독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-29198 게시물

새로고침

Fed Holds Rates at 3.75%-4%, Two Members Vote for Hike

주제 · 美印H-1B争端구조적
2026-10-29 02:00 GMT+8

The Board of Governors of the Federal Reserve System issued a statement on October 28, 2026, announcing that the Federal Open Market Committee (FOMC) decided to maintain the target range for the federal funds rate at 3.75% to 4%. The decision was approved by a 10-2 vote, with Lisa D. Cook and Beth M. Hammack dissenting; they preferred to raise the target range by 25 basis points at this meeting.

The statement noted that economic activity is expanding at a strong pace. While uncertainty remains elevated due partly to geopolitical developments, domestic spending has been resilient. The committee highlighted that productivity growth is strong and capital investment is robust. Job gains have kept pace with the workforce, and the unemployment rate has changed little.

However, inflation remains elevated. The Federal Reserve stated that the accumulation of more information will help determine the policy stance appropriate to ensure a timely return of inflation to the 2% goal. This indicates that while there is internal disagreement on immediate tightening, the committee is currently waiting for further data to assess whether supply-side strength can offset inflationary pressures.

출처:Board of Governors of the Federal Reserve System

리서치

Roblox Admits Researcher's Bot Scraped 145GB of 3D Models from Open3DLab, Denies AI Training Use

주제 · Roblox爬虫数据争议중대
2026-10-27 21:31 GMT+8

US gaming and creation platform Roblox responded on October 27, 2026, to accusations, admitting that a web crawler built by one of its researchers mistakenly scraped data from free 3D asset websites Open3DLab and its sister sites SFMLab and SmutBase between October 15 and 16. According to disclosures by Open3DLab operators, the crawler made approximately 26,000 requests over 34 hours, downloading 1,027 files across 712 projects, totaling 145.5GB, including 62.6GB of adult content.

A Roblox spokesperson told Fathom News that the agent was intended to collect publicly available 3D assets but failed to operate as designed. The company stated the agent has been decommissioned and "the data was not used for any purpose, including training any AI models." However, this claim is self-reported by the company, and there is no independent evidence confirming the data has been completely deleted or that no derivative works were produced.

The dispute centers on data usage and copyright compliance. Open3DLab noted that 423 of the scraped projects carried non-commercial-only licenses, and the crawler ignored robots.txt protocols and terms of service. Since Roblox is developing a generative 3D AI model named "Cube," and the crawler's user-agent string "trellis2-data-sourcing/1.0" points to Microsoft's open-source TRELLIS.2 project, speculation arises that the batch may have been used for training or testing AI models.

This event occurs against a backdrop of multiple regulatory pressures on Roblox. The European Commission designated Roblox as a "very large online platform" in August, requiring systemic risk assessments; simultaneously, Roblox faces child safety lawsuits in several US states, while more than one-third of its daily active users are under 13 years old. The scraping of adult content further intensifies public questioning of its data governance and ethical boundaries.

출처:Fathom News

리서치

GitHub launches Copilot Fusion: Automated model orchestration claims parity with Claude Code at half the cost

중대
2026-10-29 00:19 GMT+8

GitHub officially launched GitHub Copilot Fusion on Wednesday at GitHub Universe in San Francisco. This new feature is designed to automate the selection and orchestration of large language models within its coding assistant, aiming to relieve developers of the burden of manually choosing which model to use for different types of coding tasks, such as reasoning, code generation, debugging, or tool use.

Copilot Fusion combines the runtime model orchestration capabilities previously seen in Project HydraFusion with GitHub's existing Copilot Auto Model Selection. According to GitHub, this integration allows the system to dynamically route requests to the most appropriate model. For example, it can employ a "cascading" workflow where one model generates code and another evaluates it, or a "critique" workflow where models review each other's work. Developers retain control over the effort level—Standard, High, or Max—but the underlying model selection is handled by the system.

In a statement provided to The New Stack, Hari Pokhrel, principal software engineer at AT&T, noted that developers often inadvertently choose overly powerful models for simple tasks, leading to inefficiency. "Previously, developers would often choose a more powerful model for a complex task and forget to switch back for everyday activities," Pokhrel said. He credited Copilot Fusion Standard with improving token efficiency at AT&T and reducing unnecessary spending.

GitHub claims that Copilot Fusion Max delivers benchmark performance comparable to Anthropic's Claude Code with Opus 5.5 Max, but at about half the average API cost per task and in about 40% less time. On the Artificial Analysis Coding Agent Index, GitHub reports that Fusion Max scored 60, within two points of Claude Code Opus 5.5 Max. These figures are self-reported by GitHub and have not been independently verified. The system is generally available now as part of GitHub's broader strategy to make its platform a comprehensive software delivery system where AI handles routine decisions while humans focus on direction and outcomes.

출처:The New Stack

리서치

AI Startup Aseon Labs Develops Autonomous Vehicle Cleaning and Charging Pods with 10-Person Team and $10M Funding

주제 · Aseon Labs간단한 요점
2026-10-29 01:33 GMT+8

According to Semafor, AI startup Aseon Labs is developing parking space-sized pods equipped with robot arms designed to clean and charge autonomous vehicles. CEO George Kalligeros stated that the company, with a team of only 10 people and $10 million in funding, is accomplishing tasks that previously would have required years and a massive workforce.

Aseon Labs heavily utilizes AI tools for development and programming, with engineers even using ChatGPT prompts to program the robots. Kalligeros revealed that the company spends six figures (over $100,000) per month on AI tokens. This high reliance on AI demonstrates how startups in the AI era can leverage technology to achieve complex hardware innovations with fewer personnel and less capital.

출처:semafor.com

리서치

Investcorp and Saudi Investors Acquire Fire Safety Firm REDA Gulf Ahead of Planned IPO

주제 · REDA Gulf간단한 요점
2026-10-28 21:20 GMT+8

Bahrain-based investment firm Investcorp has joined a consortium led by Thara Future Investments and Saudi family office AlTouq Group to acquire REDA Gulf, an industrial fire safety and security business. The buyout positions the company for an initial public offering (IPO) within the next few years.

Thara Future Investments was founded by Hisham Attar, a former executive at the Public Investment Fund. Attar stated that REDA is a leader in a highly regulated industry poised to benefit from increased focus on safety. This transaction marks one of the first deals involving a foreign investor entering the kingdom since the conflict with Iranian-backed Houthi militias escalated recently.

Investcorp participated through its $500 million Saudi Pre-IPO fund. The deal represents a bet on the recovery of Saudi Arabia’s IPO market, which had slumped due to poor post-listing performance and new regulations potentially deterring investment banks from underwriting.

출처:semafor.com

리서치

US Robot AI Firm Physical Intelligence Negotiates $1 Billion Funding Round at $20 Billion Valuation

중대
2026-10-28 16:55 GMT+8

According to QbitAI, US-based robotics AI startup Physical Intelligence (PI) is negotiating a new funding round aiming to raise at least $1 billion, with an expected post-money valuation of approximately $20 billion. Existing investors Avenir and Lux Capital are considering leading the round. If completed, PI's valuation would nearly double from earlier this year's $11.5 billion, compared to just $5.6 billion in November last year.

Notably, PI has generated no commercial revenue to date and does not manufacture its own robots. Its core positioning is developing a universal robot foundation model (VLA) capable of adapting to various third-party hardware, rather than building hardware itself. The founding team includes top scholars such as Karol Hausman, Sergey Levine, and Chelsea Finn from Google DeepMind, Stanford University, and UC Berkeley.

Investors are betting on its potential in physical-world foundation models. PI recently launched a beta API for its robot foundation model, offering fine-tuning and inference services to external developers, marking a shift from closed demos to an open ecosystem. This progress has fueled investor enthusiasm, positioning PI as the "OpenAI of embodied brains."

However, competition in the sector is intensifying. Figure AI has reached a $39 billion valuation, while Skild AI is valued at $14 billion after surpassing $100 million in annual recurring revenue. PI sticks to a pure-software path, avoiding heavy asset manufacturing, but deployment in the physical world remains extremely difficult, with errors carrying higher costs than in the digital realm. PI officially warns that real-world robotic applications require dedicating full company resources to conquer each niche domain.

출처:量子位

리서치

South Korea's Seoul Viosys Starts Mass Production of Rare-Earth-Free White Micro-LEDs for AI Optical Interconnects

간단한 요점
2026-10-28 17:52 GMT+8

South Korean semiconductor manufacturer Seoul Viosys announced on October 28 that it has begun mass producing micro-LED (Micro Light Emitting Diode) chips capable of emitting white light without the use of rare earth elements. The company stated that traditional white LEDs rely on rare earth phosphors such as YAG (Yttrium Aluminum Garnet) for packaging, whereas the new technology generates white light using only Gallium Nitride (GaN). This eliminates the need for rare earth materials and complex packaging processes, reducing manufacturing costs by half.

According to the company, the first mass-produced chips measure 25 micrometers by 10 micrometers and achieve a luminous flux of 0.015 lumens within the International Commission on Illumination (CIE) standard white color coordinates. The luminous efficacy approaches 100 lm/W, with a target of reaching 200 lm/W within two years. This development overcomes the technical challenge of achieving efficient white light at ultra-small scales, which was previously limited by packaging constraints that prevented white LEDs from shrinking below 100 micrometers.

Seoul Viosys plans to extend this technology to the field of large-capacity data transmission for Artificial Intelligence (AI). By integrating approximately 1,000 10-micrometer LED chips into a matrix structure within one square millimeter, the system transmits data using photons instead of electrons. The company claims that this optical interconnect solution can be 10 times faster than traditional copper wire communications while consuming only one-tenth of the power.

출처:zdnet.co.kr

리서치

US DOE Offers $5 Billion Loan to Constellation to Expand Nuclear Power for AI Data Centers

중대
2026-10-28 23:53 GMT+8

The Trump administration's Department of Energy (DOE) has provided a conditional loan commitment of up to $5 billion to energy company Constellation Energy to support the "uprating" of eight existing nuclear power plants, aiming to add 1.1 gigawatts (GW) of generating capacity. This move directly addresses the power bottlenecks faced by tech giants building AI data centers.

According to TechCrunch, Constellation plans to use the financing to upgrade 11 existing reactors, typically involving more highly enriched uranium and equipment upgrades to handle increased power levels. Rather than building new plants, this strategy maximizes the output of current assets. Most or all of the additional power from seven of these facilities has already been purchased by Google, Meta, or Amazon. For instance, Google is the confirmed buyer of extra power from a Pennsylvania plant and four Illinois plants; Amazon is buying electricity from the Maryland Calvert Cliffs plant; and Meta is effectively buying the entire Clinton power plant in Illinois.

Constellation stated that it had already committed to spending more than $7 billion to add 1,080 megawatts (MW) of nuclear capacity specifically to deliver power to Google, Amazon, and Meta. While the loan has yet to be finalized, Constellation noted it "could lower financing costs" for the projects. By reducing the cost of capital, the DOE's loans could boost Constellation's bottom line, benefit the hyperscaler buyers, or both.

Although the Department of Energy touted the loan on Tuesday as supporting "affordable, reliable power to American homes and businesses," there is no evidence that the new nuclear capacity will lower prices for residential ratepayers. Constellation’s announcement does not specify whether the projects will maintain or lower electricity rates, and the company declined to confirm the government’s claim to TechCrunch. For investors, this event highlights the tightening financial links between utility companies and tech firms in the AI infrastructure race, with the government playing a key lever role.

출처:TechCrunch

리서치

German Automaker Mercedes-Benz Group Cuts Full-Year 2026 Guidance, Lowering Passenger Car Margin Outlook to 2-3%

주제 · 宝马重组중대
2026-10-28 09:07 GMT+8

German automaker Mercedes-Benz Group AG lowered its full-year 2026 financial forecast in an ad-hoc announcement just before releasing its third-quarter results. According to reporting by TechNews, the company originally projected an adjusted Return on Sales (RoS) of 3% to 5% for its passenger car division but has now cut this expectation to 2% to 3%. Additionally, the outlook for Earnings Before Interest and Taxes (EBIT) was revised from "better than last year" to "in line with last year."

The primary drivers for this downgrade were identified as intensifying competition in the Chinese market and anticipated sharp increases in material costs during the fourth quarter. While the new guidance is more conservative than previous estimates, the report notes it remains slightly above some analyst consensus figures. Notably, the group's Industrial Business cash flow forecast stands at €172 million, significantly below the market expectation of €420 million.

In contrast, the commercial vehicle division remained resilient, maintaining a sales return on revenue of 9.4%. This suggests that the slower adoption of electric vehicles and mature service networks for internal combustion engines have shielded this sector from similar pressures. Despite the challenges, Mercedes-Benz expects full-year EBIT to reach €1.77 billion, a substantial improvement over last year's €580 million and better than 2024 performance. This indicates that after heavy investments in the MB.OS operating system and new electric platforms, the company has stopped the bleeding, though the profit landscape remains precarious amid price wars in China's luxury segment.

출처:TechNews 科技新報

리서치

Ai2's Open Model olmOCR 2 Reduces Archaeological Handwriting Transcription Errors from 15.58% to 1.11% After Fine-Tuning

주제 · olmOCR 2중대
2026-10-28 16:00 GMT+8

Non-profit AI research institute Allen Institute for AI (Ai2) announced in a blog post that researchers are adapting its open-source document image extraction model, olmOCR 2, for historical and archaeological archives. The model supports local execution, allowing users to perform keyword searches and structured text extraction without uploading sensitive materials to third-party services.

In a collaboration between TU Wien’s Computer Vision Lab and the Austrian Archaeological Institute at the Austrian Academy of Sciences, part of the LEGION project, researchers tested the model on handwritten pottery records from the Roman site of Carnuntum in Austria. These records comprise approximately 70,000 documents, with annotations often overlapping technical drawings and using abbreviated German terminology. The team created the CENTURIA benchmark dataset containing 507 annotated records.

According to the Ai2 blog, before adaptation, olmOCR 2 had the lowest transcription error among five tested open models. The research team used Low-Rank Adaptation (LoRA) to fine-tune the model on just 57 samples. On the remaining 450 test records, the transcription error rate dropped significantly from 15.58% to 1.11%. This result enabled accurate digitization of complex layouts and specialized terminology that were previously difficult to process.

Additionally, historian Maurice Casey developed ArtificeOCR, a tool connecting olmOCR 2 with Tropy, an open-source photo management application he uses. This allowed him to search thousands of photographs of letters and newspapers related to early 20th-century European radical politics. Casey noted the model performs well with standardized forms in secret police archives and cursive scripts across multiple languages, though he also reported errors when transcribing handwriting in languages less represented in the training data.

출처:allenai.org

리서치

Google Chrome Expands Enhanced Autofill to Use Gemini for Gmail Data Extraction

주제 · Chrome增强自动填充간단한 요점
2026-10-29 01:00 GMT+8

U.S. technology company Google announced on October 28, 2026, that its Chrome browser's "Enhanced Autofill" feature is expanding. The feature now uses the Gemini model combined with the "Personal Intelligence" service to suggest form-filling information sourced directly from users' Gmail accounts. Previously, the feature relied primarily on names, passport numbers, email addresses, and physical addresses stored in user Google Accounts and services like Google Wallet.

The expanded capabilities can retrieve details such as shipment tracking numbers, travel information like flight booking or hotel reservation numbers, and order numbers from recent purchases. Google states that sensitive data presented as suggestions while completing forms remains encrypted, and users can manage what Enhanced Autofill has access to in Chrome settings or turn it off completely.

The improvements are rolling out today in the United States for desktop, Android, and iOS versions of Chrome, available to users who have enabled "Personal Intelligence" within Gemini. This change aims to save users from having to search for confirmation emails, receipts, or shipping updates themselves by hopping between browser tabs.

출처:The Verge

리서치

Meta Announces Support for Five US Universities to Accelerate Medical and Scientific Breakthroughs with Muse Models

주제 · Meta Muse模型중대
2026-10-28 08:00 GMT+8

On October 28, 2026, Meta announced support for research at five US universities: California Institute of Technology (Caltech), The Ohio State University College of Medicine, Louisiana State University (LSU), Indiana University, and UTHealth Houston. The form of support varies by institution and project, including access to Meta's latest AI models, direct collaboration with Meta researchers, and research funding for select projects.

Researchers will build using the latest Muse models, including Muse Spark Open, Meta's most capable open-weight model, the Muse Wonder and Muse Spark series, and Muse Glimmer, a 30B open-weight model that can run on a single consumer GPU. Projects will also utilize computer vision models such as Segment Anything Model (SAM), V-JEPA, and DINO, as well as SeamlessM4T, a multilingual communication model from Meta.

In the medical field, a team at The Ohio State University is investigating how open-weight Muse models could help identify patients at risk of heart failure with preserved ejection fraction (HFpEF) before clinical diagnosis. Led by Dr. Yuan Xue, an Assistant Professor of Biomedical Informatics, the project combines routine abdominal CT scans with longitudinal medical records to better utilize information already collected during standard care.

In fundamental science, Caltech's partnership tests whether large language models can function as instruments for original research. Mathematician Dr. Tony Yue Yu is building MathDB Frontier, a benchmark based on over 86,900 research-level open problems, designed to evaluate if models can formalize conjectures, search for lemmas, and generate machine-checkable proofs rather than merely pattern-matching on training data.

Meta emphasized that participating universities will run models within their own institutional environments, independently conduct research, and retain full control of all patient and clinical data, which will not be shared with or processed by Meta. This initiative signals Meta's effort to embed its open-source model ecosystem into high-value academic and clinical workflows to validate their applicability to complex real-world problems.

출처:Meta AI Research

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크