Daily AI Digest

2026-11-21

Source:橘鸦 AI 早报 · 24 items

2026-11-21
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

Mathematician claims to have proved the Riemann hypothesis with Claude

要闻

Mathematician Levent Alpöge claims to have found a proof of the Riemann hypothesis with Claude and has released Lean code in 14 files. He says the full proof manuscript exceeds 250,000 pages; it has not been independently verified.

Mathematician Levent Alpöge posted a statement on his personal website claiming to have proved the Riemann hypothesis in collaboration with Claude, but the proof has not yet been independently verified, and the source gives no publication date for the statement. According to his account, the argument has been formalized in Lean code without additional assumptions; the released code comprises 14 files, and the fully written proof manuscript exceeds 250,000 pages. He considers it inappropriate to publish an extremely long document that nobody could read through and plans to write a conventional paper after fully understanding and simplifying the proof. He also says the method can be applied to the generalized Riemann hypothesis and expects that, with more time, it could address the extended Riemann hypothesis as well.

Read original →

Thinking Machines releases and open-sources Inkling 2 Mini

要闻

Thinking Machines releases and open-sources Inkling 2 Mini

Thinking Machines has released Inkling 2 Mini, an open-weights model with 26.4B total parameters and an approximately 20GB memory footprint, with inference and customization available on Tinker. The company says its looping middle layers enable reasoning with fewer tokens.

Thinking Machines has released the weights for Inkling 2 Mini and made inference and customization available on Tinker, although the source does not specify a release date. The model uses a Mixture-of-Experts architecture with 26.4B total parameters, 3.4B active parameters, and 5.4B weight activations per token when looping is included. Its 24 layers are divided into four blocks of six layers each. The two middle blocks each run three times, while the first and last blocks each run once. According to the company, most weights are trained and deployed natively in NVFP4, with the final four layers using MXFP8. Its total memory footprint is approximately 20GB, allowing it to run on local devices such as consumer laptops.

The company reports that Inkling 2 Mini scores 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, ahead of all open-weights models with fewer than 100B total parameters on those two tests. It also released a variable-loop version for research. According to the company, increasing the number of loops improved performance on some long-context tests but did not improve results on knowledge-intensive tests such as AA-Omniscience. The model has limitations in multi-turn conversations and instructions with numerous constraints: it scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing Qwen3.8-27B on both.

Read original →

Antigravity expands cache pool, with quotas to reset once complete

要闻

Antigravity expands cache pool, with quotas to reset once complete

Antigravity is expanding its cache pool. Lead Varun Mohan says Neon’s cache hit rate is expected to improve by about 4%, enabling at least 20% more work within subscription allowances; subscriber usage will be reset after the adjustment is complete.

Antigravity is expanding its cache pool and will reset subscriber usage once the adjustment is complete; the source provides no completion date. Lead Varun Mohan says the expansion is intended to improve Neon’s cache hit rate, with an expected increase of about 4%. According to Mohan, the adjustment is expected to let users complete at least 20% more work within their subscription allowances; that figure is a projected effect, not a published measurement. The usage reset is being made because allowances had been consumed quickly, and it is scheduled to take place after the expansion is complete.

Read original →

OpenAI opens Space collaboration workspace to Plus users

要闻

OpenAI opens Space collaboration workspace to Plus users

OpenAI is expanding ChatGPT Space collaboration to Plus subscribers paying $20 per month, with a phased rollout to users aged 18 and over worldwide. Free users still cannot open shared content.

OpenAI announced that ChatGPT Space is rolling out in phases to Plus users aged 18 and over worldwide; the source did not specify a date for this expansion. The Plus plan costs $20 per month and lets users create spaces, invite others to co-edit, and work with ChatGPT to produce interactive pages, slides, and websites. Access to shared content remains restricted: recipients must sign in to a paid account with Space access, and free users cannot open it. Space was previously introduced at OpenAI Dev Day in September and was limited to paid plans such as Pro, Business, and Enterprise. According to the company, this update also adds mobile page editing, a redesigned home page, and faster loading, while pages can link to connected tools to update automatically.

Read original →

腾讯混元 open-sources flagship Hy4 model with 770B total parameters

模型发布

腾讯混元 open-sources flagship Hy4 model with 770B total parameters

Tencent Hunyuan has released the weights for its flagship Hy4 model, whose language backbone has 770B total parameters and activates 49B per token, with a 1M context window and support for understanding text, images and video. Available versions include Hy4, Hy4-FP8 and Hy4 base.

Tencent Hunyuan has made the weights for its flagship Hy4 model available for download, offering Hy4, Hy4-FP8 and the pretrained Hy4 base through GitHub, Hugging Face, ModelScope and GitCode, with a GGUF version also available. According to the company, Hy4 uses a natively multimodal MoE architecture, with a language backbone containing 770B total parameters and activating 49B per token. It supports a 1M context window and can understand text, images and video. Tencent Hunyuan also released the Hy4 System Card with an accompanying safety evaluation report.

Read original →

微信 open-sources omni-modal model family WeMM-Video

模型发布

微信 open-sources omni-modal model family WeMM-Video

WeChat has open-sourced WeMM-Video on GitHub, a Chinese–English omni-modal model family with 2B, 4B, 9B, and 35B-A3B variants that can combine video frames with original audio for understanding and reasoning, returning text responses.

WeChat has open-sourced WeMM-Video on GitHub, providing Chinese–English bilingual models for joint audio-visual understanding and reasoning. The family accepts text, images, audio, and video and produces text, with 2B, 4B, and 9B dense models and a 35B-A3B mixture-of-experts model. According to the project, 35B-A3B sets an open-source performance benchmark for omni-modal video understanding, surpassing Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 while narrowing the gap with Gemini-3.0-Flash. The project also reports leading general audio understanding performance for 35B-A3B and 9B.

Runtime requirements include Python 3.12, an NVIDIA GPU, a CUDA 12.9-compatible driver, and FFmpeg. The project provides deployment options using vLLM and SGLang, with each request supporting up to four images, two videos, and two audio items within the configured context length. Both patched backends accept video_url through POST /v1/chat/completions, with use_audio controlling whether the original soundtrack is included. The source code uses Apache License 2.0; the model license covers the WeMM-Video collection as of November 20, 2026, with each checkpoint’s license specified in its Hugging Face model card.

Read original →

Codex updates conversation threads for smoother switching and loading

开发生态

Codex updates conversation threads for smoother switching and loading

Codex has updated its conversation threads, changing state transitions, conversation loading, and animated icons. According to a team member, the update reduces redundant information in conversation histories and can improve conversation loading speed by up to 60%–70% when the app starts cold.

Codex has updated its conversation threads with changes to state transitions and conversation loading, as well as new animated icons; the source does not specify an effective date. According to a Codex team member, the changes reduce redundant information in conversation histories and make transitions between states smoother. The loading-performance figures provided by the team member apply to app cold starts: conversation loading speed can improve by up to 60%–70%. New animated icons have also been added to conversation threads.

Read original →

Codex raises WebSocket request limit to 128 MB

开发生态

Codex raises WebSocket request limit to 128 MB

OpenAI team member Charlie Marsh said Codex has raised its WebSocket limit for model requests from 16 MB to 128 MB, reducing the overall HTTP fallback rate by about 20%. The server also now immediately notifies clients when requests exceed the limit.

OpenAI team member Charlie Marsh posted that Codex has increased the WebSocket size limit for model requests from 16 MB to 128 MB and changed how oversized requests are handled. According to his explanation, clients previously often uploaded most of the content before receiving a server rejection, then retried and remained on HTTP, which does not support incremental sending. The server now immediately notifies clients when a request exceeds the limit, eliminating wasted uploads and retries and making processing about 5 times as fast. Marsh said the higher limit keeps large turns from falling back to HTTP, reducing the overall HTTP fallback rate by about 20%; that metric includes all fallback causes, including connection failures and timeouts.

Read original →

OpenAI API enables cross-turn reasoning reuse for GPT-6.2

开发生态

OpenAI API enables cross-turn reasoning reuse for GPT-6.2

OpenAI’s API documentation says prefix changes cause newer frontier models to discard subsequent reasoning items, potentially increasing token usage, cost, and latency. It also describes how to investigate these changes using cache diagnostics and removal records.

OpenAI states in its API documentation that newer frontier models remove reasoning items following a conversation prefix change as a safeguard; the documentation provides no effective date. According to the company, a prefix is the conversation history preceding a reasoning item, and shortening tool results, rewriting developer messages, or replacing older turns with summaries can change it. Removal may force the model to repeat reasoning, increasing token usage, cost, and latency; output quality may also decline if the lost reasoning cannot be recovered. OpenAI recommends an append-only integration pattern wherever possible and explains that, because caching matches requests from the beginning, prefix changes also affect subsequent cache reuse.

For investigation, the documentation identifies two information sources. Prompt Cache Diagnostics can compare a request with an earlier response by setting prompt_cache_options.comparison_response_id to that response’s ID, allowing inspection of the diagnostic reason and cached token count. OpenAI says the tool reports the first identified difference on a best-effort basis and cannot confirm whether the model reused reasoning. Meanwhile, input_transformations in a Responses API response records removed reasoning items and the reasons for removal; for prefix_mismatch, the id can be matched against an earlier response’s output. The documentation also notes that changing tools, text.format, instructions on pre-GPT-6.2 models, or earlier input items can unintentionally change the prefix.

Read original →

Google opens preview of audio editing in the Gemini API

开发生态

Google opens preview of audio editing in the Gemini API

Google has opened Gemini 3.5 Transcribe’s voice editing capability in public preview through the Gemini API, supporting voice-driven rewriting and translation of existing text with up to 64 KiB of source text per request.

Google has launched a public preview of Gemini 3.5 Transcribe’s voice editing capability through the Gemini API in Google AI Studio and Gemini for business. According to Google, developers can use it to build experiences for editing, rewriting, and translating existing text by voice. The capability is enabled by default in Smart transcription mode. For rewriting, developers can supply up to 64 KiB of existing text through source_text, and the model returns updated content based on spoken instructions. Google also says the same technology is used in Gboard’s Rambler and the Gemini App for macOS.

Read original →

Google launches early Black Friday deals with up to 50% off AI subscriptions

开发生态

Google launches early Black Friday deals with up to 50% off AI subscriptions

Google has opened early Black Friday offers to existing AI and storage subscribers, with up to 50% off annual AI Plus and Pro plans through November 26, local time. The discount depends on the subscriber’s current plan.

Google had launched its early Black Friday promotion for existing AI and storage subscribers by the time Google Gemini subscription product lead Vikas Kansal published his post, offering up to 50% off annual Google AI Plus and Pro plans. The offer runs through November 26, local time; the source does not specify the year or the exact launch date. According to Kansal, the discount varies by the subscriber’s current plan, and a subscription provides access to models including Gemini 4 Hydrogen and Gemini 4 Neon, along with additional storage. Users without a subscription must wait until Black Friday for their offers to begin. Addressing the Ultra plan, Kansal said it already provides substantial value and that further subsidies would not be appropriate.

Read original →

小米 MiMo ultra-fast model comes to TokenPlan for a limited time

开发生态

小米 MiMo ultra-fast model comes to TokenPlan for a limited time

Xiaomi MiMo has added MiMo-V2.6-Pro-UltraSpeed to TokenPlan’s individual and team Pro and Max plans for a limited time, with no additional activation required for new or existing subscribers. Uncached input and output Token usage deducts 3000 and 6000 Credits, respectively.

On November 21, 2026, Xiaomi MiMo announced that its ultra-high-speed inference model MiMo-V2.6-Pro-UltraSpeed was available on TokenPlan for a limited time starting that day, covering the Pro and Max plans for both individuals and teams. According to the company, access extends to new and existing subscribers without additional activation. For billing, uncached input Token usage deducts 3000 Credits, while output Token usage deducts 6000 Credits.

Read original →

OpenAI launches CarPlay Dot calling feature

产品应用

OpenAI launches CarPlay Dot calling feature

OpenAI developer experience lead Romain Huet said users can now call their own Dot through Apple CarPlay and get things done by voice while driving. His post on X includes a demonstration video but does not specify a launch date.

Apple CarPlay now supports calling a user’s own Dot, according to a post on X by OpenAI developer experience lead Romain Huet; the post does not specify a launch date. His description presents voice interaction while driving as a use case, saying users can accomplish tasks by speaking rather than limiting the interaction to the call itself. Huet attached a demonstration video and invited users to try the feature on their next drive to see what they could accomplish using voice alone.

Read original →

Google Gemini introduces GIF generation

产品应用

Google Gemini introduces GIF generation

Google Gemini has launched GIF generation for Google AI subscribers, with free availability in 8 countries, including Japan. According to its official help documentation, the output is a looping animation of approximately 3 seconds that can be downloaded or copied into other apps.

Google Gemini has launched GIF generation for Google AI subscribers and currently offers it free in the United States, Japan, South Korea, Australia, Canada, the United Kingdom, Germany, and France. Powered by the Gemini Omni 1.2 Flash video model, the feature can create looping photo effects, humorous memes, and other content. Users can enter a description directly in the main chat, or select the image or video tool and then choose GIF in the format selector. According to Gemini’s help documentation, the output is a looping animation of approximately 3 seconds that can be downloaded in GIF format or copied and pasted into documents, emails, messages, and social apps.

Read original →

Claude App begins rolling out a revamped mobile experience

产品应用

Claude App begins rolling out a revamped mobile experience

Anthropic designer Daniel Destefanis said Claude App’s new mobile experience has begun rolling out, with changes covering code, projects and artifacts, photo attachment, and app icons. The Android version will also roll out soon.

Anthropic designer Daniel Destefanis said Claude App has begun rolling out its new mobile experience and that the Android version will follow soon, without specifying a date. According to his explanation, the update builds on the team’s desktop work and aims to improve mobile support for code, projects and artifacts, making it easier to start conversations or resume existing ones while away from a desk. The changes also cover details across the app, including the ability to quickly add photos by long-pressing the plus button and an overhaul of the app icons.

Read original →

Meta’s personal AI agent Muse introduces email functionality

产品应用

Meta’s personal AI agent Muse introduces email functionality

Meta has introduced an email feature for its personal AI agent Muse, with the linked Help Center page titled around Muse email addresses. The supplied material consists mainly of page-loading code and provides no verifiable launch date, pricing, or feature restrictions.

Meta has introduced an email feature for its personal AI agent Muse, but the supplied material does not specify a launch date. The associated Meta Help Center page concerns Muse email addresses, although the excerpt contains a page title and page-loading code rather than the feature documentation itself. The material therefore supports only the email feature described in the title; it does not establish how addresses are created, whether email can be sent or received, which regions are covered, what usage conditions apply, or what the pricing is. It also provides no key product figures to report.

Read original →

腾讯 WorkBuddy adds computer-control capabilities

产品应用

腾讯 WorkBuddy adds computer-control capabilities

Tencent WorkBuddy has launched computer-control capabilities in public beta for macOS and Windows. Users on version 5.11.0 or later can describe tasks in chat after granting permissions, without requiring the target software to expose an interface.

Tencent WorkBuddy has launched its Computer Use feature, now in public beta for macOS and Windows, with client version 5.11.0 or later required. Before using it, users must enable the feature and grant two system permissions: accessibility and screen recording. Tasks are submitted through the chat box, and execution progress can be viewed in a picture-in-picture window on the side. According to the company, the feature can recognize the current page, understand client interfaces, and perform actions such as clicking, typing, and scrolling in the background without the target software exposing an interface.

Read original →

OpenAI says new defenses cut reasoning theft success rate to 0.2%

技术与洞察

OpenAI says new defenses cut reasoning theft success rate to 0.2%

OpenAI reported an average attack success rate of 0.2% for extracting protected reasoning from GPT‑6.2 Astra with updated safeguards in an internal adaptive evaluation, compared with 54.0% for GPT‑6 Astra with its original safeguards.

OpenAI said it has strengthened defenses against protected reasoning extraction in GPT‑6.2 Astra and GPT‑6.2 Sol. Its internal adaptive evaluation used automated red-teaming to test whether attackers could recover internal reasoning steps generated by a model while answering an earlier prompt. The company reported an average attack success rate of 54.0% for GPT‑6 Astra with its original safeguards and 0.2% for GPT‑6.2 Astra with updated safeguards. The source article did not provide a corresponding result for GPT‑6.2 Sol.

According to OpenAI, its defenses operate at two levels: individual requests and coordinated campaigns. The former combines model behavior, classifiers, and API controls to limit reasoning extraction, while the latter analyzes request and account signals to identify coordinated extraction. In a separate internal study, a student trained on exact reasoning traces generally outperformed a student trained on reasoning summaries after supervised fine-tuning. After an additional lightweight round of reinforcement learning, the former still scored higher at the largest inference budget on both benchmarks shown, while the latter scored higher at lower budgets. OpenAI interprets these results as suggesting that protecting reasoning traces may limit the transfer of cybersecurity capabilities.

Read original →

OpenAI uses thousands of agents to complete months of analysis in 32 minutes

技术与洞察

OpenAI, PerturbAI, and Scripps Research used thousands of AI agents to analyze and interpret data from millions of mouse brain cells in 32 minutes. The team estimated that comparable human analysis and literature review would take months to a year, while the resulting mechanistic leads still require experimental validation.

OpenAI, PerturbAI, and Scripps Research completed a parallel analysis of existing experimental data from millions of mouse brain cells in 32 minutes using thousands of AI agents powered by GPT‑Rosalind; the source did not disclose the completion date. According to the official account, the team assigned gene perturbations from a whole-brain CRISPR atlas to individual agents, which analyzed cellular responses, searched the literature, and proposed explanations. Their reports were then combined into a reasoning atlas connecting genes, cellular effects, relevant papers, and possible mechanisms. The atlas is available for researchers to compare related genes, examine unexpected responses, and identify leads for follow-up experiments.

According to the official account, a controlled test that kept the model and resources fixed found that, at the largest workload, a single agent cited only 11 references, compared with more than 700 for distributed agents. Human experts also preferred the latter’s interpretations. The team estimated that comparable human analysis and literature review, performed one perturbation at a time, would take months to a year. Leads proposed by the agents included a potential connection between cholesterol regulation and neurodegeneration, as well as different effects of Tsc1 and Tsc2 perturbations in the cerebellum: Tsc2 disruption changed the activity of dozens of genes involved in neuronal communication, whereas Tsc1 disruption produced little change. These mechanistic explanations still require testing in follow-up experiments.

Read original →

OpenAI releases medicinal chemistry benchmark MedChemBench

技术与洞察

OpenAI releases medicinal chemistry benchmark MedChemBench

OpenAI released MedChemBench, a medicinal chemistry reasoning benchmark that evaluates models on 99 questions drawn from 117 papers and patents. Among 11 GPT models, GPT-6.2 Astra ranked first with a score of 84.5% at the xhigh reasoning setting.

OpenAI released the medicinal chemistry reasoning benchmark MedChemBench and published test results for 11 GPT models. GPT-6.2 Astra ranked first with a score of 84.5% at the xhigh reasoning setting. Performance differed across tasks: GPT-6 Astra scored 95.4% on molecular structure reconstruction, but only about half on compound selection and numerical prediction. The benchmark contains 99 questions drawn from 117 papers and patents, covering 761 molecular structures. Scoring spans four categories: structure reconstruction, scientific reasoning, compound ranking and selection, and numerical prediction. OpenAI said the tests showed progress in its frontier models’ ability to help scientists screen promising drug candidates.

Read original →

US House report: Cursor considers phasing out Chinese AI models

行业动态

A report by two U.S. House committees says Cursor, SpaceX’s AI coding platform, is considering phasing out Chinese-developed open-weight models over geopolitical risks. The report also recommends restricting Chinese AI models in federal systems and contract performance.

Two U.S. House committees released a report stating that Cursor, SpaceX’s AI coding platform, had told the committees it was considering phasing out Chinese-developed open-weight models because of geopolitical risks. According to the report, Cursor’s Composer 2 coding model is built on Moonshot AI’s Kimi K2.5. On policy, the report recommends establishing a U.S. national open-weight initiative. Building on an order barring DeepSeek from federal civilian agency systems, it also recommends prohibiting Chinese-developed AI models in federal information systems and in the performance of federal contracts.

Read original →

Isomorphic Labs teams up with MSD to develop cancer drugs

行业动态

Isomorphic Labs teams up with MSD to develop cancer drugs

Isomorphic Labs announced a strategic research collaboration with MSD, combining its AI drug discovery approach with MSD’s proprietary data and research experience. The partners aim to advance therapeutic candidates; the source did not disclose the collaboration’s financial terms or timeline.

Isomorphic Labs announced a strategic research collaboration with MSD, although the source did not specify an effective date. The collaboration will combine Isomorphic Labs’ AI drug discovery approach with MSD’s proprietary data resources and drug discovery experience. On the technology side, Isomorphic Labs has built IsoDDE, a drug design engine composed of foundational AI models that is not restricted to a particular drug modality. According to the company, this computational drug design system can provide the predictive accuracy needed to explore novel biological systems; the collaboration aims to combine AI, proprietary data and biopharmaceutical experience to advance therapeutic candidates for patient populations with unmet medical needs.

Read original →

Recently circulated research report says 智谱 will launch a 3-trillion-parameter flagship model

前瞻与传闻

Recently circulated research report says 智谱 will launch a 3-trillion-parameter flagship model

A circulating JPMorgan report says Zhipu’s management expects to launch GLM-5.5, a model with approximately 3 trillion parameters, in the fourth quarter of 2026, roughly four times the size of its existing base model. A separate model with approximately 700 billion parameters has no confirmed release schedule.

Zhipu’s management expects to launch its flagship GLM-5.5 model, with approximately 3 trillion parameters, in the fourth quarter of 2026, according to a circulating JPMorgan report. The report describes it as roughly four times the size of the existing 745-billion-parameter base model and says it is intended to return Zhipu to the global top tier. On development, a separate new model with approximately 700 billion parameters is undergoing post-training as a validation platform for larger-scale training. According to the report, its performance already exceeds GLM-5.3, but its release plans remain undecided. On the business side, the report says Zhipu reiterated its target of reaching $3 billion in ARR by year-end and assesses visibility into that target as relatively high. It also says most of the company’s compute capacity is located in China, limiting the short-term impact of new U.S. chip regulations.

Read original →

Switching harnesses with a Gemini subscription? Google team member responds

前瞻与传闻

Switching harnesses with a Gemini subscription? Google team member responds

Responding to a community proposal to make Gemini subscriptions usable in any harness, Google Gemini team member Logan Kilpatrick said a first step would come on December 2. He did not explain the specific plans, and the source did not specify a year.

Google Gemini team member Logan Kilpatrick said in a reply to community user Kun Chen that a first step would be taken on December 2; the source neither specified a year nor provided details of that step. The reply addressed a suggestion about how subscriptions could be used, rather than describing an already announced feature. After praising the experience of talking with Gemini 4 Neon, Kun Chen proposed allowing Gemini subscriptions to be used in any harness and said this would make returning easier. Kilpatrick responded on the same day the suggestion was posted but offered no further explanation.

Read original →