독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-19129 게시물

새로고침

Apple Integrates MCP Protocol into iWork Suite, Allowing External LLMs to Control Pages, Numbers, and Keynote

중대
2026-11-19 19:08 GMT+8

In a surprising move within its version 16 update for the Mac office suite, Apple has integrated the Model Context Protocol (MCP) into Pages, Numbers, and Keynote. This addition allows external Large Language Models (LLMs) to directly control these applications. Users can now connect tools such as ChatGPT, Claude, or locally running AI models via this interface to perform tasks like document editing, data analysis, and presentation generation. Previously, MCP support was rare in Apple software, available only in Safari and Xcode, while other creative tools like Final Cut Pro and Logic Pro lacked it.

Alongside MCP support, the update introduces an "Intelligent Assistant" feature that leverages Apple Intelligence, OpenAI, and Google Gemini models to generate text and image content. However, both the advanced AI functions and MCP access are locked behind the monthly €13 Creator Studio subscription bundle. The free base versions receive only bug fixes and performance improvements. Furthermore, the AI assistant is currently limited to US English and requires devices supporting Apple Intelligence running iOS 27, iPadOS 27, or macOS 27.

For developers and enterprise users, the integration of MCP signifies that Apple's native productivity tools are no longer confined to internal AI ecosystems but are open to a broader range of AI agents. This could accelerate the penetration of third-party AI tools in office scenarios but also raises new considerations regarding data privacy and the security of automated operations, as external models can now directly read and write user document and spreadsheet data.

출처:Mac & i

리서치

Beijing AI Startup Norizon AI Completes Hundreds of Millions RMB Angel and Angel+ Rounds

간단한 요점
2026-11-19 13:19 GMT+8

According to a report by Zhidongxi, Beijing-based AI startup Norizon AI announced on November 19 that it has completed angel and angel+ funding rounds totaling hundreds of millions of RMB. The round was participated in by investors including Five Yields Capital, Sequoia China, Lenovo Star, Zhongguancun Science City Fund, and Gobi Ventures, with FlowCapital serving as the financial advisor.

Norizon AI specializes in Real-Time Recursive Self-Improvement (RSI) technology, aiming to enable large models to continuously learn and update their own parameters after deployment. One of its core self-developed technologies is the SHINE hypernetwork, which attempts to complete lightweight parameter updates through a single forward calculation, thereby reducing the computational overhead of continuous learning and embedding new knowledge directly into model capabilities rather than relying solely on expanding context windows or external retrieval.

The company's founder and chief scientist, Zhang Muhan, previously served as an assistant professor at the Institute for Artificial Intelligence at Peking University and worked on large-scale graph machine learning systems at Meta AI. The current team size exceeds 20 people. The funds raised will be primarily used for related technology research and development and product implementation.

출처:zhidx.com

리서치

AI Researcher Claims Discovery of New Maya Site; Archaeologists Point Out It Was Documented in 2025 Paper

중대
2026-11-19 19:00 GMT+8

On November 10, 2026, artificial intelligence researcher Pietro Miotti shared on social media platform X that he had used AI to "discover" a previously unmapped ancient Maya civic center. Miotti stated that while on vacation in Peru, he prompted dozens of AI agents to sift through publicly available NASA lidar (light detection and ranging) data to look for signs of hidden structures. The AI could not identify any evidence that the site had been documented in academic papers, leading Miotti to contact Mexican officials.

However, the claim was quickly challenged. Sarah Parcak, a professor at the University of Alabama at Birmingham, noted that the site had been documented at least once before—by Francisco Estrada-Belli, a professor at Tulane University, and his colleagues in a 2025 paper published in the Journal of Archaeological Science: Reports. "This is why peer review exists," Parcak said.

Marcello Canuto, a professor of anthropology at Tulane and co-author of the 2025 study, added that their work built on lidar data analyzed as early as 2016. Ivan Šprajc, head of the Institute of Anthropological and Spatial Studies at the Research Center of the Slovenian Academy of Sciences and Arts, who helped conduct a visit to the site, described it as a "peripheral group" of a larger center.

Miotti later updated his blog post and posted a correction on X, crediting the archaeologists' survey. The incident sparked a debate over the role of AI in science. Some commentators argued it demonstrated AI's shortcomings, terming it "artificial mediocrity." Parcak pointed out that AI relies on information available online, whereas millions of articles, books, and archaeological datasets are not contained within any publicly accessible databases. "Not everything is online," she said. "That's where human expertise—and collaboration—comes in."

출처:Scientific American

리서치

Siemens TIA Portal V21.1 Decouples Licensing from Installation and Converts Engineering Data to Text for AI Integration

중대
2026-11-18 20:33 GMT+8

German industrial giant Siemens released a major update to its automation engineering software, TIA Portal V21.1, on November 19, 2026. The core change lies in a structural shift in software delivery: Siemens has decoupled software installation from licensing. Within the V21 version line, users no longer purchase specific version numbers but gain continuous access to the latest updates. This means engineers do not need to upgrade and revalidate entire projects with each new release, significantly reducing repetitive maintenance work, simplifying lifecycle management, and lowering operational risks associated with frequent upgrades.

Beyond the delivery model, V21.1 introduces features designed to promote artificial intelligence integration. Siemens expanded Simatic Source Documents to cover hardware configuration and converted these engineering contents into text-based source code formats. This transformation allows external tools like Git and code generators to integrate more easily and, crucially, makes engineering information structured and machine-readable. Siemens states this is a key step toward making TIA Portal "increasingly AI-ready," laying the foundation for AI-supported engineering workflows.

Additionally, for high-end applications, the new version introduces multicore programming for S7-1500 controllers, allowing motion control and safety programs to run independently, thereby improving determinism and runtime performance. The version will continue to connect with the Eigen Engineering Agent, an AI-driven agent presented by Siemens at Hannover Messe in April 2026. TIA Portal V21.1 is scheduled for download availability at the end of November 2026 and package delivery in December 2026.

출처:press.siemens.com

리서치

Microsoft AI Agent Framework Hit by Supply Chain Attack Stealing Cloud Credentials

중대
2026-11-18

Microsoft's official AI agent development framework, "agent-framework," suffered a supply chain attack on the Python Package Index (PyPI). Independent analyses by security firms Wiz, Socket, Aikido, and StepSecurity revealed that versions 1.26.1 and 1.26.2 contained malicious code designed to steal cloud service and Continuous Integration/Continuous Deployment (CI/CD) credentials.

The compromise occurred between 20:35 and 21:20 UTC on November 17, 2026. During this window, the infected versions were available for download. The malware exploited Python's .pth file mechanism to execute automatically whenever the Python interpreter started, without requiring an explicit import of the library. Once active, it scanned environment variables for sensitive information (such as those containing TOKEN, SECRET, KEY, AZURE, or AWS) and local configuration files (like ~/.azure, .npmrc, and .git-credentials), exfiltrating this data to external servers.

The impact extends to any development environment that installed or updated agent-framework during that timeframe. As a core tool for building multi-agent workflows, the package sees approximately 190,000 to 200,000 monthly downloads, creating a significant potential exposure surface. The latest version on PyPI has since been reverted to the clean 1.26.0 release. Microsoft has not yet issued a formal statement, though related GitHub repositories temporarily returned 404 errors, indicating ongoing investigation.

Teams potentially affected should immediately take several steps: first, check lockfiles, CI logs, and container images for versions 1.26.1 or 1.26.2; second, if infection signs are found, completely rebuild virtual environments and images rather than just deleting files; finally, rotate all secrets accessible from the affected systems, including Azure service principals, GitHub tokens, Docker credentials, and API keys. Notably, according to Wiz's analysis, stolen Azure tokens may remain valid for up to 24 hours, so even after revoking federated credentials, Entra ID login activities must be checked to confirm no unauthorized access occurred.

출처:KIberblick

리서치

Hugging Face Researcher Benchmarks Show Open-Source OCR Costs Drop to $65-$130 Per Million Pages

중대
2026-11-19 17:28 GMT+8

Daniel van Strien, a researcher at Hugging Face, benchmarked eight open-source Optical Character Recognition (OCR) models, revealing a significant drop in the cost of extracting text from scanned books and records. The tests were run on an Nvidia RTX PRO 6000 GPU, rented through Hugging Face for $2.75 per hour.

Data shows that compared to RolmOCR, released in April 2025, which cost about 51 cents per 1,000 pages, the newest model LightOnOCR-3-0.8B, released in October 2026, brought this cost down to just under 10 cents. This means digitizing one million pages currently costs between $65 and $130 in GPU rental fees, depending on the source material.

Van Strien noted that optimizing how the models ran—such as feeding pages more evenly, using smaller scans, updating software, and running faster GPUs—saved as much money as switching to newer models, if not more. While the newest models read book pages at least as accurately as the oldest ones, some models exhibited issues like hallucinating text on blank pages or repeating content.

출처:The Decoder

리서치

Climate Change AI Co-founder: Climate Applications Need 'Bees' Rather Than 'Elephants'

중대
2026-11-19 17:08 GMT+8

David Rolnick, Associate Professor at the School of Computer Science, McGill University in Canada, and co-founder of Climate Change AI, recently put forward a core judgment at the "Green Computing Power and Climate Digital Technology" forum hosted by Taiwan's Ministry of the Environment during the United Nations COP31 period: When discussing AI and climate change, the real question is not "whether to use AI," but "what kind of AI should be used for what problem."

He uses the metaphor of "elephants and bees" to illustrate two directions in current AI development. Large generative models such as ChatGPT, Gemini, and Claude are "elephants"—massive in scale, requiring large data centers, and primarily handling general tasks such as text generation, image synthesis, and code writing. The other type of AI is the "bee"—specialized models used for prediction, classification, optimization, and scientific computing, such as analyzing satellite imagery to identify land-use changes, predicting wildfire risks, or optimizing grid operations, which do not require support from large data centers.

Rolnick emphasizes a key distinction: these two types of AI are fundamentally different and cannot simply be understood as "shrinking a large model" to obtain specialized AI. Citing research cases, he points out that while large language models may have trillions of parameters, some climate application models only require millions of parameters—a difference of several million times. For specific scientific problems, smaller models can actually be more efficient.

His argument is that AI applications for many social issues should be a combination of a few "elephants" and multiple "bees." Most problems are solved by specialized small models, which improves efficiency while reducing unnecessary energy consumption. Taking the climate sector as an example, if the goal is to predict solar power generation for the next hour, what is needed is a prediction model combining meteorological and energy data, rather than a large language model.

Rolnick further criticizes a spreading deployment pattern: high-energy-consuming AI functions are preset into daily services. He cites Google Search's introduction of generative AI summary features as an example—users do not actively request generative AI summaries for every search, yet the system automatically generates them. He estimates that if billions of searches globally adopt this feature, the overall electricity consumption could be comparable to the residential electricity usage of a large city.

He believes the problem lies not with the power consumption of a single model, but with preset deployment. Generative AI should be prioritized for scenarios that truly require reasoning and content generation, rather than becoming the default solution for all applications. He suggests that AI development should adjust in several directions: promoting more medium-sized computing resources so users can use task-appropriate AI through personal computers or edge devices; increasing AI energy transparency so developers and users understand the energy costs of different tools, hardware, and tasks; and continuously improving the energy efficiency of models, hardware, and data centers.

출처:iThome

리서치

Meta Says AI Agents Have Handled Over 60% of Work in Its Ad Ranking Technology Propagation

중대
2026-11-19 01:00 GMT+8

Meta’s ad ranking system is composed of numerous heterogeneous models, each with its own architecture, training data, and inference constraints. An ML technique (such as Wukong factorization layers or Interformer interaction networks) validated on one model can rarely be directly copied to the next. Previously, Meta organized models into reusable components using Standard Model Templates (SMT), allowing technology validation to propagate automatically to all models based on that template; this delivered over 5,000 technology-model combinations within two years and reduced per-model engineering time by 92%. However, template automation only resolved deterministic steps; remaining critical decisions—which technology is worth propagating to which model, how wide the hyperparameter search should be, why a model failed to improve, and how to detect and recover from failure—still relied on human judgment by engineers.

Meta describes AI Agents taking over these judgment decisions in three stages. In the discovery stage, the Agent evaluates the entire model fleet simultaneously based on a shared knowledge base (accumulated historical training runs, model architectures, and technology results), generating a list of technology-model candidates ranked by expected impact, replacing the previous method where domain experts assessed models one by one. In the generalization stage, the Agent leverages results from millions of historical training runs, starting from configurations most likely to be effective, narrowing the search space for the next round based on each round's results, and deciding which experiments are worth continuing and which should be terminated early. In the adoption stage, the Agent manages hundreds of concurrent training jobs, prioritizing them by revenue importance, technological potential, and deadlines, and reclaiming GPU capacity from stalled or low-value runs.

The efficiency figures reported by Meta each have measurement boundaries. The 60% coverage rate is an estimate calculated since the Agent began driving execution in March 2026. The 3x productivity gain is measured against the same work prior to Agent intervention. The 4x compute savings is a median figure, explicitly limited to select technology-model combinations. The $45 million in savings is an estimate through year-end. Meta also notes that engineer-led custom adaptation is still used for the highest-revenue models, where the Agent acts as an assistant that learns from engineers' decisions before assisting in propagation to other models. Currently, it takes over a year for a verified technology to reach all applicable models; Meta’s goal is to shorten this to three months, meaning agent-driven workflows must achieve 5x higher end-to-end throughput.

출처:Engineering at Meta

리서치

Chinese AI Firm DeepSeek Adjusts API Pricing: In-Model Search Now Token-Billed Only

주제 · DeepSeek API Pricing중대
2026-11-19 16:51 GMT+8

Chinese AI company DeepSeek adjusted its API billing rules effective 12:00 Beijing time on November 18. According to a report by Zhidx, when developers execute tasks using DeepSeek models that trigger server-side search or web scraping, these operations will no longer incur separate per-call fees. Instead, they are included in the standard token usage billing. This significantly reduces the marginal cost for developers building AI Agents that require frequent online information retrieval.

It is important to note that this adjustment does not make all search and scraping interfaces free. If developers bypass model calls and directly request the independent Search API or Fetch API, they must still pay based on the number of successful requests. The official pricing page lists both services at 5 yuan per 1,000 successful requests. This distinction aims to encourage developers to integrate search capabilities naturally within model inference rather than using DeepSeek solely as an external search engine.

Accompanying this billing change, DeepSeek Harness released version v0.2.2-rc.5. According to the GitHub changelog, when users employ DeepSeek models within this framework, Harness defaults to DeepSeek's server-side search to benefit from the waived per-call fees. However, when users employ other models, Harness defaults to calling the independent DeepSeek Search API, where search costs are still charged according to interface rules.

This cost optimization for developers follows a tightening of consumer-side experience. On November 16, some users of the DeepSeek app and web interface encountered issues such as interrupted long conversations, reduced input limits, and daily quota restrictions, with system prompts citing "insufficient computing power." Many users called for paid services on social media to retain their previous experience. The contrast between resource throttling for general users and tool-call cost reductions for developers reflects DeepSeek's attempt to allocate limited computing resources toward the high-value developer ecosystem, aiming to attract more enterprises and individuals to build applications on its models by lowering barriers.

출처:zhidx.com

리서치

California Proposes Emergency Shutdown Powers for High-Risk AI Models in New Safety Report

주제 · California AI Emergency Shutdown Regulation중대
2026-11-19 15:52 GMT+8

The California state government released a policy report titled "AI Safety in 2026 and Beyond" on November 18, 2026, outlining legislative recommendations to strengthen artificial intelligence oversight. The proposal grants the state government emergency powers to directly order companies to halt high-risk AI models under specific conditions, without requiring a prior court injunction. This marks a significant step toward more coercive local regulation of potential AI out-of-control scenarios.

The report mandates that large advanced AI developers prepare emergency response measures, including stopping AI agent operations, revoking system access, and isolating computing resources. If a model poses an immediate risk of catastrophic accident or severe loss of control, and the company fails or refuses to address the danger, the state can intervene to force a shutdown. Additionally, the plan suggests establishing continuous third-party safety verification, requiring independent assessment bodies to monitor training, development, and deployment phases and submit full results to the government.

Addressing the growing concern over Recursive Self-Improvement (RSI)—where AI systems autonomously develop stronger successor models without human intervention—California proposes specific restrictions. The report recommends requiring companies to disclose the extent of AI involvement in research and the computational resources dedicated to automated development. If AI-driven progress exceeds current safety measures and evaluation capabilities, posing a substantial risk of catastrophic harm, the state would have the authority to limit the proportion of compute resources allocated to such automated R&D.

Regarding incident reporting, the report suggests expanding the definition of major AI safety events to include model failures and attempts to steal model weights, while creating channels for near-miss incidents. The deadline for reporting most major events is proposed to be shortened from the current 15 days to 48 hours, with a 24-hour requirement maintained for events with imminent catastrophic risk.

It is important to note that these measures remain at the policy recommendation stage and require passage through the California legislative process to become law. Previously, California enacted SB 53 in 2025 and signed SB 813 and AB 1405 in 2026, establishing certification for independent verification organizations and registration for AI auditors. This report follows Governor Gavin Newsom's September directive to study emergency shutdown mechanisms, presenting a concrete legislative blueprint as regulators seek a new balance between fostering innovation and preventing systemic risks.

출처:iThome

리서치

Why AI Still Cannot Autonomously Adjust Process Parameters in Semiconductor Manufacturing

주제 · AI Autonomy in Semiconductor Manufacturing중대
2026-11-19 16:01 GMT+8

At the SEMICON West 2026 conference, a recurring question was discussed: If AI can already identify process solutions invisible to human engineers, why is it still unable to autonomously adjust production parameters?

David Fried, Chief AI Officer at Lam Research, described a specific scenario: The potential recipe space for an etcher involves approximately 10^23 combinations. Constrained by historical experience, human engineers tend to revert to parameter regions where problems were solved ten years ago. In contrast, AI can point to operating regions that engineers have never considered. Vikas Sharma, Senior Director of Pathfinding for R&D at Intel Foundry, provided an even more extreme example: A process layer involving 350 parameters has a theoretical combination count of approximately 10^700. Engineers and equipment suppliers spent a year developing a minimally acceptable recipe, but after introducing AI-assisted optimization, multiple solutions superior to the original baseline were found within just a few months.

The capability has been proven, yet the authority remains ungranted. Where are the obstacles?

The first obstacle is that models identify statistical relationships rather than physical mechanisms. Eric Guichard, Head of AI Manufacturing Business at Silvaco, pointed out that surrogate models are reliable when interpolating within the region represented by their training data, but become less reliable when extrapolating beyond it. The fact that AI identifies new operating regions—precisely what makes it most valuable—also means that engineers have the least experimental evidence on hand to validate its predictions. The relative importance of pressure, temperature, voltage, and gas flow may differ completely across different operating regions; a model that performs well in one region cannot be assumed to hold true in another.

The second obstacle is data fragmentation. Information in semiconductor manufacturing is often scattered across different companies and systems. In advanced packaging, a foundry provides the die, another company produces the substrate, and an OSAT (Outsourced Semiconductor Assembly and Test provider) completes assembly and testing; these three parties typically cannot share the same manufacturing history. It is even more difficult because physical properties such as substrate warpage, thermal stress, and material behavior change continuously during manufacturing. These changes are often not captured by existing measurements, and models relying on nominal material properties may miss them entirely. Kenneth Larsen, Senior Director of Product Management at Synopsys, summarized the core issue: If the training range, calibration source, and approved use are not transmitted along with the model, it may be sufficient for early exploration but unsafe for design closure. What is missing is not more data, but the context that tells the next tool or team how the data should be used.

The third obstacle is the time lag between model updates and production. Sharma explicitly stated that the industry is far from continuous learning; the actual practice is updating models quarterly or monthly. A fab may process 10,000 wafers between two updates, and the data generated by these wafers is not rapidly used to improve the model or agent learning. More perplexingly, if the model adjusts equipment settings between updates, engineers need to know exactly what changed and whether the updated behavior remains within verified limits. When investigating failures that emerge weeks later, it may be necessary to reconstruct the state of the model when it processed specific batches.

출처:semiengineering.com

리서치

US Industrial Automation Giant Rockwell Automation Signs Deal to Acquire Austrian AMR Maker AGILOX

주제 · Rockwell-AGILOX Acquisition중대
2026-11-18

US-based industrial automation company Rockwell Automation (NYSE: ROK) announced on November 18, 2026, that it has signed a definitive agreement to acquire AGILOX, an Austria-based manufacturer of autonomous mobile robots (AMRs). The transaction is subject to regulatory approval and is expected to close in the second quarter of Rockwell’s fiscal year 2027.

Founded in 2017, AGILOX specializes in high-end AMRs driven by self-developed artificial intelligence and swarm networking technology, with over 2,500 forklift AMRs deployed globally. Rockwell stated that upon completion, AGILOX’s proven forklift AMR portfolio will complement OTTO by Rockwell Automation’s existing flatbed AMR platform to create a comprehensive autonomous material movement offering. This integration elevates AGILOX from a best-in-class point solution to a cornerstone of a world-class production logistics capability.

Strategically, the acquisition combines OTTO’s strong presence in North America with AGILOX’s established footprint across Europe. AGILOX was previously backed by private equity firm Carlyle, which made a majority growth investment in 2021, during which time AGILOX quadrupled its revenue. Rockwell noted that the deal will more than double its AMR business and provide AGILOX access to Rockwell’s deep customer relationships in strategic verticals such as consumer packaged goods and life sciences.

Blake Moret, Chairman and CEO of Rockwell Automation, said that combining the technologies creates a best-in-class platform setting a new standard for autonomous material movement. Upon closing, AGILOX will report to Rockwell’s Intelligent Devices operating segment.

출처:Rockwell Automation

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

It needs to be pointed out that this is an attribution argument made during a forum speech, not a peer-reviewed research conclusion. Rolnick's distinction between "elephants and bees" is a useful heuristic framework, but the boundaries between the two types of AI are not clearly distinguishable in all tasks. Some climate problems may require the general world knowledge of large models to integrate cross-domain information, in which case "bees" may not be sufficient. His core contribution lies in reminding practitioners: choose tools based on the problem, rather than looking for application scenarios based on the model.

These three obstacles collectively point to a conclusion: The autonomy of AI in semiconductor manufacturing is not a single technical problem, but a verification infrastructure problem. Fried admitted that teams spend more time deciding whether to use full physical simulation, surrogate models, machine learning, deep learning, or agentic AI to solve specific problems than they do actually building and implementing these systems. The packaging inspection case presented by Ingu Yin Chang, Executive Vice President of ASE Group, illustrates the current status: AI can identify a thousand shades of gray difference in X-ray images, but the system can only trigger equipment responses based on this once humans physically confirm through traditional manufacturing methods that a certain shade corresponds to a real defect. Between identifying anomalies and autonomously correcting them lies a whole set of verification, governance, and traceability mechanisms that have not yet been built.