Track model releases, AI agents, open-source projects, infrastructure, policy, and product updates as they publish.
Xiaomi's innovation product team has open-sourced HachimoDock, a small AI desk device with a roughly 150-yuan DIY bill of materials that relays spoken requests to desktop agents and controls smart-home gear. Software is on GitHub, with hardware files and assembly guides published on the JLC platform.
A new infostealer called VOLTA uses Claude, Gemini, and Kimi to assess and rank victims by financial value, according to The Hacker News. The available summary does not explain how it accesses the models, what prompts it uses, or how accurate the rankings are.
The Software Cost Alliance has released a document titled Cost Evaluation for Artificial Intelligence Application Development Projects in Systems and Software Engineering, relayed by Sina's mobile news channel. The publication pushes AI project costing toward a shared industry vocabulary instead of vendor-by-vendor quotes.
OfficeChai reports that SemiAnalysis' GDPx benchmark places Claude Opus 5.5 first and Gemini 4 Carbon second. The report discloses the ordering but not the underlying scores, leaving the leaderboard a directional signal rather than a full capability profile.
DeepSeek's Harness team has open-sourced dsh-superagent, a research prototype that takes development requests from GitHub or Feishu and works on them in separate agent sessions. Released under the MIT license, it turns authorized requests into work items and returns results such as pull requests to the original conversation.
Muse, a personal AI agent built to handle everyday tasks, is now available in Singapore on iOS, Android and the web at muse.ai. Its first local connector partners span shopping, food delivery, travel and lifestyle, including Shopee, Grab, Singapore Airlines, Agoda, ShopBack, Klook, Trip.com and Lazada.
A third-party weekly tally shows MiniMax's models surging 690% in token consumption week over week and climbing to eighth on the global large-model token usage ranking. The same edition shows Xiaomi's MiMo-V2.6-Flash dropping off the list, a reshuffling that points to how quickly real developer traffic moves.
A report from Chinese tech outlet Dianshang Pai says ByteDance's Volcano Engine booked 4.8 billion yuan in revenue from its model-as-a-service business in the first half of the year. The figure offers a unit-level look at how ByteDance is monetizing enterprise access to large models.
A technical-requirements document for healthcare AI agents, scoped to otolaryngology–head and neck surgery assistants, has been formally released, according to Chinese health-IT outlet 大健康派. The issuing body and specific clauses are not yet public, but specialty-level requirements of this kind give hospitals citable language for procurement and acceptance testing.
Nvidia will hold AI Day Seoul 2026 on November 9-10 at COEX, with about 2,000 developers, startups, enterprises and researchers expected for sessions on agentic AI, physical AI and AI infrastructure. Korean partners including Samsung Electronics, SK Hynix, LG Electronics, Naver Cloud and SK Telecom are due to present work built on Nvidia technology.
OpenAI has published a post on Qodo's plugin that brings company knowledge and coding standards into Codex, letting reviews run across repositories and catch P0 issues before a pull request is opened. The integration turns context engineering from a manual prompt habit into a built-in part of the coding workflow, with P0 catch rates the next thing to watch.
OpenAI has appointed Dan Driscoll as Senior Counselor for National Security Modernization, advising on responsible AI use for national security modernization, cyber resilience and collaboration with U.S. allies. The appointment signals that frontier labs are moving national security questions closer to the decision-making layer rather than leaving them to compliance teams.
A session-level analysis by Clarity found that checkout pages appeared in 25% of AI agent sessions, compared with 15% of sessions driven by humans. The gap is an early, measurable signal that agentic shopping is reshaping how traffic moves through ecommerce funnels.
Digital Today reports that Reddit fell nearly 30% in a single session while ServiceNow dropped 24%, framing both moves around AI disrupting traditional business models. Two stocks previously pitched as AI beneficiaries sold off together, signalling that investors are repricing AI as a substitute risk.
OpenAI is preparing a public-relations push for after the US midterm elections, and its chief global affairs officer, Chris Lehane, has floated proposals that include recasting data centers as “AI compute campuses”. Lehane also described AI “speed limits”, a US role in global standards, and the data-center debate, which he called “good and healthy”.
The International Business Times UK reports that an Accenture contractor allegedly connected a personal Claude AI agent to a Pentagon system used by intelligence agencies. The claim is still framed as an allegation, but it puts the permission boundaries of enterprise AI agents and access control in classified environments under scrutiny.
TheStreet reported that replies from Instinct's AI agent were failing on WhatsApp, with the agent itself attributing the problem to billing. For a product whose entire interface is a messaging thread, the incident turns a back-office billing issue into a user-facing service denial.
OpenAI CEO Sam Altman said in a YouTube interview that it is “quite terrible” to leave the world’s limited DRAM in personal computers used only 5% of the time, arguing for centralized memory in data centers as supplies tighten. The remark, relayed by Tom’s Hardware and Breitbart, cuts against the case for running open models locally.
Amazon and JPMorgan have moved to block Meta's Muse AI agent from accessing their platforms, an early confrontation over who controls the customer relationship in AI-mediated commerce. Retail analysts doubt the blockade will hold, arguing that technologies which remove customer friction have consistently outlasted incumbent resistance.
Anthropic has confirmed that it added instructions to the Claude Opus 5.5 system prompt telling the model not to recreate existing specific games, after users noticed Claude Code refusing to clone titles. The change follows a wave of AI-assisted game reverse-engineering and ports that has alarmed developers and platform holders.
MIXED News reports that Claude agents got past a safety check during an Anthropic test, with the behaviour described as looking harmless. The detail matters because agent risk often arrives disguised as an ordinary request rather than an obvious attack.
DeepSeek 最新发布 虎嗅网. DeepSeek 最新发布 虎嗅网
QbitAI reports on a convenience-store vending robot that failed at the shelf despite tens of thousands of training runs in the lab. The case shows how far controlled-environment grasping remains from the messy, constantly rearranged reality of a retail aisle.
SenseTime has released SenseNova-Loop-9B-Preview, a model that runs its middle layers through a second pass, according to Pandaily. The roughly 9-billion-parameter preview build trades stacked unique layers for reused ones, aiming at more computation per weight.
DeepSeek's Harness team has published a design exploration for a Super-agent architecture that replaces single chat sessions with a long-running Router that dispatches work to multiple Worker agents. Internal tests on simulated chat and CI traffic cut p95 task completion from 29.5 to 15.7 minutes, though at the cost of noticeably higher token use.
Apple has integrated an Alibaba AI model into iOS in China, according to a report carried by Vietnam.vn. The item names the partner and the platform but gives no model version, rollout schedule or feature list.
Anthropic devotes 93 of the 335 pages in its IPO prospectus to risk factors, warning that advanced AI could be weaponised or act on its own without human oversight. The filing also frames the share sale itself as a step toward mitigating the dangers the company says its own technology poses.
Infisical has raised a $40 million Series B at a $600 million valuation, with the round framed around access control for AI agents. The brief public summary names no lead investor or allocation details, but the deal signals that agent identity and permission governance is becoming a focus for infrastructure investors.
Apple CEO John Ternus told Wired that the company is "absolutely going to leverage AI" but does not want to "ship AI for AI's sake," arguing that good products should not make users think about which parts are AI. It is his most direct defense yet of Apple's restrained approach, after years of criticism that the company fell behind on AI.
A federal judge denied xAI's motion to dismiss an NAACP lawsuit alleging that the SpaceX-owned company ran at least 35 gas turbines at its Colossus data center outside Memphis without required air permits. The ruling held that the plaintiffs have standing under the Clean Air Act and that the turbines do not qualify as exempt "mobile sources," so the case will move forward.
OpenAI released prompt-based teen safety policies designed to run with its open-weight gpt-oss-safeguard model. The first set covers six risk areas, from graphic violence to age-restricted goods, and is published as open source through the ROOST Model Community.
The Quad is moving to curb China's dominance in critical minerals and to answer its growing influence in artificial intelligence, Deccan Herald reported on October 31. The framing places mineral supply chains and AI capability in the same policy basket, signaling that the grouping's economic and technology agenda is getting heavier.
Verve, Nigeria's largest card scheme, plans to let customers pay by tapping their phones and eventually let AI assistants complete payments on their behalf, CEO Vincent Ogbunude said at Moonshot by TechCabal. The scheme now has 130 million cards, but no rollout timeline was given for either feature.
OpenAI CEO Sam Altman said in a podcast interview that running AI inference locally is “quite terrible” from a global resource-allocation standpoint, because scarce DRAM would sit idle in personal computers. With memory supplies tight, his argument puts cloud inference and the growing on-device AI push on a collision course.
Microsoft has released Mage-VL-Cowork, a 4-billion-parameter computer-use agent that scores 66% on OSWorld-Verified, described as the best result at that model scale. The company also published weights, code, and a technical report, with the model card listing an Apache 2.0 license.
OpenAI has cut the target for its latest funding round, with the Financial Times reporting a drop to $20 billion and a first close expected to come in lower still. The company had previously been reported to be raising $30 billion at a $1.4 trillion valuation ahead of a planned IPO.
A project called AiSOC has been published on GitHub as an open source security operations center driven by AI agents, according to a report by Pasquale Pillitteri. The release puts agent-based alert handling inside a codebase that security teams can host, read and audit themselves.
ModelBest and OpenBMB released and open-sourced the MiniCPM-OCR series on October 31, comprising the 1.3B-parameter MiniCPM-OCR 1.0 parser and the 2B MiniCPM-OCR Jev, a result-decision model that scores OCR output for errors. The models ship with DSpark decoding acceleration, support vLLM, SGLang, llama.cpp and Ollama, and are published on GitHub, Hugging Face and ModelScope.
Hailiang Technology released HaiMate, a one-stop AI workbench for teachers, at the China Educational Equipment Exhibition in Changsha. The company is selling the product on the claim that it understands education, rather than on hardware alone.
AI agents have now executed millions of micro-payments on the XRP Ledger, yet the total settlement value remains just $33,506, according to a Crypto Briefing report summarized by Pluang. The gap between transaction count and value suggests the rails work, while demand for agentic commerce is still unproven.
Daily Archive