Daily AI Digest

2026-10-08

Source:橘鸦 AI 早报 · 34 items

2026-10-08
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

Anthropic releases Claude Haiku 5.5, cutting average costs by about 75%

要闻

Anthropic releases Claude Haiku 5.5, cutting average costs by about 75%

Anthropic has released Claude Haiku 5.5, saying its average running cost is around 75% lower than Haiku 4.5’s. Alongside the release, it halved Sonnet 5.5’s cache reads pricing and introduced a monthly API credit for Claude Max and Team subscribers.

With this release, Anthropic introduced Claude Haiku 5.5 and changed Sonnet 5.5 pricing and subscribers’ API benefits. According to the company, Haiku 5.5 costs around 75% less to run on average than Haiku 4.5; halving Sonnet 5.5’s cache reads pricing reduces its running cost by around 20% for most agentic work. The new monthly API credit is for Claude Max and Team subscribers and is intended to support building agents and applications on the Claude Platform. The source does not specify the credit amount.

According to Anthropic, Haiku 5.5 is designed for high-volume, cost-sensitive tasks and can handle summaries, context compaction, database queries, and classification requests. It can also work as a subagent alongside Opus 5.5 and Sonnet 5.5 on coding tasks. The company describes it as its fastest model to date and says it suits speed-sensitive uses such as live customer support and browser use. It is also the first Haiku-series model with an adjustable effort setting, allowing users to choose whether to prioritize optimizing for cost or intelligence.

Read original →

OpenAI rolls out GPT-6 to all users with an intelligent interaction interface

要闻

OpenAI rolls out GPT-6 to all users with an intelligent interaction interface

OpenAI is rolling out GPT-6 and Intelligent UI in ChatGPT, with paid users receiving the update on launch day and Free and Go users following the next day. According to its evaluations, Instant began answering 44% earlier in web-search scenarios.

OpenAI began rolling out GPT-6 and Intelligent UI to ChatGPT Plus, Pro, Business, and Enterprise users on launch day, with Free and Go users receiving the update the following day; the source does not specify a calendar date. The release changes only the Chat experience, leaving the models used by Work and Codex unchanged. Paid plans use GPT-6 Sol, while Free and Go use GPT-6 Luna. According to OpenAI, GPT-6 can progressively output answers while continuing to think. In its internal evaluations, GPT-6 Extra High took about as long to begin answering as GPT-5.6 Medium, while scoring higher overall than GPT-5.6 Extra High. For queries involving web search, GPT-6 Instant began answering an average of 44% earlier than GPT-5.6 Instant.

The interaction updates include combining text, images, charts, buttons, forms, and interactive components according to the question, as well as directly generating task tools such as calculators and mini-games. OpenAI says Intelligent UI uses native components and a compiler that support streaming output, allowing the interface to build progressively alongside the answer rather than waiting for the entire response to finish. Model safety training has also been updated; according to the company, training against high-risk misuse, including cyberattacks, biological threats, and violence, has been strengthened.

Read original →

Codex reset usage limits based on votes yesterday and is offering additional manual resets today

要闻

Codex reset usage limits based on votes yesterday and is offering additional manual resets today

OpenAI Codex lead Tibo announced that all paid accounts will receive one additional usage-limit reset that can be saved for later, expected by the end of the campaign’s third day in Pacific Time. He had already confirmed an earlier reset after 76% of voters supported it.

OpenAI Codex lead Tibo announced on day three of the 28-day update campaign that all paid accounts would receive one additional usage-limit reset that can be saved for later, with delivery expected by the end of that day in Pacific Time. According to his explanation, the offer celebrates GPT-6 arriving in Chat and the combined active user count for Codex and ChatGPT Work reaching a new high of 40 million. Earlier, after releasing four updates on day two of the campaign, Tibo held a vote on whether to reset usage limits, with 76% of voters choosing a reset. He subsequently confirmed that the reset had been carried out and that all four improvements remained in place.

Read original →

Claude Sonnet 5.5 cuts cache read prices in half

要闻

Claude Sonnet 5.5 cuts cache read prices in half

Anthropic has cut Claude Sonnet 5.5’s cache-read price by 50% to $0.10 per million tokens. The company estimates that this change reduces running costs by approximately 20% for most long-running tasks.

Anthropic announced that it is halving Claude Sonnet 5.5’s cache-read price, bringing it to $0.10 per million tokens. The announced price change applies to cache reads, and the company also provided an estimate of its impact on task running costs: Sonnet 5.5 is expected to cost approximately 20% less to run for most long-running tasks. This cost reduction is the company’s estimate and applies to most long-running tasks, rather than all tasks.

Read original →

Claude grants monthly API credits to Max and Team subscribers

要闻

Claude grants monthly API credits to Max and Team subscribers

Claude’s Max and Team subscriptions include monthly API credits that users can claim by linking a Claude Console organization, without adding a payment method. Team credits are pooled across all seats, with a monthly cap of $500 USD.

Claude now provides monthly Claude API credits with Max and Team subscriptions, which users must link to a Claude Console organization to claim. According to the official documentation, credits are issued each billing cycle, or monthly for annual subscriptions, and can be used to build and run users’ own applications and Agents without adding a payment method to Claude Platform. Team allocations depend on the seat configuration at the start of each billing month, including the first allocation: 3 Standard seats and 2 Premium seats receive $260 USD per month, and adding 1 Premium seat brings the following month’s credits to $360 USD. The credits can only be used on Claude Platform and do not change the rules for rate limits or spend limits. After claiming them, users can check Promotional credits under Settings > Billing in the Console, then create an API key to send requests.

Read original →

谷歌’s SynthID detection website opens to users worldwide

要闻

谷歌’s SynthID detection website opens to users worldwide

Google has opened SynthID Detector to users worldwide after previously limiting access to media professionals. It is currently available only in English and lets users upload images, videos or audio to check for invisible watermarks, though detection results are not 100% accurate.

Google announced that its AI-generated content detection platform, SynthID Detector, is now available to users worldwide, with an English version currently offered; the early version was restricted to media professionals. The platform accepts image, video and audio uploads and checks whether files contain invisible SynthID watermarks from Google or partners including OpenAI, NVIDIA and Kakao, with Apple also set to add support soon. According to Google, since SynthID launched in 2023, it has watermarked more than 180 billion images and videos and 240,000 years of audio content; verification features built into Search, the Gemini app and Chrome currently process more than 1 million requests per day. Watermarks remain identifiable after edits such as adding filters, but the tool’s detection results are not 100% accurate.

Read original →

A Claude Code team member launches html-plan skill for public testing

开发生态

A Claude Code team member launches html-plan skill for public testing

Anthropic Claude Code team member Thariq has made the html-plan skill available to try through the community plugin marketplace. It generates HTML plan pages containing code snippets, questions needing clarification, and mockups, with user feedback being collected ahead of a wider rollout.

Anthropic Claude Code team member Thariq announced on X that his html-plan skill is now available to install through the community plugin marketplace. Users first run `claude plugin marketplace add anthropics/claude-plugins-community` to add the marketplace, then run `claude plugin install html-plan@claude-community` to install it. According to his explanation, the skill presents plans in concise language alongside code snippets, questions needing clarification, and mockups; an accompanying lint mechanism is intended to reduce common Claude failures, with some rules targeting chart and flowchart generation. Thariq said he particularly likes the call-stack display and code-snippet annotation features, noting that the call-stack display draws on developer @dillon_mulroy’s approach. He wants to collect user feedback before a wider rollout and turn it into a plugin that replaces Claude Code’s default planning mode with HTML plans.

Read original →

Claude SDK now includes built-in computer use and browser use toolsets

开发生态

Claude SDK now includes built-in computer use and browser use toolsets

Anthropic has provided minimal Python and TypeScript examples of Claude’s computer toolset on GitHub, using VNC to execute desktop actions through a driver that implements 10 tools. The examples demonstrate the interface and are not production code.

Anthropic has provided a minimal VNC example of computer_toolset_20260801 in its claude-quickstarts repository on GitHub, with Python and TypeScript implementations. Developers declare one toolset entry in tools, and the model issues desktop actions such as screenshot, left_click, and type as separate tool calls that the driver executes. The driver consists of a few hundred lines, controls one desktop through VNC’s RFB protocol, and implements 10 tools covering screenshots, clicks, mouse movement, scrolling, typing, key input, and waiting. The SDK provides tool schemas, rejection handling for unimplemented tools, and the tool runner. According to the official explanation, the code demonstrates the interface and is not production code.

Execution safety and call verification are also part of the example. According to the official explanation, screen content influences the model’s next actions, and a focused terminal executes what the model types, so the execution environment should be a disposable desktop inside a sandbox. The VNC server has no password, connections must be restricted to localhost, and container ports must be published only to the host’s loopback interface. Before executing any action other than screenshot, run uses confirm to require the user to enter y; exercise instead approves calls automatically and requires no API key. exercise checks 6 valid calls and verifies that both a left_click one pixel outside the screen and the unimplemented zoom return is_error. It exits with a non-zero status if any result differs from the expected outcome.

Read original →

OpenRouter launches the full range of ElevenLabs voice models at half price for a limited time

开发生态

OpenRouter launches the full range of ElevenLabs voice models at half price for a limited time

OpenRouter launched nine ElevenLabs Text to Speech models and two Speech to Text models on October 7, 2026. All users can access these models at 50% of the platform’s list price during a limited-time promotion.

OpenRouter announced the addition of nine ElevenLabs Text to Speech models and two Speech to Text models on October 7, 2026, with a 50% discount off list prices for all OpenRouter users running through October 19, 2026, at 8 a.m. Pacific Time. According to the company, developers do not need a separate ElevenLabs plan and can call POST /api/v1/audio/speech and POST /api/v1/audio/transcriptions using an OpenRouter API key. The launch does not include Realtime Scribe v2 over WebSocket. Users with an existing ElevenLabs account can add their own key through BYOK to access their own cloned voices, pronunciation dictionaries, and higher-quality formats such as mp3_44100_192.

Capabilities and request limits vary by model. According to the company, Eleven v4 and v3 can interpret audio tags as instructions for tone and delivery, with per-request limits of 10,000 and 5,000 characters respectively; Flash v2.5 supports up to 40,000 characters per request. Scribe v2 can produce transcripts with timestamps and speaker labels, while Scribe v2 Medical is fine-tuned for medical terminology. Transcription uploads are capped at 25 MB, equivalent to about 27 minutes of 128 kbps MP3 audio. Providing a public URL bypasses that upload cap, but long files may still reach the 180-second upstream request timeout. Text to Speech is billed by input Unicode code points, with audio tags included in the character count. Scribe is billed per second of audio, and usage.cost in the transcription response reports the cost of the request.

Read original →

GitHub Copilot opens local sandboxes, with hybrid reasoning previewing at the end of the month

开发生态

GitHub Copilot opens local sandboxes, with hybrid reasoning previewing at the end of the month

GitHub Copilot will preview automatic routing between local and cloud inference by the end of the month, while also allowing developers to select local models directly. Microsoft says the quantized on-device version of MAI Code 1.1 Flash is 53GB, an 80% reduction in size compared with the Bfloat16 cloud version.

GitHub Copilot will offer automatic routing between local and cloud inference by the end of the month; the source does not specify the month or year. GitHub Copilot CLI, the Copilot app, and VS Code will offer two ways to use local models: letting Auto decide whether a task uses local or cloud inference, or having developers explicitly select a local model. According to the official explanation, Auto considers task context and cache state during multi-turn sessions, retaining reusable cached work as it switches models. Direct selection supports MAI Code 1.1 Flash through the Windows ML provider, as well as connections to OpenAI-compatible local endpoints and selection of the models they expose. Execution permissions are separately constrained by sandboxing. Copilot uses Microsoft Execution Containers (MXC), an open-source library from the Windows team, to control access to files, networks, credentials, system capabilities, and other resources by processes and local services launched by an Agent.

The local model was developed by Microsoft AI using a mixture-of-experts architecture, with 137 billion total parameters and 6.8 billion active parameters, and uses quantization and speculative decoding. Microsoft says the quantized on-device version is 53GB, an 80% reduction in size compared with the Bfloat16 cloud version. Its published figures for the first shipping version on Surface Laptop Ultra show peak memory usage of 75.5GB at 256k context, with prompt-processing throughput of 923.5 and 769.8 tokens per second at 64k and 128k context, respectively. The device is built around NVIDIA RTX Spark, with up to 128 GB of unified memory and up to 1 petaflop of AI compute. The source also explains that the operating system, applications, inference runtime, and key-value cache require memory, and that keeping a model loaded does not guarantee constant response times.

Read original →

Microsoft MXC is now generally available, defining execution permissions for AI agents

开发生态

Microsoft MXC is now generally available, defining execution permissions for AI agents

Microsoft announced general availability of Microsoft Execution Containers (MXC) on October 7, 2026, enabling developers and IT administrators to set policy-based limits on Agent access to files and networks. Windows 365 support is also generally available.

Microsoft announced general availability of Microsoft Execution Containers (MXC) and its Windows 365 support on October 7, 2026. MXC targets untrusted code and dynamically generated workloads and can contain model-generated output, plugins, tools, an Agent harness, or an entire Agent. According to Microsoft, developers and IT administrators can define resources that a workload may access, including files and network destinations, and MXC enforces those policies at runtime using the appropriate container. The policies remain outside the Agent workload’s control, so an Agent or generated code cannot grant itself additional access.

Developers integrate MXC through a unified JSON configuration schema and a multi-language SDK, with MXC mapping the required controls to backends on Windows, macOS, or Linux. Lightweight containment uses AppContainer, Seatbelt, and Bubblewrap, respectively. The session container and WSL container support Windows 11 only, while MicroVM supports Windows 11 and Linux and remains experimental. Microsoft also said Windows will support distinguishing Agent activity from user activity through Microsoft Entra and extend Microsoft Agent 365 controls to local Agents on devices. These capabilities are intended to let IT teams manage MXC containers, apply policies, and monitor Agent activity, but the source provides no specific launch dates for these forthcoming features.

Read original →

Liquid AI open-sources decision models d1-3B and d1-omni-600M

模型发布

Liquid AI open-sources decision models d1-3B and d1-omni-600M

Liquid AI has released the open-weight decision models d1-3B and d1-omni-600M on Hugging Face. The company reports scores of 48.57 and 15.95, respectively, on the public split of Decision Index v0.2.1.

Liquid AI has released two open-weight decision models, d1-3B and d1-omni-600M, on Hugging Face, with llama.cpp support available at launch. According to the company, d1 returns an answer in a single forward pass rather than generating output tokens. The base model for d1-3B was created by averaging the weights of LFM2.5-2.6B and the text backbone of LFM2.5-VL-3B, followed by fine-tuning and merging. d1-omni-600M is an experimental checkpoint based on LFM2.5-Encoder-350M, with audio and vision capabilities added in stages. It supports text, image, and audio inputs and remains under development.

The company reports that d1-3B scored 48.57 on the public split of Decision Index v0.2.1, matching Decider 35B-A3B, while d1-omni-600M scored 15.95. Across seven public text benchmarks, their mean scores were 82.9 and 78.4, respectively. The company measured d1-3B's input-to-output latency with one request at a time: a single question took 8 ms on an NVIDIA GeForce RTX 4090, 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and 50 ms on a Jetson Orin Nano. This release did not report results for the private vision split of Decision Index v0.3 or inference latency data for d1-omni-600M.

Read original →

Perplexity releases two multimodal embedding models

模型发布

Perplexity releases two multimodal embedding models

Perplexity has made the 0.6B and 9B multimodal embedding models in its pplx-embed-v2-late series available on Hugging Face. They support retrieval of text, images, and visual documents and share an embedding space, allowing queries across models’ indexes.

Perplexity has released the pplx-embed-v2-late series, comprising 0.6B and 9B multimodal late-interaction embedding models, now available on Hugging Face. Both are built on Qwen3.5, output a 128-dimensional vector for each token, and use MaxSim for scoring. They support retrieval of text, images, and visual documents, with no OCR required to process PDFs, slides, or scanned documents. Both were distilled from the same internal 18B ColBERT teacher model and share an embedding space, so the 0.6B model can directly query an index built by the 9B model. According to the official evaluation, the 9B model achieved a Recall@1000 of 74.8% on Q2D-Web, compared with 69.3% for nemotron-embed-8b.

Read original →

生数科技 releases flagship video model Vidu Q4 Preview

模型发布

生数科技 releases flagship video model Vidu Q4 Preview

Shengshu Technology has released its flagship audio and video model Vidu Q4 Preview through its web product and API, with launch promotional pricing starting at $0.014 per second. The model supports up to 15 image references, 3 audio references, and 2K and 4K output.

Shengshu Technology launched Vidu Q4 Preview, the first public preview of its next-generation flagship audio and video model, on October 7, 2026, making it available through Vidu’s web product and API platform that day. Launch promotional pricing starts at $0.014 per second, with output resolutions including 540p, 720p, 1080p, 2K and 4K, alongside support for up to 15 image references, 3 audio references and 10-bit color depth. According to the company, comparable output specifications and billing conditions allow up to 5 times as much output for the same budget. Final pricing, supported resolutions, feature availability and usage terms may vary by plan and region.

According to the company, the model focuses on improving coordination among character expressions, emotions, body movements and voices, as well as between camera movements, editing and on-screen action, while making effects such as explosions, fireworks and particles fit more naturally into scenes. Its target users include independent creators, small studios and production teams working on narrative and commercial projects. The preview precedes the full Q4 model and lets creators use it in real projects. Feedback on performance, creative control and everyday production needs will inform subsequent improvements.

Read original →

Musk: Grok Bot will no longer use only its own models

产品应用

Musk: Grok Bot will no longer use only its own models

Musk announced that SpaceXAI’s Grok Bot will move beyond in-house models and select backends by task, with access to third-party APIs including Claude Opus 5.5, Midjourney and Suno. Models will also be assigned according to question complexity.

Musk announced on X that SpaceXAI’s Grok Bot will select backend models by task going forward, rather than remain limited to its in-house Grok series; the source did not specify an effective date. According to his explanation, available third-party APIs include Claude Opus 5.5, Midjourney and Suno, with selection based on which model is most likely to deliver the best result for the user. He subsequently added that simple questions will go to small, fast models, while only questions requiring complex answers will be assigned to large models.

Read original →

Grok Bot 0.68.1 released with slide creation and formatted email sending

产品应用

Grok Bot 0.68.1 released with slide creation and formatted email sending

SpaceXAI has released Grok Bot 0.68.1 with slide creation and formatted email sending. It can deliver presentations as PowerPoint files or Google Slides, and the screen resolution used for computer operation has increased from 1280×800 to 1920×1200.

SpaceXAI has released Grok Bot 0.68.1 with updates to slides, email, and computer operation. Users can work with the Bot to create slides, choose from examples it presents, and receive the results as PowerPoint files or Google Slides. Email features support sending formatted messages containing headings, bold text, and lists from draft cards, as well as searching Gmail’s spam folder and moving messages back to the inbox. According to the company, the Bot operates computers faster, and the screen resolution it uses has changed from 1280×800 to 1920×1200. In one-on-one chats, user messages take on the Bot’s color. This feature requires Grok Bot 0.64.0 or later and can be disabled by setting the accent color to black in settings.

Read original →

Grok Bot can now search and read X, available to everyone with no configuration required

产品应用

Grok Bot can now search and read X, available to everyone with no configuration required

Grok Bot announced that all users can now search, read, and monitor X content without setting up an X connector. A team member also demonstrated a workflow that routes feature requests and bug reports from product feedback to development tools.

Grok Bot announced that its X content search, reading, and monitoring capabilities are now available to all users, with no X connector setup required. For handling product feedback, team member lauren demonstrated a workflow in which Grok Bot monitors user feedback about her product on X, sends feature requests to an issue tracker, and passes bug reports to Cursor cloud agent or a project for classification and fixes. Other uses listed by the official team include following breaking news and compiling weekly summaries of industry developments.

Read original →

Grok team member: The @bot mention feature is now live on X

产品应用

Grok team member: The @bot mention feature is now live on X

SpaceXAI Grok Bot team member Larsen Cundric said X has launched @bot mentions, allowing instructions in posts to be sent to a user’s Grok Bot. Users must first link their X account to a grok.com account with bots enabled.

SpaceXAI Grok Bot team member Larsen Cundric said the @bot mention feature on X is now live. According to his explanation, users must first link their X account to a grok.com account with bots enabled. They can then mention @bot and include instructions in replies, posts, or quote posts, sending tasks directly to their Grok Bot. Example instructions include adding content to a Notion reading list, setting a reminder to read it later, summarizing a post thread, and drafting a reply.

Read original →

OpenAI will bring a college application planning tool to ChatGPT for Teens

产品应用

OpenAI will bring a college application planning tool to ChatGPT for Teens

OpenAI announced that College Planner is coming to ChatGPT for Teens, which applies automatically to accounts belonging to users under 18. The tool will consolidate college application and financial aid tasks, initially serving US students in grades 10–12 who plan to attend four-year colleges.

OpenAI outlined updates to ChatGPT for Teens, with the college application planning tool College Planner set to launch; no specific launch date was announced. The tool will bring school application requirements, deadlines, tasks, and financial aid application steps into a plan that can be continually updated. It will initially serve US high school students in grades 10–12 preparing to apply to four-year colleges. According to the company, coverage will later expand to more countries and to institutions including two-year colleges and technical colleges. ChatGPT for Teens applies automatically to accounts belonging to users under 18.

Among the learning features, iOS now supports photographing multiple pages in succession and combining them into a PDF, while the Android version remains in development. Flashcards and easier-to-create quizzes have also been introduced. OpenAI said teenagers spend an average of less than 15 minutes per day using the service, with nearly 1.2 million teenagers using learning visualization features and more than 180,000 using study mode within one week. OpenAI will also support College Advising Corps in expanding to additional states and establish an AI scholarship program. It will support the student advisory council at Boston Children’s Hospital’s Digital Wellness Lab over the next three years.

Read original →

Codex Cloud supports Tailscale from launch

产品应用

Codex Cloud supports Tailscale from launch

OpenAI launched Codex Cloud with support for Tailscale, currently its only VPN provider, without a joint integration project between the companies. The feature lets cloud-based Codex access resources inside a tailnet, but supports only specific connection methods.

OpenAI included Tailscale support when Codex Cloud launched and is gradually making the product available to paid ChatGPT accounts on Plus and higher tiers. Codex Cloud works on GitHub repositories in an environment that is not tied to a particular device; connecting Tailscale also enables access to internal APIs and testing services in a tailnet. According to Tailscale, there was no partnership agreement, custom API, or integration project between the companies, and OpenAI used the same Tailscale available to all users. It is currently the only VPN provider supported by Codex Cloud.

Configuration is available under Advanced > VPN in the cloud environment, and users must generate an auth key in the Tailscale admin console. OpenAI’s instructions require enabling Reusable and Ephemeral and tagging the key so that Codex VMs join the network with tag:codex and receive the corresponding rules rather than inheriting the permissions of the user who generated the key. Supported destinations currently include HTTP, HTTPS, and arbitrary TCP through HTTP CONNECT on proxy:8088. Private IPv4 subnet routes are supported, but require IPv4 addresses or hostnames that resolve to IPv4. MagicDNS, split DNS, UDP, ICMP, inbound connections, SSH, and Tailscale SSH are not supported. Tailscale says that, by default, Codex can access nearly all resources in a tailnet.

Read original →

Google launches experimental game-generation platform Playground

产品应用

Google launches experimental game-generation platform Playground

Google has launched Playground, an experimental game-generation platform, for U.S. users aged 18 and older, enabling game creation, testing and sharing through text. Creation access will roll out in tiers based on Google AI subscriptions; Unity Spark integration is not yet available.

Google has launched Playground at playground.google for U.S. users aged 18 and older, with creation access rolling out in tiers based on Google AI subscriptions. According to Google, the experimental platform requires no coding experience: users can enter text through a conversational interface, start with a blank project, starter prompts or guided support, and immediately test their games. They can also adjust physics, rules, characters and environments. The platform runs in a browser, allowing users to play community games on a phone or laptop. Games can remain private, be shared through a link or be published to the Playground Explore gallery.

Community features include in-game leaderboards and multiplayer gameplay in select genres. A Play Games profile lets users set a custom handle, like games, compete in rankings and follow creators. Google says the gallery recommends games based on player ratings and play activity, and every published game undergoes safety screening aligned with its Community Guidelines, alongside user reporting. Playground will later integrate with Unity Spark, which Google says provides professional-level game mechanics, high-fidelity 3D capabilities and the flexibility of the Unity runtime. Unity Spark is currently in testing, with a closed beta coming soon.

Read original →

谷歌 introduces AI Edge Foresight, an on-device AI meeting assistant for Mac

产品应用

谷歌 introduces AI Edge Foresight, an on-device AI meeting assistant for Mac

Google has launched AI Edge Foresight, an on-device meeting assistant for Mac that uses EmbeddingGemma 2 and Gemma 4 for meeting and file retrieval. EmbeddingGemma 2 has 740M parameters and, according to Google, maps text, images, video frames, and audio into a unified vector space.

Google has launched Google AI Edge Foresight, an experimental meeting assistant for Mac; the source does not specify an exact launch date. According to Google, the app runs on EmbeddingGemma 2 and Gemma 4, handling note-taking assistance and the indexing and retrieval of conversation transcripts and private files locally on the device. It connects directly to system audio and the microphone, works with any meeting platform, and can operate entirely offline. Google says Foresight maps text, audio, and visual data into a unified vector space to search a personal knowledge library, with sensitive data remaining on the device. The open-weight EmbeddingGemma 2 model, released alongside it by Google DeepMind, has 740M parameters. According to the company, active RAM requirements on a Google Pixel 11 Pro can be as low as approximately 191MB for text-only weights and approximately 567MB for the full multimodal model.

Google AI Edge Gallery has added two showcases using the same model: Instant Media Search and Video Moments Finder. The former searches local images and videos using natural language or example images, stores media vectors in a local SQLite database, and returns results based on cosine similarity. The latter locates visual moments in local videos without transcribing audio or generating intermediate text descriptions. Google says it will provide the model as a service on Android through ML Kit in the coming weeks, supporting NPU acceleration on devices where available and automatic model updates. MediaPipe Tasks is adding EmbeddingGemma 2 support to Embedder and Semantic Retriever to abstract media preprocessing and data transformations. MediaPipe Decision Task can use the model to classify images, text, or audio on-device without fine-tuning.

Read original →

DGX Station will support Windows, enabling trillion-parameter-class models to run locally

产品应用

DGX Station will support Windows, enabling trillion-parameter-class models to run locally

NVIDIA and Microsoft previewed DGX Station with Windows support. According to the companies, its 748GB of memory and up to 20 petaFLOPS of FP4 AI compute enable local execution of trillion-parameter-scale models, allowing developers to handle large AI workloads within Windows.

NVIDIA and Microsoft previewed DGX Station for Windows at the Windows AI and Surface event in San Francisco, bringing the previously Linux-based DGX Station into the Windows environment. The system uses the GB300 Grace Blackwell Ultra Desktop Superchip and provides 748GB of coherent memory and up to 20 petaFLOPS of FP4 AI compute; according to the companies, this configuration can run models of up to a trillion parameters locally. They say developers and researchers can fine-tune and run inference on large models on their primary machine, while running always-on AI Agents that connect directly to existing Windows applications and infrastructure. Linux AI toolchains remain available through WSL when needed. The previous deployment arrangement required enterprise developers to maintain separate environments: Linux for heavy AI workloads and Windows for productivity tools, applications and workflows.

Read original →

Surface Laptop Ultra preorders open, starting at $2,599

产品应用

Surface Laptop Ultra preorders open, starting at $2,599

Microsoft has opened pre-orders for Surface Laptop Ultra and Surface RTX Spark Dev Box, with the laptop starting at $2599. According to the company, both devices use NVIDIA RTX Spark and support running AI workloads locally.

Microsoft opened pre-orders for Surface Laptop Ultra and Surface RTX Spark Dev Box on October 7, 2026, with the former starting at $2599. Both devices are built around NVIDIA RTX Spark Superchip and offer up to 128 GB of unified memory. Laptop Ultra includes an NVIDIA Blackwell RTX GPU with up to 6,144 cores and an NVIDIA Grace CPU with up to 20 cores. According to Microsoft, the laptop can run AI models exceeding 120B parameters locally, delivers up to 1 petaflop of AI performance, and supports Agent coding tasks using local models with GitHub Copilot. It has a 15-inch PixelSense Ultra touchscreen and user-removable storage. Designed for desktop development, Dev Box belongs to the Project Zenith device family and comes with preinstalled tools and environments including Visual Studio Code, Git, GitHub CLI, GitHub Copilot, WSL, Python, and Node.

Read original →

Copilot on Windows will be able to read local context and perform computer tasks on users’ behalf

产品应用

Copilot on Windows will be able to read local context and perform computer tasks on users’ behalf

Microsoft announced on October 7, 2026, that MXC is generally available on Windows 11 and that local models are being integrated into Windows hybrid AI. HydraFusion-powered GitHub Copilot features will enter experimental preview later that month.

Microsoft announced on October 7, 2026, that Microsoft Execution Containers (MXC) is generally available on Windows 11. Hybrid AI features powered by HydraFusion will enter experimental preview later in October 2026 across the GitHub Copilot app, GitHub Copilot CLI, and Visual Studio Code. According to Microsoft, MXC lets organizations specify which files and networks an Agent can access and enforces those policies at runtime. Windows offers containment options ranging from process and session isolation to WSLc, virtual machines, and Windows 365 for Agents. Products already supporting MXC include OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA’s OpenShell, and Unsloth AI. Microsoft also announced llama.cpp support in Windows ML.

For local models, Microsoft is bringing MAI Code 1.1 Flash to devices. The model has 137 billion total parameters and 6.8 billion active parameters. Microsoft says 3-bit precision reduces its size by nearly 80% while preserving coding quality and supporting a 256K context window locally. Microsoft also listed an upcoming NVIDIA Nemotron model with over 70 billion parameters, using just over 20GB of memory with 2-bit quantization, alongside the 284-billion-parameter DeepSeek V4 Flash and other models running locally on RTX Spark. According to the company, HydraFusion will expand from routing tasks in the cloud to accessing models running on devices. Microsoft also announced that Copilot on Copilot+ PCs will gain Windows hybrid AI support across the Home, Code, and Autopilot experiences.

Read original →

Talk to it like a colleague: A Claude Code member shares prompting insights

技术与洞察

Talk to it like a colleague: A Claude Code member shares prompting insights

Anthropic Claude Code team member Boris Cherny shared his prompting approach, emphasizing clear tasks, effort levels, and verification methods. He said a few short prompts to Opus 5.5 produced 16 fix PRs during formal verification with Lean.

Anthropic Claude Code team member Boris Cherny shared how he prompts Claude, saying the current priority is to explain what to do, how much effort to invest, and how to verify the result. He compared this interaction to communicating with a colleague, arguing that most tasks do not require overly structured or exhaustive prompts and that Claude can handle the work once given a goal. By contrast, he said prompt wording had a greater effect on results in the Sonnet 3.5 era. As a practical example, he described using Opus 5.5 to formally verify the Claude Agent SDK with Lean, saying a few short prompts produced 16 PRs fixing bugs and race conditions. In replies, he also stated that he does not use a custom CLAUDE.md and sets the reasoning effort for Opus 5.5 to medium by default.

Read original →

Artificial Analysis releases a leaderboard for music model evaluations

技术与洞察

Artificial Analysis releases a leaderboard for music model evaluations

Artificial Analysis has published music-generation model leaderboards, separating instrumental and vocal generation into two v1.1 benchmarks. Scores rely exclusively on blind preference votes from recruited panelists and include 95% confidence intervals; public votes do not count toward the rankings.

Artificial Analysis has published the AA-Music-Instrumental v1.1 and AA-Music-Vocal v1.1 leaderboards for music-generation models; the source does not specify a publication date. Each benchmark provides an overall ranking that can be filtered by genre. According to the organization, scores come from blind pairwise comparisons by recruited panelists: each panelist listens to two tracks generated from the same prompt and modality, then selects the preferred track. Public votes are excluded from the calculation. Scores are calculated using Bradley-Terry Maximum Likelihood Estimation and then converted to an Elo-like scale for readability. Each score includes a 95% confidence interval, which narrows as the model receives more votes.

According to the organization, the instrumental benchmark includes only models whose endpoints offer controls for generating music without vocals; endpoints that cannot separate instrumental from vocal generation are evaluated only on the vocal benchmark. The vocal benchmark requires models to generate both lyrics and singing from a prompt, without lyrics supplied by the evaluators. Before tracks enter the Arena, their loudness and delivery format are standardized, and generation settings remain fixed within each model. When a specified default setting is unsupported, the closest supported value or the provider’s default is used. Prompts come from a curated, reproducible set and also include reviewed user submissions of 25–200 characters. The organization states that it receives no compensation from providers for inclusion or favorable evaluation outcomes.

Read original →

Unsloth: Train Decision models locally with 4GB of VRAM

技术与洞察

Unsloth: Train Decision models locally with 4GB of VRAM

Unsloth has published a method for training LLMs as Decision models that score supplied options and return a choice with a probability. The company reports that training with a Clef head and LoRA for 1 epoch raised downstream accuracy from 30–37% to as high as 78%.

Unsloth has documented how to train and connect Decision models, supporting LLM fine-tuning through its Desktop app or code, as well as further fine-tuning of Laya and Clef. These models do not generate text: they read the input, questions and options once, then use a small head with the same design as Cloudflare’s Clef to score options and handle all questions together. The Desktop app supports macOS, Windows and Linux. Once trained, a model can be connected to the Decision API, where requests specifying laya, default or jev-latest are routed to that model.

The company reports that training with a Clef head and LoRA rank 64 for 1 epoch raised downstream accuracy from 30–37% to as high as 78%. Training combined 12 sources with typed-decisions, and the test set was decontaminated against the training data. It contained 3,000 rows: 2,000 from typed-decisions and 500 each from BANKING77 and CLINC150. The documentation separately gives Desktop app settings of 2 epochs, LoRA rank 16 and a learning rate of 2e-4, and states that the app evaluates accuracy on held-out decisions before and after training and calibrates confidence. The company also reports that Qwen3.5-4B reached 76% accuracy after 10 minutes of training on an L4.

Read original →

South Korea announces a $3.5 billion frontier AI project to rival China’s open-source models

行业动态

South Korea plans to support domestic frontier AI model development with 4.7 trillion won (about $3.48 billion) in government equity investment, aiming to compete with China's leading open-source models. The funding will enter the 2027 budget proposal and, according to media reports, still requires parliamentary approval; participating companies must provide matching private funding.

South Korea's government announced plans to launch a state-backed frontier AI project in 2027, using government equity investment to support domestic frontier model development. The government said the project aims to compete with China's leading open-source models, with an investment of 4.7 trillion won (about $3.48 billion), approximately nine times the 530 billion won previously pledged to five companies. According to media reports, the funding will be included in the 2027 budget proposal and still requires parliamentary approval. The anticipated allocation includes 3.9 trillion won to purchase 10,000 Nvidia Vera Rubin GPUs and 800 billion won to purchase training data, with participating companies required to provide corresponding private funding. The government will select companies through an open application and review process and is considering a special-purpose company to hold the investment so that returns can be shared with the public. The existing national foundation model project competition will continue but shift toward industry-specific models; the two finalist teams selected at the end of 2026 may choose to join the frontier AI project or focus on industry applications.

Read original →

Biohub, the U.S. Department of Energy, and NIH invest $1.8 billion to expand an AI biological data initiative

行业动态

Biohub, the U.S. Department of Energy, and NIH invest $1.8 billion to expand an AI biological data initiative

Biohub, the U.S. Department of Energy, NIH, and new partners announced an expansion of the Virtual Biology Initiative, committing a combined $1.8 billion in funding, data, computation, and measurement technology resources to build open, AI-ready biological data for the research community.

Biohub, the U.S. Department of Energy (DOE), the National Institutes of Health (NIH), and new funding partners announced the expansion of the Virtual Biology Initiative on October 7, 2026, committing a combined $1.8 billion in funding, data, computation, and new measurement technology resources. Announced in April 2026, the initiative aims to organize data generation across institutions and disciplines to establish an open data foundation for predictive AI models of biology. Not all of the resources represent new funding: DOE will invest more than $500 million over five years, while NIH will coordinate contributions of relevant datasets, repositories, and knowledge bases developed through more than $500 million in previous federal investment. Biohub will work with NIH to standardize these data for AI model training.

Google DeepMind, Isomorphic Labs, and Meta are collectively investing $300 million to develop the technologies and multimodal datasets needed for predictive models of biology. Biohub's initial commitment totals $500 million, with $400 million supporting new measurement technologies, including cryo-electron tomography, imaging of cells in living tissue, and engineering tools for building and perturbing biological systems at molecular, cellular, tissue, and whole-organism levels. The remaining $100 million funds research outside Biohub. According to the official announcement, the initiative will expand data on responses to interventions across different cell types and conditions, and develop and validate technologies for studying cells and their interactions, enabling researchers to use AI models to pose, predict, and answer biological questions digitally.

Read original →

Google signs a 20-year nuclear power purchase agreement with Constellation Energy

行业动态

Google signs a 20-year nuclear power purchase agreement with Constellation Energy

Google and Constellation Energy have signed a 20-year power purchase agreement to add 890 MW to the PJM grid through upgrades to existing nuclear units, with the first capacity expansion expected to be completed in 2028.

Google and Constellation Energy have signed a 20-year power purchase agreement, with the first nuclear capacity expansion expected to be completed in 2028. Constellation will invest more than $4.3 billion to upgrade 11 existing nuclear units, adding 890 MW of nuclear capacity to the PJM grid. The companies also have a separate 15-year energy supply agreement covering 2,700 MW. According to Google's explanation cited by Reuters, this supply is not tied to a specific generation source. The agreements include a demand response mechanism to reduce non-essential electricity use when the grid is under strain. Beyond the power supply arrangements, the companies have also expanded their five-year technology collaboration. Constellation will adopt Google Cloud and Gemini Enterprise to support capacity delivery, generation scheduling and critical infrastructure protection.

Read original →

Nous Research raises $90 million in Series B funding

行业动态

Nous Research raises $90 million in Series B funding

Nous Research has raised $90 million in Series B funding, with investors including NVIDIA, Microsoft’s M12, and Samsung. The company says it will launch Hermes for Businesses, allowing companies to choose models and control their own intelligence stacks.

Nous Research announced $90 million in Series B funding in an announcement without a publication date and said it would develop and deliver Hermes for Businesses. Investors include NVIDIA, Microsoft’s M12, Samsung, Robot Ventures, Union Square Ventures, Y Combinator, and Menlo Ventures. According to the company, the business product will allow customers to choose models to control costs and support private, safe scaling of AI use and management of organizational knowledge. The company positions the plan as extending AI users’ control to businesses.

Nous Research was founded in 2023. Its early work included fine-tuning Meta’s Llama models around users’ viewpoints and publishing YaRN to extend the context windows of open models. The company developed an Agent stack in 2025 and subsequently released Hermes Agent under an MIT license, but the source does not provide a determinable full release date. According to the company, users can choose the runtime environment and models, while the Agent can remember projects, generate reusable skills, and retain that knowledge across sessions. The company says Hermes has been cloned over 24 million times. Its internal estimates put Hermes at approximately 2.5% of global token usage, with thousands of contributors participating in development and thousands more building projects on top of it.

Read original →

Report: SpaceX is negotiating to borrow $40 billion to purchase Nvidia chips

前瞻与传闻

Report: SpaceX is negotiating to borrow $40 billion to purchase Nvidia chips

SpaceX is in talks to borrow $40 billion to buy Nvidia chips, Bloomberg reported. The Financial Times said the financing is expected to be completed in 2027, but people familiar with the matter said discussions remain at an early stage and the deal may not go ahead.

Elon Musk’s SpaceX is discussing $40 billion in debt financing with banks and investors to purchase Nvidia chips, Bloomberg reported, citing people familiar with the matter; no deal has yet been reached. On the financing structure, the Financial Times previously reported that the plan comprises approximately $10 billion in bank loans and approximately $30 billion in investment-grade bonds, led by Apollo Global Management, with completion expected in 2027. PIMCO is among the investment managers assessing the transaction. People familiar with the matter said talks remain at an early stage and may ultimately produce no deal; if completed, the financing would rank among the largest debt financings ever for AI computing infrastructure.

Read original →

16-year-old boy stranded after following Claude’s directions downhill is rescued by a search team

其他

16-year-old boy stranded after following Claude’s directions downhill is rescued by a search team

CTV News reported that a 16-year-old in Canada left a trail while following Claude’s directions downhill and became stranded beneath a steep rock face on Crown Mountain. With a helicopter unable to approach, two rescuers rappelled down and moved him before extracting him.

According to CTV News, North Shore Rescue personnel rescued a 16-year-old who became stranded after leaving a trail while following Claude’s directions downhill in British Columbia, Canada. He had taken a gondola to the summit of Grouse Mountain, at more than 4000 feet, before becoming trapped beneath a steep rock face nicknamed “Widowmaker” on Crown Mountain and calling for help. Search and rescue manager Paul Markey said a fall at that location could cause serious injury or death. Because the steep, vertical rock face prevented a helicopter from approaching, two rescuers rappelled down, established anchors and moved the teenager to a location where he could be hoisted out. Markey urged the public not to rely on chatbots such as Claude for this type of information, but to research routes themselves and supplement that preparation with training and navigation.

Read original →