the platform. By the time the case was uncovered, the interface had been called more than 290,000 times without authorization, generating over RMB 34,000 in illegal proceeds.
Zong first used AI to write an automated registration script, bought large numbers of email accounts, and registered nearly 300,000 free accounts in batches to collect promotional credits and create an account pool. After the platform introduced IP-based registration limits, he used AI tools to bypass the blocks. When the platform added electronic-signature verification in January 2026, he created an automated script to forge signatures. The company found that more than 4 million users had registered on two applications from the same IP address within a short period and reported the case in March 2026. The Xuhui District People’s Procuratorate said the platform had never offered paid API access to the public and that its user agreement prohibited automated scraping and verification bypasses. Bypassing safeguards without authorization to obtain compute, data, or generated content for profit cannot be excused as merely “搬运接口,” or moving an interface elsewhere.
Industry insiders say DeepSeek V5 is planned to be pretrained on Huawei Ascend
前瞻与传闻
Multiple people in the industry supply chain said that DeepSeek plans to use Huawei Ascend chips for pretraining its next-generation large model, V5. Ascend chips have previously been used mainly for inference tasks involving DeepSeek’s large models, and the new arrangement remains at the planning stage.
According to a report by The Science and Technology Innovation Board Daily, multiple industry supply-chain sources said DeepSeek plans to use Huawei Ascend chips to pretrain its next-generation large model, V5. The plan concerns the model-training stage, which differs from the main role Ascend chips have previously played in work related to DeepSeek’s large models. The sources said Ascend had mainly been used for inference tasks, while the V5 plan would extend its use to pretraining. The report did not disclose the specific chip models, timeline, or scale of the V5 pretraining plan.
Tencent Hy’s official account posted a Hy4 teaser video on X, ending with “Tencent Hy 4,” but the post did not disclose a specific release date, model specifications, or further product details.
The official Tencent Hy account (@TencentHunyuan) posted a video on X teasing Hy4. The video ends with the words “Tencent Hy 4,” while the post carries the caption “4 Real Productivity.” It does not provide a specific release date or disclose any model specifications. An accompanying notice states that the content was created with AI assistance and may contain hallucinations or errors.
