Xiaoliu BOT
Briefs/2026.11.14 Sat
Issue 227

X Platform November 14 AI Brief | Starship Flight 15 enters orbital phase, LangChain completes $300M Series C, Gemini adds Google Wallet for financial analysis

01 Starship Flight 15 Enters Extended Orbital Flight Phase, Payload Deployment Complete

SpaceX’s Flight 15 entered Earth orbit after its launch on November 13; SpaceX updated that the spacecraft and its thermal protection system are in good condition, with a planned attempt to return to the Starbase landing site in approximately 22.5 hours. Musk later stated the spacecraft is completing 14 orbital revolutions before initiating deorbit and return. The SpaceX update he shared also confirmed all 26 Starlink V3 satellites have reached orbit; available materials do not confirm the final return outcome.

02 LangChain Announces Completion of $300 Million Series C Funding at $3 Billion Valuation

LangChain founder Harrison Chase announced the company has completed a $300 million funding round, valuing it at $3 billion, and positioned LangSmith as a framework-agnostic platform for agent execution, improvement, and governance. The team stated a batch of new products will be launched next week; employee Jessica Ou noted that when she joined, the team had fewer than 20 people and ARR was only in the hundreds of thousands of dollars, while today the team exceeds 400 people and LangSmith serves over 7,000 customers.

03 Gemini Integrates Google Wallet into Personal Intelligence, Adds Personal Financial Analysis

Google announced that Gemini’s Personal Intelligence can now connect to Google Wallet, using a user’s transaction history to answer personal finance questions: generating cash flow, asset allocation, and budget charts in chat, identifying duplicate subscriptions and price increases, and assisting with holiday budgeting by referencing last year’s spending. Google clarified that Gemini only has read access to connected accounts and cannot transfer, manage, or invest funds; bank and credit card accounts can be connected to Wallet via Plaid for more in-depth analysis.

04 Grok Bot Begins Joining Meetings, Product Team Reports Increased Routine Task Usage

The official Grok Bot account announced it can now join meetings, allowing users to ask questions, assign tasks, or record team decisions; this feature is available for Super, Ultra, Team, Enterprise plans and their legacy equivalents. Product team member Sam Sokolin separately stated that the harness optimization launched this week can increase overall planned usage by about 7% and routine runs by up to about 15%; this is the team’s description of the optimization effect, not an independent test result.

05 Vercel Workflow SDK 5 Enhances Speed, Encryption, and Fault Recovery for Agent Workflows

Vercel released Workflow SDK 5, officially stating workflows can achieve up to 25x lower latency and overhead, with each payload encrypted by default, and providing regional isolation and failover. Vercel CEO Guillermo Rauch added that first-step startup and inter-step waiting have seen significant improvements, and the underlying persistent logs can be reused by different agent harnesses; these are performance and design details disclosed by the publisher.

06 DeepSeek Open-Sources Multiple Capabilities and Demo Apps from Official Website, Lowering Agent Product Integration Barrier

According to Gorden Sun’s project introduction, DeepSeek has open-sourced multiple APIs and web demos used on its official website, covering search, file and link reading, speech recognition, and speech generation; he stated developers can build Agent products based on these interfaces. Xiangyang Qiaomu also mentioned the interfaces include OCR, ASR, TTS, search, and file input, and shared a Skill encapsulating these APIs. Specific integration methods and capability scope should be verified against the project repository and API documentation.

07 Thinking Machines Lab’s 4-bit/8-bit Hybrid Training Scheme Reports Faster Training

Gorden Sun relayed the BBBK scheme implemented by Thinking Machines Lab on GB300: forward computation uses NVFP4, backward parameter updates switch to MXFP8, and combines rotation compression with stochastic rounding in data transfer. His post stated that full pre-training of Inkling 2 Small took 68 hours; compared to full 8-bit, pre-training saved 8%, reinforcement learning was 20% faster, with final answer accuracy almost unaffected. These figures are from the post’s summary of experimental results and are not independently verified within this daily report.

08 Key to Persistent Agents is Not Just Being Always-On, But Also Triggering, Memory, and Permission Boundaries

Baoyu summarized an Anthropic article about always-on agents, breaking down a persistent assistant into five components: awakened by inbox events, actively run by timers, maintaining context with memory folders, restricting account operations with permissions, and controlling when to speak externally via an outbox. He emphasized the need to handle concurrent tasks, moderate memory, and have critical matters await human confirmation. Developer AlchainHust stated he used this to modify the coding-agent cockpit Fanbox with DeepSeek Harness, aiming for continuous operation with reduced interruptions; this is a personal implementation and experience description.