독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-22168 게시물

새로고침

Delft University Team Develops MonoRace: AI Drone Beats Human Champions at Over 100 km/h Using Single Camera

중대
2026-10-22 07:20 GMT+8

A research team from Delft University of Technology in the Netherlands has unveiled a breakthrough result: their artificial intelligence drone system, named "MonoRace," defeated professional human pilots in the 2025 Abu Dhabi Autonomous Racing League. The study was published on October 14, 2026, in the journal Science Robotics.

The core advantage of the MonoRace system lies in its minimal hardware dependencies. It operates without GPS, external motion capture systems, stereo cameras, or pre-loaded track maps. Instead, it relies solely on a single front-facing rolling-shutter camera, an inertial measurement unit (IMU), and a small onboard computer to perceive the environment and control flight in real time. In the final race, the system completed two laps of an 11-gate course in 16.56 seconds, reaching a top speed of approximately 102 km/h, surpassing all competing human champions.

This performance was achieved through reinforcement learning-based training. Given the prohibitive cost of millions of trial-and-error flights in the real world, researchers conducted extensive training in simulation environments. Guido de Croon, a professor at Delft University of Technology and co-author of the paper, noted that the system learned to fly from scratch. Initial training focused on obstacle avoidance and basic navigation, gradually progressing to higher speeds and complex maneuvers. To enhance generalization, the training process randomized variables such as drone weight, voltage, drag, and sensor noise, ensuring the neural network could adapt to physical differences across various hardware units.

In actual deployment, MonoRace was mounted on a lightweight drone weighing 966 grams, with its internal micro-neural network issuing motor control commands 500 times per second. This achievement not only demonstrates AI's capability for real-time decision-making in high-speed dynamic environments but also provides significant empirical evidence for the development of "physical AI": robots can achieve precise motion control exceeding human capabilities even under constrained sensor and computational resources.

출처:zdnet.co.kr

리서치

Gartner Forecasts Over $3.5 Trillion in AI Infrastructure Investment at Risk of Impairment by 2030

주제 · AI基建投资与营收缺口구조적
2026-10-22 13:52 GMT+8

Research advisory firm Gartner released a forecast on October 14, 2026, indicating that the artificial intelligence industry faces significant pressure to recoup investments. Using a 12% return on invested capital (ROIC) as the threshold for viable investment, Gartner estimates that AI data center operators need to generate a cumulative $8.5 trillion in revenue by 2030 to cover announced capital expenditures and financing costs while achieving this return. However, enterprise and consumer AI revenue attributable to AI data centers during the same period is projected to be at most $3.6 trillion, leaving a gap of $4.9 trillion.

Will Sommer, Senior Director Analyst at Gartner, stated that AI companies are expected to invest nearly $1 trillion in building AI infrastructure in 2026, with an additional $6 trillion planned between 2027 and 2030. Because most AI-native companies operate with asset-light models, firms owning large amounts of AI infrastructure will capture less than half of the AI revenue. According to the analysis, total revenue from AI software, services, and consumer products is estimated at only $4.4 trillion by 2030, with at most 30% available to recover underlying capital investments.

Gartner warns that if AI companies maintain their current pace of capital expenditure, more than $3.5 trillion in AI infrastructure spending could face asset impairment risk by the end of 2030; if investment continues along its current trajectory, assets facing write-down risks could exceed $4 trillion by 2032. Additionally, Gartner predicts that cost pressures and undifferentiated business models will force over 80% of Neoclouds (new cloud service providers centered on AI computing) to cease independent operations by 2030, leading to market consolidation before that year.

출처:TechOrange 科技報橘

리서치

Meta Pauses Muse Sharing After Security Firm Reports DNS Flaw Exposed Chats

중대
2026-10-22 10:10 GMT+8

US tech company Meta has temporarily suspended the conversation-sharing feature of its AI assistant, Muse. The decision follows a report by security firm Claroc, which identified a design flaw in how Muse generated shareable links. Originally designed to allow anyone with the link to view the content, the system embedded conversation titles directly into the URL's subdomain. Researchers demonstrated that these hostnames could be harvested from public passive DNS history, allowing third parties to reconstruct valid links to private chats.

Claroc stated that they recovered approximately 6,800 shared links through a single query and accessed hundreds of pages. While not all contained sensitive information, many exposed personal and business secrets, including internal revenue reviews, loan applications, and confidential notes prepared for calls with AI lab executives. The firm termed these externally discoverable pages 'GhostShares'.

Meta confirmed the pause and stated it is reviewing the report to adjust link discoverability. The company emphasized that only user-initiated shares were affected, not all private Muse conversations. Claroc noted that while the pages are now inaccessible, some hostnames may remain in passive DNS records, indicating the privacy risk may not be fully eliminated.

출처:TechNews 科技新報

리서치

Ex-OpenAI Research Chief Launches AI Manufacturing Startup Arda with $230M Funding at $1.5B Valuation

주제 · Arda融资动态중대
2026-10-22 08:10 GMT+8

According to TechNews, Arda, an AI manufacturing software startup founded by former OpenAI research chief Bob McGrew, was unveiled on October 20 and announced the completion of a $230 million funding round, reaching a $1.5 billion valuation. The round was co-led by Altimeter Capital, Accel, and Khosla Ventures.

Arda develops software combining user interfaces with AI agents to automate parts of the manufacturing process. McGrew stated that the product serves manufacturing engineers similarly to how Claude or Cursor serve programmers. When designs for weapons or drones change, adjusting automated equipment traditionally takes three to six months, whereas Arda's software helps engineers set up assembly lines faster. Defense technology company Anduril is already using Arda's AI models.

Technically, Arda combines proprietary models, open-source models, and a few cutting-edge models. Co-founders include Augustus Odena, formerly of Google and Meta and involved in founding Adept, along with two former Palantir engineers, Alex Mark and Jakob Frick. McGrew revealed that robotics is the company's next step.

Although foundation models are becoming widespread, factory AI must handle sensor data, robotic systems, and supply chain variables, making it far more complex than training language models on web text. Many industrial AI startups have previously stalled at the pilot stage; whether Arda can successfully move toward mass production deployment will be key to validating its $1.5 billion valuation.

출처:TechNews 科技新報

리서치

Amazon Stockpiles Billions in NVIDIA GB300 Racks Due to Power Bottlenecks, Turning Free Cash Flow Negative

주제 · 亚马逊AWS电力瓶颈중대
2026-10-21 16:30 GMT+8

According to a report by Business Insider citing Amazon internal documents and recordings of AWS employee meetings, Amazon Web Services (AWS) is facing a severe power supply bottleneck. Although Amazon imports billions of dollars worth of NVIDIA's latest GB300 rack systems every month, a lack of connected power at data centers prevents timely deployment, forcing the company to temporarily stockpile large amounts of equipment in intermediate warehouses.

To gain an edge in the fierce AI chip competition, Amazon launched a new "Early Rack Receipt" process in August 2026, importing approximately $5 billion in equipment that month alone, with plans to purchase about $3 billion in GB300 racks monthly thereafter. This model shifts away from the traditional pay-after-installation structure, creating a financial situation where capital is spent upfront while revenue generation is delayed. The AWS supply chain team acknowledged there is currently no clear path to installing these racks into data centers. In response, Amazon stated that completing equipment deployment before power is ready is an industry norm.

This backlog has directly impacted financial performance. During internal meetings in September, AWS CFO John Felton and Head of AI Infrastructure Prasad Kalyanaraman discussed the company's negative free cash flow (FCF). Kalyanaraman noted that the AI business faces massive cash flow losses, expected to remain negative for approximately the next two years. Data shows that as of the second quarter of 2026, Amazon's trailing twelve-month free cash flow deficit reached $7.6 billion, while capital expenditures for 2026 are projected at $220 billion. Amazon originally expected to resolve the shortage by 2027 but now predicts demand will continue to exceed deployed capacity through that year.

Because GPUs sitting in warehouses cannot generate computing power, the true scarcity in the market is energized compute resources, which has further driven up rental prices. Goldman Sachs data indicates that the index for on-demand B200 GPU rentals has risen to $6.01 per GPU-hour, up approximately 40% from the low of $4.29 in January 2026, approaching the historical high set in March. This suggests that significant upfront capital investments require months or even years to translate into actual output, with the supply-demand imbalance unlikely to ease in the short term.

출처:TechNews 科技新報

리서치

OpenAI Reports 97.9% Accuracy in Legacy Code Migration Benchmarks with New Models

중대
2026-10-21 14:40 GMT+8

US artificial intelligence company OpenAI announced on October 19 that its latest generation models demonstrate significantly better performance in legacy programming languages such as COBOL, Fortran, and Visual Basic. These languages have less training data available but still support critical infrastructure in finance and government sectors globally.

According to TechNews, OpenAI designed two internal evaluations. The first, MultiLangBench, required models to build a SQLite-compatible database engine from scratch using 20 different programming languages. Results showed that GPT-6 Astra achieved an average pass rate of 87.4% under maximum reasoning settings, far exceeding the previous generation GPT-5.6 Sol's 26.8%. Specifically, COBOL performance improved from 10.9% to 72.1%, Fortran from 21.4% to 83.5%, and Visual Basic from 27.5% to 89.4%.

The second evaluation focused on code migration, requiring models to rewrite COBOL programs into C++, Java, Kotlin, Python, or Rust while maintaining original behavior. OpenAI stated that these tasks involve more than simple translation; they require preserving original calculation logic and business rules. Hidden tests were used to check for byte-for-byte consistency. Data indicated that GPT-6.1 Sol achieved 86.6% accuracy at low reasoning settings, up from 5.4% for GPT-5.6 Sol; with increased reasoning intensity, accuracy rose to 97.9%.

OpenAI mentioned that the lower-cost version, GPT-6 Luna, also achieved an average pass rate of 40.3% on MultiLangBench, with lower estimated API costs per test sample compared to the previous generation. However, the company emphasized that these results represent improvements in a test environment. For actual production use, expert review and careful verification against specific business rules and dependencies are still required.

출처:TechNews 科技新報

리서치

US Semiconductor Firm Allegro Launches 48V Power Management Chip, Claims 50% Cost Reduction for Humanoid Robot Systems

주제 · Allegro机器人芯片간단한 요점
2026-10-21 15:03 GMT+8

As humanoid robots accelerate in development, US semiconductor company Allegro MicroSystems announced on October 21, 2026, an expansion of its robotics product portfolio. The launch includes a 48V power management integrated circuit (PMIC) and high-precision position sensors targeting joint control and functional safety needs in humanoid robots. The vendor claims that the new A81481 PMIC can reduce the footprint and material costs of related power systems by approximately 50%.

Compared to stationary industrial robots, humanoid robots must integrate numerous motors, sensors, and control components within extremely limited physical space, placing higher demands on power efficiency and component size. As 48V power architectures are gradually adopted to reduce cable weight and power transmission losses, Allegro's A81481 can directly provide system power from a 48V bus without requiring additional front-end voltage regulators, simplifying circuit design. Housed in a 6x6 mm package, the chip integrates two buck converters, five low-dropout (LDO) regulators, and a triple watchdog diagnostic mechanism supporting ASIL-D level functional safety designs.

Beyond power management, Allegro also introduced inductive position sensors for high-torque joints such as shoulders and hips, alongside tunneling magnetoresistance (TMR) and Hall-effect angle sensors for fine motor control in hands and fingers. In terms of application partnerships, Allegro has collaborated with robotics firm Sarcomere Dynamics to integrate these controllers, sensors, and PMICs into the Scorpion robotic gripper, with plans to expand to the next-generation ARTUS Dex robotic hand featuring 24 degrees of freedom.

출처:TechNews 科技新報

리서치

Google Open-Sources TorchTPU: Enables Native PyTorch Execution on TPUs to Challenge NVIDIA Ecosystem

주제 · 英伟达AI生态垄断중대
2026-10-21 11:54 GMT+8

On October 21, 2026, U.S.-based Google announced at the North American PyTorch Conference that it has officially open-sourced a project named TorchTPU. This tool allows developers' existing PyTorch artificial intelligence applications to run directly on Google's custom Tensor Processing Units (TPUs) without rewriting underlying code. The move is seen as a key step by Google joining competitors like Amazon in an attempt to break the long-standing dominance of NVIDIA's AI hardware and software ecosystem.

PyTorch is one of the most popular open-source machine learning frameworks, with data suggesting that approximately 85% of AI applications are built upon it. Since PyTorch was originally developed by Meta in 2016 primarily for NVIDIA GPUs, many of its functions default to relying on the CUDA ecosystem. TorchTPU provides native support, enabling existing PyTorch-based applications to be easily ported to TPUs and Google Cloud services; developers can achieve this simply by changing basic configurations.

Notably, just two days prior on October 19, U.S.-based Amazon also launched a similar tool called TorchNeuron, designed to allow AI applications to migrate to its custom Trainium chips and AWS platform without re-coding. The successive actions by these tech giants indicate that competition in AI chips has extended from pure hardware performance and cost to control over the software ecosystem. In the past, NVIDIA established a deep moat through the tight binding of CUDA and PyTorch; now, Google and Amazon are attempting to attract more workloads to their custom chips by lowering migration barriers.

Beyond basic compatibility, Google emphasized close collaboration with Meta, Hugging Face, and the development teams behind vLLM and SGLang to ensure that TorchTPU's function libraries integrate smoothly with original PyTorch. Additionally, Google released the TPU agent suite, a helper tool to assist developers in fixing errors and optimizing performance during the migration process. While this initiative lowers technical barriers, actual performance and stability in production environments still require broader independent verification.

출처:TechNews 科技新報

리서치

NVIDIA Customizes HBM Base Dies via TSMC, Experts See New Moat

주제 · 英伟达NVHBM标准化중대
2026-10-22 11:15 GMT+8

According to reports from South Korea's Seoul Economic Daily, NVIDIA is advancing its next-generation customized High Bandwidth Memory (HBM) solution, known as NVHBM. The plan requires the three major DRAM manufacturers—Samsung Electronics, SK Hynix, and Micron Technology—to uniformly adopt the most advanced 6th-generation process technology for producing DRAM core dies. As part of this initiative, Samsung will introduce an external foundry partner for the first time in its HBM manufacturing stage, with Taiwan Semiconductor Manufacturing Company (TSMC) contracted by NVIDIA to produce the HBM base dies using its 3-nanometer process.

Semiconductor industry experts note that this arrangement is not Samsung independently ordering from TSMC, but rather NVIDIA directly commissioning TSMC to manufacture the base dies. This strategy aims to reduce performance discrepancies and validation burdens among HBM products from different suppliers. Previously, SK Hynix and Micron had collaborated with TSMC on base dies, while Samsung largely relied on its own process. By mandating uniform DRAM processes and specifying TSMC for base dies, NVIDIA seeks to control data flow efficiency between GPUs and DRAM through standardization.

Analysts suggest that amid severe memory shortages driven by global AI demand, NVIDIA's dominant position in the data center ecosystem allows it to compel these memory giants to align with its customization scheme. If NVHBM volumes increase, it could crowd out access to HBM capacity for other camps such as Advanced Micro Devices (AMD), Intel, and Application-Specific Integrated Circuit (ASIC) designers. This deep integration with upstream capacity and manufacturing standards is viewed by experts as a potential new competitive moat for NVIDIA against rising rivals. Details remain based on media reports and expert speculation, without full technical disclosure from NVIDIA.

출처:TechNews 科技新報

리서치

Musk Attempted to Acquire TSMC Wafer Fab for Chip Manufacturing Plan Terafab, Pivots to Partnership After Rejection

주제 · Terafab芯片项目중대
2026-10-21 15:10 GMT+8

According to a report by Korean outlet THEELEC, Tesla and SpaceX CEO Elon Musk attempted to directly acquire wafer manufacturing facilities from Intel, TSMC, and Micron before launching his ambitious Terafab chip manufacturing initiative. Following these rejections, Musk pivoted to seeking cooperation with global semiconductor giants.

Intel has already formed a direct alliance with Terafab, assisting the project through its Intel 14A manufacturing technology. However, the partnership with TSMC remains contentious. While earlier reports suggested negotiations were underway, Musk later clarified that TSMC could only serve as a "subcontractor" in the plan. In response, TSMC President C.C. Wei confirmed discussions with Musk during an earnings call but emphasized that TSMC must "personally operate the wafer fab" for any collaboration to proceed.

The Terafab plan aims to reach an initial capacity of 100,000 wafers per month by 2030, with a long-term goal of 1 million wafers per month. For context, TSMC alone expects its leading-edge 2-nanometer process capacity to reach 120,000 wafers per month by the end of 2026. Industry observers note that despite Terafab's massive targets, its share of global chip manufacturing remains limited.

Musk's primary motivation for building Terafab is to secure sufficient chip supply for Tesla's Full Self-Driving (FSD) vehicles and Optimus humanoid robots. This episode highlights the practical constraints facing AI giants attempting vertical integration: even with capital and clear demand, bypassing traditional foundries' insistence on operational control proves difficult.

출처:TechNews 科技新報

리서치

Former OpenAI Safety Transparency Lead David Robinson Resigns, Argues Company’s Safety Culture Is Broken

주제 · OpenAI安全文化争议중대
2026-10-05 19:50 GMT+8

David Robinson, former OpenAI Safety Transparency Lead, resigned after more than three years and published an essay in The Atlantic stating that the company’s safety culture is broken. He criticized OpenAI’s reactive approach to safety, which focuses on fixing problems only when they emerge, arguing that this stance effectively guarantees failures.

Citing incidents such as the July 2026 HuggingFace hack executed by an AI model and a recent failure of an AI “kill switch” to stop a rogue agent, Robinson said Silicon Valley lacks “wisdom about what it means to care for people.” He wrote that this moment needs a degree of humility that isn’t natural for people who have succeeded through their extreme confidence.

He argued that the industry should rely on safety experts already existing in fields like nuclear engineering or aviation. These industries learned about safety, redundancy, and planning through disasters that claimed thousands of lives across hundreds of incidents, building systems and procedures to ensure no single machine failure or human error results in tragedy. Robinson contends that OpenAI and other labs are deploying frontier AI with far less redundancy and rigor, even though the harm from an irreversible loss of control would be much greater than any single meltdown.

Robinson did not focus on rules and regulations but targeted the core culture of safety and OpenAI’s penchant for “iterative deployment,” commonly known as trial and error. He stated that he should have stayed inside the company to fight for a fundamental shift in thinking, but this was next to impossible due to how fast things were moving, so he left to tackle the problem from the outside.

출처:Tom's Hardware

리서치

The Wall Street Journal's AI Tool Doppler: From Polling Numbers to the Context of Public Discourse

중대
2026-10-22 12:57 GMT+8

Following the conclusion of the 2024 United States presidential election, traditional polling was once again proven unable to accurately capture the public's true concerns. Sampling biases from telephone interviews and automated voice response (ARS), coupled with an expanding non-response group, alongside media coverage focused on 'who leads by how many points' in a horse-race format, collectively created a structural blind spot: what people are actually discussing, and in what context, cannot be answered through a multiple-choice question.

The United States-based The Wall Street Journal (WSJ) launched the AI analysis tool Doppler News AI ahead of the 2026 United States midterm elections, attempting to address this issue from another direction. Instead of asking voters questions, Doppler enters the spaces where people engage in daily conversation—social media and podcasts—to build a discourse database. According to ZDNet Korea's retelling of Nieman Lab reporting, this database includes over 25 million TikTok videos, 500,000 podcast episodes, as well as audio content from Instagram videos and cable television broadcasts.

The technical core is Vector Embedding: mapping the emotions and themes contained within text and video into a high-dimensional numerical space. This differs from keyword search—keywords can only find content that explicitly mentions a specific word, whereas vector embeddings can capture semantic aggregation and divergence. Through this, Doppler constructs a new form of visualization, showing where and in what manner public discourse aggregates.

A specific case reveals what this method can uncover. WSJ used Doppler to track the topic of 'opposition to AI surveillance cameras,' analyzing 4 million TikTok videos and podcast data. The results showed that content creators discussing this topic had a disagreement rate of 12% on typical political issues, which is three times the disagreement rate for general political topics. In other words, on the issue of AI surveillance, groups that usually hold opposing stances instead exhibited unusually consistent opposition. Traditional polling's approval/disapproval ratios cannot present the formation of such cross-party consensus.

Another case involves prices. Former President Donald Trump dismissed the 'cart crisis' (grocery price inflation) as fake news, but Doppler's analysis of over 3,000 hours of audio data indicated that right-wing podcast hosts have recently discussed price issues with a frequency approaching that of left-wing podcast hosts. Carefully crafted phrasing by political figures and standardized questionnaire items cannot capture the flow of this 'life-adjacent public opinion.'

Perhaps the most notable design choice of Doppler is not the technology itself, but that it is not an autonomous agent. As retold by ZDNet Korea, the biggest problem with AI analysis is hallucination and error; Doppler overcomes these structural limitations through human intervention and transparent design. WSJ's Doppler does not automatically write and publish articles. Instead, it is a collaborative system composed of data journalists, machine learning engineers, and political desk reporters: together they propose hypotheses weekly, decide what questions to ask the data, and continuously adjust the model's classification standards. As summarized in Nieman Lab reporting, 'the usefulness of AI depends on the level of questions you ask.'

출처:zdnet.co.kr

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

Transparency is another key design element. At the end of every report using Doppler, WSJ explicitly lists the algorithmic principles, data extraction methods, and known limitations of the AI model. This is not a decorative disclaimer, but part of the methodology itself: when analytical tools shift from 'measurement' to 'exploration,' readers need to know the boundaries of the exploration to judge the credible scope of the conclusions.

It must be noted that all the specific figures above—the 25 million videos, 500,000 podcast episodes, 12% disagreement rate, and 3,000 hours of audio—come from ZDNet Korea's retelling of the Nieman Lab article. The original Doppler system's embedding model selection, cluster validation methods, and sample definitions have not been independently verified. Doppler is currently in an experimental phase, serving the 2026 United States midterm election coverage, and whether its methodology can be generalized to other fields remains to be seen.

However, the Doppler experiment demonstrates at least one thing: the value of AI in public discourse analysis lies not in replacing human judgment, but in freeing humans from the measurement mode of 'counting heads,' shifting instead toward understanding why people care, what they care about, and in what contexts consensus is reached or lost. The prerequisite is that humans remain responsible for asking the right questions.