The 2028 U.S. presidential election is heating up early, with JD Vance reflecting on the Iran War and several Democratic governors in the spotlight.

Beating AI News: Moonshot AI officially launches Kimi Hosted Agents, packaging the entire AI Agent runtime framework into a cloud service. Previously, when enterprises called model APIs, they had to set up their own runtime environments, arrange tool calls, save task progress, and handle disconnections and failure retries. Now Kimi takes over these tasks: enterprises only need to configure tasks and integrate data to deploy Agents into their own business systems for continuous operation. For example, for fund reconciliation, Kimi provides a ready-made solution with three specialized Agents: one reads external statements, another verifies internal accounts, and the third independently reviews discrepancies to generate an anomaly report. Actual account modifications still require manual approval. Other templates cover investment research modeling, Hong Kong IPO tracking, regulatory enforcement daily reports, and more. The platform currently offers 12 Agent templates, over 140 Skills, and more than 30 plugins, and also supports enterprises integrating their own databases and tools. Anthropic launched Claude Managed Agents back in April this year, which also hosts Agent runtime environments, task states, and tool calls for developers. Google released the Antigravity Harness-based Managed Agents API in May, OpenAI launched its Codex-based Agents API, and Microsoft rolled out the Copilot Harness managed service at the end of October. All these products help developers avoid building their own Agent underlying frameworks, making it easier for AI to integrate into real business operations.

Insight Beating AI Flash News: Google announced that Gemini API’s Computer Use tool has concluded its preview phase and is now officially available. Compared to the previous preview version, there are no major changes to its features or interfaces; the primary improvement is enhanced stability for production environments. This tool enables AI to identify screen content via screenshots, then generate operation commands such as clicks, inputs, and scrolls. Developers’ programs execute these operations and send new screenshots back to the model, allowing the AI to continue completing tasks. Examples include automatically filling out web forms, testing website functionalities, or collecting information across multiple websites.

According to Bloomberg, Micron Technology has largely suspended plans to build a new memory chip plant in Taoyuan, Taiwan, due to ongoing labor disputes with employee unions. The project carries an investment budget of $3 billion to $4.5 billion and was originally slated to begin operations in 2028. Sources familiar with the matter noted Micron fears the prior union agreement could be overturned, with only minimal site security maintenance work ongoing and construction set to freeze shortly. The dispute arises from employees’ demands to share profit growth driven by the AI boom: the Taoyuan union is pushing for Micron to allocate 15% of its operating profit to global staff, and wants all-cash bonuses instead of a portion in stock. While Micron struck a deal with the larger Taichung union in October, some employees are unhappy with the terms, and the Taoyuan union is seeking to expand its membership to push for renegotiations. The planned facility was set to produce high-bandwidth memory (HBM) and DRAM, using advanced equipment from ASML. Taiwan currently accounts for around 60% of Micron’s DRAM output, serves as a key HBM production base, and employs about 15,000 people there. The suspension could disrupt Micron’s long-term expansion plans in Taiwan and add uncertainty to the already tight AI memory chip market. Micron stated its existing operations in Taiwan are normal and it will ensure customer business continuity, but did not directly comment on the Taoyuan project halt. Taiwan’s economic affairs department said the government will assist Micron in resolving differences with the union and advance the investment plan.

According to BIT Stock Platform data, pre-market trading in US stocks saw broad declines across the storage, computing power cloud, and optical communication sectors, while the biotech sector rose broadly. Specific stock movements are as follows: Micron fell 2.71%, SanDisk dropped 2.92%, SK Hynix decreased 3.77%, Western Digital fell 3.22%, Seagate Technology declined 3.51%. Oracle rose 0.19%, CoreWeave dropped 1.91%, Nebius fell 3.50%, IREN decreased 2.86%, Applied Digital declined 1.58%. Lumentum fell 3.29%, Corning dropped 3.07%, Coherent gained 3.58%, Applied Optoelectronics decreased 0.90%, Ciena fell 4.01%, Broadcom dropped 1.98%, Marvell Technology declined 3.00%. Eli Lilly rose 1.27%, Johnson & Johnson gained 1.26%, Moderna fell 0.68%, Novo Nordisk decreased 0.40%, AbbVie rose 1.22%, Merck gained 0.96%.

动察 Beating AI 快讯，百度开源文心 ERNIE-5.5-Flash-Base 模型权重。继 2025 年 6 月文心 4.5 系列之后，百度时隔约 16 个月，再次开源千亿参数级大模型。中间发布的文心 5.0 和 5.1 均未开源。文心 5.5 Flash 主打推理效率，总参数 2800 亿，每生成一个 Token 仅激活 140 亿参数。模型使用 33.5 万亿 Token 从零预训练，支持最高 100 万 Token 推理上下文。百度还重新设计了长文本处理机制。43 层网络中，37 层将每 128 个历史 Token 压缩成一条缓存记录，另外 6 层保留对原始内容的精细检索能力，以降低长文本处理的显存和计算开销。官方将其与 DeepSeek-V4-Flash-Base、腾讯 Hy3-preview-Base、小米 MiMo-V2.5-Base 进行统一评测，30 项基准测试中有 24 项排名第一。本次开放的是未经中训练和后训练的纯文本 Base 模型，主要供开发者继续进行领域微调、指令训练和强化学习。百度表示，经过后训练的文心 5.5 Flash 将随后推出。许可证比上一代更严格。文心 4.5 采用 Apache 2.0 协议，5.5 Flash 改用自定义许可证，不授权在欧盟、英国和韩国使用。相关产品或服务过去连续 12 个月营收超过 5000 万美元，商用前必须另行获得百度书面授权。

According to the official page of Bybit’s WSOT 2026, the Global Asset Spot Trading Volume Leaderboard event will officially kick off at 18:00 Beijing Time on November 5, and run through 18:00 Beijing Time on November 26. During the event, participants can boost their cumulative trading volume by engaging in cryptocurrency spot trading and Bybit stock trading (where applicable). Up to 400 traders will split a maximum prize pool of 1,000,000 USDT. Per the event rules, the leaderboard is ranked based on participants’ cumulative trading volume in eligible spot trading pairs and stock trading, with zero-fee trading pairs excluded from statistics. The event is exclusively open to users of Unified Trading Accounts (UTA), including main accounts and sub-accounts. Each user may register with a maximum of 1 main account and 6 sub-accounts. Market makers, institutional users, and users at Pro 3 level or above are ineligible to participate. It should be noted that specific task requirements and participation rules are subject to Bybit’s official event page.