독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-1392 게시물

새로고침

openJiuwen Releases RSI-Native AgentOS: Enabling Agents to Add Features to Themselves via Layered Extension Points

주제 · openJiuwen중대
2026-11-13 14:54 GMT+8

When using WorkSwarm to handle tasks, can you simply tell the Agent about a new feature you want, letting it build and install it for a test? openJiuwen’s answer is yes, and this mechanism is called RSI-Native AgentOS. It provides a runtime environment for Recursive Self-Improvement (RSI), advancing the extension mechanism from "conveniently allowing developers to integrate new capabilities" to "enabling Agents to proactively improve themselves based on task feedback."

Architecturally, openJiuwen divides the system into three layers: the Web client handles the interface display and receives user operations; the Gateway handles request access and forwarding; and the Agent layer consists of the AgentServer managing sessions and the Harness organizing task execution and tool calls. Each of these three layers retains extension entry points. Following this architecture, extension capabilities are divided into five segments—the Client can add buttons, panels, and settings items; the Gateway can integrate authentication or adjust request routing; the AgentServer can handle sessions, messages, and server-side commands; Swarm can customize multi-Agent collaboration rules; and the Agent can adjust prompts, integrate tools, or sub-Agents.

These modifications are delivered via extension packages. The manifest.json within the package specifies the package ID, version, dependencies, and code entry point; when involving interfaces, it can also declare where components should be displayed, such as in the input area, sidebar, or settings page. During installation, the system reads and verifies the manifest, saving the extension package by its ID and version. During loading, the Gateway and AgentServer respectively select packages running under their own control, determine the versions and order, and register the processing functions from the extension code into their respective event buses. When a user clicks an interface button, the gateway receives a request, or the Agent begins assembling a prompt, the corresponding component emits an event, the event bus finds the matching processing function, and invokes it.

In practical use, WorkSwarm does not default to a desktop pet form. A user tells the Agent the requirement—'Add a desktop pet to WorkSwarm so I can chat with you by clicking on the desktop'—and the Agent generates an extension package based on the description, then installs it into WorkSwarm as a system plugin. Once enabled, the workbench can be collapsed, leaving only the desktop pet on the screen; double-clicking the pet returns the user to the workbench, preserving the previous conversation. Another example is Token Weather: the user says 'Add a status bar above the input box showing the current session's token usage,' and the Agent writes the code, packages it, and installs it. After trying it out and finding it takes up too much space, the user says 'Make it more compact, but keep the usage percentage.' The Agent modifies it according to the feedback and reloads it. Throughout this cycle, new features are hot-pluggable, requiring no changes to the application's original code upon integration; extension packages are managed by version, and if a new version causes issues, it can be rolled back to a saved checkpoint.

출처:量子位

리서치

Google DeepMind Releases AlphaProof Code to Verify Signal Encryption Library Down to Machine Code

중대
2026-11-13 14:39 GMT+8

Google DeepMind recently announced its AI agent formal verification framework, AlphaProof Code, which was used to verify libsignal, the core program library of the Signal encryption communication protocol. This work extends the scope of security proofs from source code to compiled machine code.

According to iThome, five DeepMind researchers collaborated with AI agents to produce approximately 6.5 million lines of mathematical proof code written in the Lean language over four weeks. These proofs aim to verify key security requirements such as forward secrecy, authentication, and post-compromise security recovery.

Formal verification is a technique that proves software correctness through mathematical methods, typically requiring significant time and specialized expertise. DeepMind's progress suggests that AI agents can automate the large-scale generation of complex verification proofs, potentially lowering the barrier and cost for auditing the security of high-assurance software like end-to-end encrypted applications.

출처:iThome

리서치

Google's New Smart Glasses First to Support 'Hey Gemini' Wake Word, Phasing Out 'Hey Google'

주제 · 谷歌Gemini语音唤醒간단한 요점
2026-11-13 20:15 GMT+8

According to German tech outlet t3n.de, Google has introduced "Hey Gemini" as a new voice activation word on its latest smart glasses, developed in collaboration with Warby Parker, Gentle Monster, and Samsung. Users can now wake the built-in Gemini AI assistant using this dedicated phrase, while traditional commands like "Hey Google" and "OK Google" remain functional.

This update signals Google's ongoing strategy to replace the Google Assistant brand with Gemini across its ecosystem. The company previously forced a transition on smartphones, despite some user resistance. By supporting both wake words on the new eyewear, Google appears to be adopting a gradual approach, allowing users to acclimate to the new branding before a potential full switch on other devices. It remains unclear when this feature will expand to existing hardware.

출처:t3n.de

리서치

Report: Anthropic IPO Six Times Oversubscribed Despite Tech Pullback; Goldman Forecasts $600B Revenue by 2030

중대
2026-11-13 19:48 GMT+8

According to a Semafor report published on November 13, 2026, artificial intelligence company Anthropic's Initial Public Offering (IPO) has been reportedly six times oversubscribed. This high level of demand comes amidst a significant correction in the technology sector, where investors have pulled out from tech stocks due to fears of excessive spending and soaring debt piles, with the Nasdaq index registering its biggest monthly loss in over a year in October.

The report notes that while the resignation of Nvidia's CEO amplified concerns about the sustainability of the global AI buildout, some investors appear undaunted. Goldman Sachs, one of the banks leading Anthropic's IPO, expects the firm's revenue to exceed $600 billion by 2030, more than 130 times its sales last year. These upbeat predictions for future growth likely underpin the strong subscription interest.

출처:semafor.com

리서치

AWS Releases Bedrock AgentCore Runtime V2: 4x Compute Limits, Cold Starts Under 2 Seconds

중대
2026-11-13 21:10 GMT+8

Amazon Web Services (AWS) has announced that its managed execution environment for AI agents, Amazon Bedrock AgentCore Runtime, has been upgraded to version V2, which is now the default configuration for new execution environments across 19 AWS regions globally. The new version increases the virtual CPU and memory limits for a single session by four times compared to the previous generation, reaching up to 8 vCPUs and 32 GB of memory. The container image size limit has also been raised to 10 GB. This expansion allows AI agents to compile large program projects, operate browsers, and load local models within the same execution environment, meeting the growing computational demands of complex agent tasks.

In terms of performance optimization, V2 adopts a snapshot startup mechanism introduced in September this year. The system pre-completes agent initialization and creates a snapshot, allowing new sessions to restore directly from the snapshot, thereby eliminating the time required for repeated program loading and environment initialization. According to AWS test data, when using a 2 GB container image with a simple agent that does not call AI models, 75% of sessions have a startup time of less than 2 seconds, compared to 29.7 seconds for the previous generation. Additionally, each session continues to use an isolated Firecracker microVM to ensure hardware isolation, and the system can scale down computing resources to zero when the agent is idle.

Regarding memory management, V2 reclaims resources within minutes after the agent releases memory, rather than waiting for the entire session to end. AWS states that for AI agents requiring multi-turn interactions, a single session can reduce billed memory usage by up to 68%, though this does not imply that overall costs will decrease by the same proportion. Existing execution environments will not automatically upgrade due to the change in default versions; developers can switch to V2 or continue using V1 by setting the platformVersion parameter.

출처:iThome

리서치

Google Gemini launches 'Memory summary' to let users view and edit AI-stored personal details

주제 · Gemini记忆摘要功能간단한 요점
2026-11-13 01:58 GMT+8

Google has introduced a new feature called "Memory summary" in its AI assistant Gemini, enabling users to view, add, or remove personal details that Gemini has stored about them. The feature is now available across the Gemini website, iOS and Android mobile apps, and Windows and Mac desktop applications.

Users can access this function via the "Personal Intelligence" section in settings. Memories are organized by categories such as Essentials and Business & Career. Users can directly edit these entries using a pencil icon or instruct Gemini to modify specific memories through conversation. For instance, if Gemini incorrectly records a user's location or profession, they can manually correct it, and changes sync across all devices.

This update addresses growing concerns about AI privacy and data control. While Gemini's memory feature aids personalization, users previously had limited visibility into what information was being stored. The new "Memory history" option also allows users to track when memories were changed.

출처:ZDNET

리서치

Bloomberg: SpaceX and Tesla Merger Talks Advance, Targeting December Announcement

구조적
2026-11-13 14:43 GMT+8

According to Bloomberg, people familiar with the matter revealed that merger negotiations between SpaceX and Tesla have made substantial progress. Boards and advisory teams from both companies are currently discussing specific merger terms, equity arrangements, and the governance structure of the new entity. The report states that the parties are targeting an announcement as early as December 2026, with an alternative plan to delay until early 2027.

Currently, the combined market capitalization of these two U.S.-listed companies is approximately $3.4 trillion. If the merger is completed, it will directly break the global record for mergers and acquisitions, becoming the largest M&A deal in history. The post-merger company's market cap is expected to rank around 5th among U.S. listed companies, surpassing Amazon. This would place Elon Musk at the helm of a massive corporate group spanning rockets, satellite internet, AI, chips, computing power, electric vehicles, robotics, and energy storage.

Previously, during earnings calls in October 2026, Musk confirmed that both companies had established special committees of independent directors to discuss a potential merger, but he emphasized at the time that this was not a certainty and faced high hurdles, including requiring support from 50% of all Tesla shares. This news regarding specific term negotiations marks a shift in the process from preliminary exploration to a more advanced execution phase.

출처:量子位

리서치

IBM Releases LLMON Markup Language and Experimental Granite Preview Models to Combat Prompt Injection

주제 · LLMON중대
2026-11-13 03:07 GMT+8

IBM has released LLMON (LLM Object Notation), a new markup language designed specifically for large language models, along with two experimental preview models: granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. These models are derived from the Granite 4.1 base checkpoints and are trained to consume and generate text marked with LLMON tags.

The core innovation of LLMON is its use of reserved special tokens to explicitly mark structural components like instructions, data, and requirements. Unlike JSON, which relies on characters that can appear in normal text, LLMON uses unique delimiters that allow the model to distinguish between executable commands and passive data payloads. This structural separation aims to mitigate indirect prompt injection attacks, where malicious instructions are hidden within retrieved documents or user data.

According to IBM's internal evaluation on the OPI benchmark, the LLMON models demonstrated significantly higher robustness against static indirect prompt injections compared to other open-source and closed-source models. The company reports that these models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while maintaining general capability comparable to the standard Granite 4.1 Instruct models.

The release includes a GitHub repository with conversion tools that allow users to translate existing JSON pipelines into LLMON format losslessly. Additionally, IBM provides notebooks demonstrating how to fine-tune LoRA adapters on top of the LLMON models to understand custom domain-specific vocabularies. These are experimental previews intended for research and evaluation, not for production deployment.

출처:huggingface.co

리서치

AWS Quick Introduces Smart Approvals: AI Classifier Filters Routine Actions, Escalating Only High-Risk Operations for Human Review

중대
2026-11-13 10:55 GMT+8

Amazon Web Services (AWS) has introduced a new feature called "Smart Approvals" within its business AI workspace, AWS Quick. This mechanism aims to address "confirmation fatigue," a phenomenon where users, overwhelmed by frequent prompts asking for permission to execute AI agent actions, often approve requests without careful review.

Smart Approvals employs a two-stage processing flow. First, the system applies existing permission settings (such as Custom Permissions, Quick Guardrails, and connector action permissions) for initial determination. Operations explicitly denied or requiring mandatory human approval remain unchanged. Only actions set to "confirm every time" by the user, but lacking other strict approval requirements, proceed to automatic review.

During automatic review, a separate AI model referred to as a "classifier" evaluates the operation's content, data destination, and affected individuals against the user's original request and conversation context. For instance, if a user asks to send materials to two colleagues, but a referenced note contains instructions to send them to all 2,000 sales staff, the classifier identifies this discrepancy and returns the action for human confirmation. This design also helps mitigate indirect prompt injection attacks.

High-risk operations, such as external transmissions, permission changes, record deletions, or payment initiations, are escalated to human review even if they align with the user's request. Automatic approval is valid only for a single instance with matching parameters; if the classifier times out or fails, the action defaults to human review. All related permission updates are logged in AWS CloudTrail for subsequent auditing.

AWS explicitly states that model-based reviews can be erroneous and recommends maintaining strict human approval requirements for operations involving confidential data. Enterprises are advised to start deployment with pilot teams, gradually expanding the scope of automatic approvals.

출처:テクノエッジ TechnoEdge

리서치

Microsoft Cloud Outage Disrupts Outlook, Teams and Copilot for Hours; Resolution Attempts Fail

주제 · 微软云服务中断事件중대
2026-11-13 14:22 GMT+8

US technology giant Microsoft experienced a severe multi-hour outage of its cloud service platform on November 13, 2026, rendering email service Outlook, collaboration platform Teams, and AI assistant Copilot unusable. The disruption affected a large number of businesses and consumers who rely on these tools for daily operations and data exchange.

According to German tech outlet t3n, Microsoft posted updates about the interruption on the online platform X. Initially reported as issues with Exchange mailboxes in the US, Microsoft subsequently stated that the disruption could affect Microsoft 365 cloud users worldwide. The company attempted to route data traffic around potentially problematic servers but achieved no breakthrough success initially.

Microsoft engineers identified a process on affected servers that paralyzed them and attempted to bring some servers back online. Although there were brief reports of stability, the situation deteriorated an hour later: servers hung again as soon as data traffic reached them. Microsoft then tried to start up parts of the infrastructure in isolation, seeing "positive results" briefly before admitting the measures did not work as expected. As of the report, the duration of the ongoing disruption remained unclear.

This is not Microsoft's first major service interruption recently. In late August 2026, its email services suffered a hours-long widespread outage due to a problem in the cloud infrastructure, preventing users from logging in or sending and receiving messages. Such incidents highlight the vulnerability of major cloud providers when dealing with complex infrastructure failures and the operational risks posed to enterprise customers heavily dependent on a single cloud platform.

출처:t3n.de

리서치

Japanese Startup higurashi Open-Sources Full-Duplex Voice Model Haru 4B

주제 · Haru 4B중대
검증됨 2026-11-13 20:55 GMT+8

Japanese AI startup higurashi released Haru 4B, a full-duplex speech dialogue model, on November 13, 2026. The company has made the pre-trained weights and inference code available on Hugging Face under the Apache 2.0 license. With 4 billion parameters, the model is specifically designed to address the needs for naturalness and low latency in real-time voice interactions.

Unlike traditional "cascaded" architectures that sequentially process Automatic Speech Recognition (ASR), Large Language Models (LLM), and Text-to-Speech (TTS), Haru 4B employs an end-to-end full-duplex LSLM (Large Speech Language Model) technology. This architecture allows the model to simultaneously process three data streams—user voice, model voice, and text—synchronizing information approximately every 133 milliseconds. This enables the model to "listen while speaking," facilitating low-latency responses to user interruptions, backchanneling, and topic shifts, while preserving paralinguistic features such as tone and emotion.

According to higurashi's official announcement, Haru 4B was pre-trained on 40 million hours of multilingual data including Japanese and English, currently supporting nine languages. The model also incorporates the VibeVoice continuous speech tokenizer to balance audio quality with computational efficiency. Despite its fluent conversational capabilities, higurashi explicitly notes that due to its 4B parameter size, the model's logical reasoning ability is roughly equivalent to GPT-2 levels, highlighting that its primary value lies in its interaction mechanism rather than complex task processing.

출처:higurashi

리서치

Ex-Anker CMO's Memoket Launches 11g AI Wearable Recorder with Privacy-First Manual Trigger

주제 · Memoket Gem간단한 요점
2026-11-13 08:08 GMT+8

Memoket, a startup founded by former Anker Chief Marketing Officer Terrence Wang, has launched its first AI wearable device, the "Memoket Gem." Weighing approximately 11.3 grams, the device aims to serve as a "context layer" between the physical world and AI. It records meetings and daily conversations, utilizing its proprietary long-term memory engine, "KITE," to connect disparate dialogue fragments and provide background information to AI assistants like ChatGPT or Claude.

Unlike many traditional recording devices, Memoket explicitly chose not to develop an auto-start recording feature. CEO Terrence Wang stated that while technically feasible, the decision was made due to high sensitivity regarding data security and privacy among consumers in Europe, the US, and Japan. The company decided to leave the control of starting and stopping recordings entirely to the user. The device features a red LED light during recording to ensure those nearby are aware, balancing convenience with social etiquette.

The Memoket Gem supports multiple wearing options, including attaching directly to Apple Watch bands, addressing the issue of limited wrist space for users. Its core selling point is the "KITE" engine, which does not rely on traditional vector retrieval but instead organizes speech into facts and establishes links, significantly reducing token consumption when answering queries. The product is currently available in the Japanese market for ¥31,980 (tax included), offering 300 minutes of free monthly transcription.

출처:テクノエッジ TechnoEdge

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

Looking further ahead, openJiuwen believes that as more segments become open to extensions, individual usage habits, team collaboration rules, and enterprise operational requirements can all serve as bases for Agents to adjust themselves. Initially, it has a default appearance; what it becomes later, it can modify itself through actual work tasks piece by piece. openJiuwen summarizes this state as 'What the Agent thinks is what you get.'