読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-1448 投稿

更新

NVIDIA xBeam Moves Radio Interferometry Correlation to GPUs, But Does Not Perform Imaging

トピック · NVIDIA xBeamクイックテイク
2026-11-14 05:17 GMT+8

The Event Horizon Telescope (EHT) links 13 radio observatory stations located in Hawaii, Chile, Spain, Mexico, and Antarctica via very-long-baseline interferometry (VLBI). Each telescope records high-speed voltage streams in two polarizations at rates up to 64 Gbps, generating petabytes of data. Correlation processing—multiplying voltage signals from any two stations channel-by-channel and accumulating them—is one of the most time-consuming steps in the entire data pipeline. In a technical blog post released in November 2026, NVIDIA introduced how cuPhoton.xBeam moves this step onto GPUs.

xBeam adopts an FX correlator architecture, using the same mathematical framework as the open-source CPU correlator DiFX. Its pipeline consists of five steps: reading VDIF frames and building a .vdix index to support random access; performing integer alignment and fringe rotation on each station according to the delay model in the job file; executing an FFT on the voltage stream to split it into frequency channels; calculating the product of the complex spectrum of each station pair's channel with its conjugate and accumulating over the integration time; and finally writing visibilities in DiFX SWIN or xbeam-v1 format. Among these, the FFT and cross-multiplication are computationally intensive and independent across channels, making them naturally suitable for the GPU's parallel execution model.

Multi-GPU scaling is achieved by sharding observation time. On a single node, multiple GPUs are launched simultaneously using the --devices parameter, with each GPU processing a segment of integration time. Across nodes, each process specifies its own time shard using --shard-count and --shard-index. After completion, the merge-shards command verifies that all shards originate from the same job, cover the complete time range, and have consistent SWIN headers before merging. NVIDIA reported that processing 10 TB of data on a Vera Rubin NVL72 system using 64 GPUs and GPUDirect Storage took 70 seconds, approximately 460 times faster than the x86 CPU baseline of 31,992 seconds (about 8.89 hours); read bandwidth increased from 20 GB/s on 4 GPUs to 170 GB/s on 64 GPUs. These figures are benchmark results self-reported by NVIDIA, and the article did not mention independent reproduction.

The article also delineates what xBeam does not do. It does not calculate delays based on station coordinates but instead directly uses the delay polynomials provided in the job file; it does not trace light paths around black holes; it does not perform calibration or image reconstruction. Downstream processes are still completed by tools such as HOPS, rPICARD, and eht-imaging. The Gaussian splatting method mentioned in the article—modeling emitting gas as compact glowing clumps moving along orbits to fit visibilities—is explicitly labeled as an independent component not currently included in cuPhoton. The paragraph describing AI agents orchestrating end-to-end pipelines outlines potential application directions rather than implemented features.

ソース:NVIDIA Technical Blog

リサーチ

GeekPark's Zhang Yongyi: The Window for AI Recording Hardware May Be Only Six Months Long, as Smartphone Makers Integrate All-Day Memory into Systems

トピック · AI录音硬件窗口期重大
2026-11-13 12:57 GMT+8

On November 4, 2026, Apple introduced two new AI features to the Apple Watch in the watchOS 27.3 developer beta: Live Rewind, which allows users to review conversations from the past 15 seconds at any time, and Siri Recap, which records conversations throughout the day and compiles them into summaries. These features require no additional hardware and have virtually no learning curve; Apple states that audio is neither recorded nor stored, and speaker identification is not performed. GeekPark author Zhang Yongyi draws a conclusion from this: the core value of AI recording hardware is being eroded by smartphone manufacturers, as recording and transcription no longer form a competitive moat. What is truly scarce is filtered context and the Agent entry point. Before smartphone makers fully integrate all-day memory into their systems, hardware companies may have a window of only about six months.

This judgment rests on the premise that recording hardware originally filled gaps that smartphones could not hear. Smartphones spend most of their time in pockets, and the current Android architecture does not allow apps to independently activate the microphone from the background. Carl Pei, founder of Nothing, once illustrated this with an example: while having coffee, someone says something worth noting, but by the time you take out your phone, the moment has passed. Chen Yusen, CEO of Qwen Office, also noted that today's AI speech-to-text technology is accurate enough that recording is no longer necessary. In other words, the remaining task for hardware is primarily to capture sound and convert it into text. However, this gap is currently being bridged by system-level solutions: Apple has placed audio capture on the watch, while OPPO shares a memory engine between its Xinliqiu (Heart Sphere) device and the smartphone system. Since smartphones accompany users all day—housing accounts, messages, schedules, and payments—the final output from Agents must ultimately return to the smartphone screen for user review.

A more critical change is that the entity reading these records is shifting from humans to Agents. Zhang Yongyi shared his own practical usage data: after connecting recordings from the OPPO Xinliqiu to Feishu (Lark) and importing them into tools such as Codex and DeepSeek Harness, he filtered thousands of dialogue segments down to approximately 200. Of these, about 30% were background noise, 40% were pleasantries like 'I'm sharing my screen,' and 20% were context useful only at that specific moment, meaning effective information accounted for less than 10%. This implies that filtering may be scarcer than simply remembering more. Which Agent reads these records depends on who holds the true entry point to intelligence: system Agents from smartphone makers do not need to wait for others to open interfaces, whereas companies without phones can only read from the outside. For hardware companies, becoming 'hardware peripherals for big tech's Personal Agents' is becoming a more pragmatic position.

Based on this, Zhang Yongyi provides a timeline judgment: by mid-2027, at least one domestic smartphone manufacturer should have integrated all-day recording as a built-in system capability, eliminating the need for users to purchase separate recording hardware. He also notes boundaries: the domestic gap may be slightly longer than in the United States because Recap is not yet available in China, and products like Bee and Muse have not entered the market; the variable lies in whether Huawei, OPPO, vivo, or Xiaomi will first integrate all-day memory into their systems. For those considering buying hardware, professional recording cards remain effective for meeting recordings; however, for those considering adding a new device solely for all-day recording, it may be wise to wait, as smartphone makers' watches and earbuds may fill this gap faster. It took smartphones over a decade to absorb cameras, maps, wallets, and contact lists into a single device; all-day recording will likely follow the same path, appearing first as standalone hardware before being integrated into smartphone systems.

ソース:GeekPark

リサーチ

AgentCore Runtime V2 Snapshot Mechanism: Recommendations to Reverse Cold-Start Optimization

重大
2026-11-14 06:02 GMT+8

Traditional cold-start advice is to delay imports and minimize initialization so the first request triggers actual work as quickly as possible. In an engineering blog post published on November 13, 2026, Amazon Web Services states that this advice needs to be reversed for AgentCore runtime V2: initialization runs only once and is frozen into the snapshot, while deferred static work in the handler incurs a cost with every new session.

The V2 mechanism works as follows: when creating or updating a runtime, the platform starts the container, waits for the health check to pass (within 120 seconds), clears the operating system file cache, freezes the microVM, and saves memory and disk state as a snapshot. Subsequently, each new session no longer pulls the image or runs initialization code; instead, it restores the snapshot into a new microVM, allowing the process to resume from the frozen point, with imported libraries, created objects, and clients already present in memory.

This brings three direct implications. First, complete memory restoration: any data held during the initialization phase, including content on RAM-backed file systems, becomes available at memory speed from the first access. Second, file reads do not carry over: because the file cache is cleared before freezing, files read during initialization must still be fetched from the image in restored sessions. Third, image files are loaded on demand: the first time a specific data block is read, it is pulled from snapshot storage and then cached closer to the host.

Therefore, the correct refactoring direction is to move static work into initialization. Create objects rather than merely importing modules, because object creation loads their dependencies and keeps them in active memory. Create boto3 clients during initialization; endpoint resolution and credential resolution are restored with the snapshot, and sockets automatically reconnect on the first call. Read static data such as prompts, configurations, and small indexes into memory-held objects rather than reading from the image with every request. For files that cannot be held within the process—such as independently running CLI toolchains or browser binaries—they can be copied to RAM disks like /dev/shm, or pre-warmed using build hooks during snapshot construction.

However, there is a trade-off that must be measured: data retained in memory during initialization increases snapshot size, which in turn increases restore time, and you pay for that memory in every session. The article explicitly recommends measuring end-to-end session start times before and after changes to confirm that latency is indeed reduced rather than merely shifted from the handler to the restore phase.

At the same time, certain values must remain in the handler. Random numbers, timestamps, credential refreshes, dynamic tool catalogs, and feature flags—if created during initialization, they would be frozen into the snapshot, causing every session to carry identical random states or expired time bases. The article draws an analogy to known pitfalls in AWS Lambda SnapStart and offers a module-level caching pattern with expiration times as a compromise for dynamic data.

ソース:Amazon Web Services

リサーチ

AWS Launches Kiro Android App for Mobile Cloud Coding Sessions and Integrates Grok 4.7

トピック · AWS Kiro安卓应用重大
2026-11-13 11:05 GMT+8

Amazon Web Services (AWS) launched 'Kiro for Android' on November 13, 2026, an application for its AI development environment Kiro. The app enables paid users to start, check, and continue instructions for AI coding tasks running in the cloud via smartphone. This update addresses the limitation of managing long-running AI agent tasks when away from a desktop.

The core mechanism is the 'Cloud Session'. All coding work executes in a cloud sandbox, with conversation history and state synchronized in real-time across Kiro CLI, Web interface, IDE, and the Android app. Background tasks do not stop even if the phone screen is off. On Android, users can view lists of running, completed, or pending-response sessions. New tasks can be started by connecting a repository and entering instructions, with support for attaching screenshots. When Kiro requires a decision (such as branch selection), users can reply directly on their phone. Change diffs are displayed in mobile-optimized cards, and upon confirmation, Kiro automatically creates a Pull Request.

For enterprise customers using IAM Identity Center, administrators must manually enable cloud sessions. Currently, the service is only available in the US East (N. Virginia) region, with an iOS version in preview.

Additionally, AWS announced the integration of SpaceX's 'Grok 4.7' model into Kiro's IDE, CLI, Crew, and Web interfaces. According to SpaceX-published data, compared to the previous Grok 4.6, Grok 4.7 improved success rates on Terminal-Bench 4.0 from 20.3% to 37.6%, and on DeepSWE v1.1 from 65.2% to 71.0%. The model is designed for tasks taking hours that require step-by-step verification, compatible with Kiro's 'spec-driven development' mode. Pricing is $2 per million input tokens and $6 per million output tokens, with a 500,000-token context window. In Kiro Pro and higher plans, Grok 4.7 is offered as an experimental feature consuming 2.7x credits.

ソース:テクノエッジ TechnoEdge

リサーチ

AMD Engineers Port 50 Models to Client GPUs in Three Days Using HRX Loom and AI Agents

トピック · AMD AI代理模型移植クイックテイク
2026-11-13 23:53 GMT+8

Engineers Jeremy Fowers and Ben Vanik at U.S. semiconductor company AMD led an internal experiment leveraging the newly released ROCm HRX Loom v0.1 software stack and an AI agent workflow to port 50 popular Hugging Face models across 18 architectures to AMD client GPUs in approximately three to four days.

The experiment aimed to demonstrate the potential of accelerating the adaptation of new and niche models to AMD hardware through HRX Loom and agentic AI. According to AMD's report, the process achieved a medium speed-up of 1.5 to 2.1 times for 32K token prompt processing compared to Llama.cpp's ROCm backend, and a 1.9 times speed-up compared to its Vulkan backend.

HRX Loom remains an experimental project supporting CPU, GPU, and NPU execution. These results highlight AMD's effort to narrow the compatibility gap between its software ecosystem and mainstream open-source models via automated toolchains, though the specific performance figures have not been independently verified.

ソース:phoronix.com

リサーチ

US Corporate Pension ROA Assumptions Stabilize at 5.9%: Interest Rate Cycles and Asset Allocation Logic

トピック · 美国养老金ROA假设重大
2026-11-12

The Return on Assets (ROA) assumption for US Defined Benefit (DB) pension plans is not merely a figure in actuarial reports; it is a critical variable driving hundreds of billions of dollars in asset allocation, influencing the supply and demand of fixed-income markets, and shaping the bond yield curve. Recently, Wellington Investment Insights released its annual update on ROA assumptions for US corporate and public DB plans, revealing signs of stabilization or even structural reversal in this key metric at year-end 2025 after a 15-year period of continuous downward adjustment.

Data shows that as of year-end 2025, the average ROA assumption reported by Russell 3000 index constituent companies for their DB plans was 5.9%, unchanged from 2024; the average ROA assumption for public pension plans also remained unchanged at 6.9%. This stabilization is no coincidence but a correction to the dual shocks of the macroeconomic interest rate environment and balance sheet rules over the past two decades. In 2006, the Financial Accounting Standards Board (FASB) introduced mark-to-market balance sheet accounting for pensions, and Congress passed the Pension Protection Act, forcing many plan sponsors to reevaluate their investment strategies. From 2006 to 2022, US corporate plan ROA assumptions underwent a 15-year structural decline, with a cumulative drop of 200 basis points.

The core mechanisms behind this long-term decline are twofold: passive tilting of asset allocation toward fixed income and the sustained decline in risk-free rates. In 2006, only about 10% of companies selected an ROA assumption below 7.0%; by 2022, this proportion surged to 90%. Over these 16 years, corporate pensions increased their allocation to fixed-income assets by more than 20 percentage points to hedge liability risks. However, as interest rates continued to fall between 2008 and 2021, the expected yields of fixed-income assets were significantly compressed, leading to a decline in the expected returns of the overall portfolio.

However, this trend reversed after 2022. The broad-based stock and bond market sell-off in 2022 forced companies to reset their forward-looking return expectations for both equities and bonds. More importantly, the rise in the global interest rate center directly boosted the coupon yields of fixed-income assets, providing underlying support for the stabilization of ROA assumptions. Wellington’s investment strategy team notes that as yields recover, corporate plans with higher fixed-income allocations may modestly increase their ROA assumptions in the coming year; for plans with higher equity allocations, their ROA assumptions will remain relatively stable.

Regarding long-term capital market expectations, Wellington continues to predict that long-term equity and bond returns will struggle to match the levels seen over the past 50+ years. This concern about the downward shift in the long-term return center explains why plan sponsors remain cautious despite the rebound in interest rates. Although rising yields create room for adjusting ROA assumptions upward, the process of derisking into fixed-income assets has not stopped, though the pace has become more measured. Some plans are approaching their end-state, while some overfunded plans are reevaluating their surplus asset investment objectives.

ソース:Wellington

リサーチ

Former DeepMind Researchers' Startup Mirendil Raises $1B Series A to Automate AI Model Training

トピック · Mirendil融资动态重大
2026-11-14 05:22 GMT+8

Mirendil, an AI infrastructure startup co-founded by former Google DeepMind and Anthropic researchers Behnam Neyshabur and Harsh Mehta, announced it has raised $1 billion in Series A funding. According to SiliconANGLE, the round was jointly led by Kleiner Perkins and Clearlake Capital, with participation from more than six other investors including Google LLC. Neyshabur is a co-inventor of SAM, an algorithm that improves AI model performance on tasks they were not specifically trained for, while Mehta previously worked on pre-training projects at DeepMind and Anthropic.

The company's flagship product, Lor, is a platform designed to automate the trial-and-error phases of model training. Mirendil states that Lor uses hundreds of AI agents to handle code writing, infrastructure setup, algorithm selection, and data collection. The company cited an example where a single researcher used Lor to build a competitive coding model in 3.5 months for $10 million, with over 99% of the code generated by AI agents.

Building on this technology, Mirendil launched a specialized coding model called Fern last month. The company claims Fern can process 10 million input tokens per prompt, approximately ten times the capacity of Claude Opus 5.5, equivalent to about half a million lines of code. Fern is based on a version of Mirendil-Base with 100 billion to 500 billion parameters. The company also previewed upgraded versions ranging from 500 billion to 10 trillion parameters, planning to use these algorithms to develop more capable LLMs. Specific post-money valuation terms remain undisclosed.

ソース:SiliconANGLE

リサーチ

Trump Threatens to Double Tariffs on Mexico Over Migrant Caravan

トピック · 特朗普对墨关税威胁重大
2026-11-13 19:49 GMT+8

According to Semafor, US President Donald Trump has threatened to double tariffs on Mexican goods if Mexico does not stop a migrant caravan heading toward the US border. This statement escalates tensions with the United States' largest trading partner.

Illegal crossings at the US-Mexico border have plummeted during Trump's second term, partly because Mexico, which sends the vast majority of its exports to its northern neighbor, has deployed troops to stem the flow and win over the White House. However, recent electoral victories by the Democratic Party have lifted hopes among would-be migrants for a warmer welcome in the US. These hopes have been dashed by Trump's elevation of Stephen Miller, the architect of Washington's draconian immigration approach, to become the president's chief of staff.

For investors, Mexico is a critical source of US imports. If the tariff threat is implemented, it would directly increase costs for US businesses relying on Mexican manufacturing and agriculture, potentially passing these costs through to consumers and reigniting concerns about inflationary pressure. The measure remains a threat at this stage, with no formal executive order or legislative action yet taken.

ソース:semafor.com

リサーチ

Cloudflare Open-Sources Kitesurf, a Stateless Browser Engine for AI Agents

トピック · Kitesurf重大
2026-11-14 04:20 GMT+8

US cloud computing company Cloudflare announced on November 13, 2026, that it has open-sourced Kitesurf, its stateless browser engine designed specifically for AI agents, on GitHub under the Apache 2.0 license. This fulfills the company's promise since the tool's initial release in August, allowing developers to inspect the code, run it locally, modify it, and deploy their own instances to the Cloudflare Workers platform.

Kitesurf aims to provide AI agents with a lightweight, efficient browser environment that runs entirely on Cloudflare Workers without traditional browser containers. It supports HTML parsing, CSS application, page JavaScript execution, and screenshot rendering, and is compatible with the Chrome DevTools Protocol (CDP), enabling connections with clients like Puppeteer. Compared to full browsers like Chromium, Kitesurf uses less CPU and memory and scales on demand across Cloudflare's global network.

The engine is built on Cloudflare's Developer Platform and a series of open-source Rust packages compiled to WebAssembly. Core components include Blitz for HTML parsing and layout, Stylo for CSS processing, Boa for running JavaScript inside WebAssembly, and Linebender libraries and Krilla for text layout and graphics generation. Kitesurf consists of three main parts: the Engine managing sessions, PageScript running page scripts, and PageRenderer drawing results.

Developers can clone the repository to start the engine locally for testing or deploy it to their own Cloudflare account to replace the Browser Run managed APIs. Note that deployment requires a paid Workers plan because Kitesurf uses Dynamic Workers to create isolated page workers. Although the project remains under heavy development with a focus on improving compatibility, reliability, and efficiency, its open-sourcing marks a significant advancement in server-side infrastructure for AI agents.

ソース:Cloudflare Blog

リサーチ

AI Infrastructure Startup Drogue Raises $20 Million to Power Floating Data Centers with Ocean Heat

トピック · Drogue浮动数据中心クイックテイク
2026-11-14 04:07 GMT+8

Washington, D.C.-based startup Drogue has announced it completed a $20 million funding round backed by investors including XYZ Venture Capital, Toyota Ventures and Tamarack Global. Founded six months ago, the company plans to use ocean thermal energy conversion to power floating data centers, with a first multi-megawatt prototype expected at sea in early 2028.

ソース:TechCrunch

リサーチ

Bloomberg: SpaceX and Tesla in Intensive Merger Talks, Deal Possible by Dec 2026

構造的
2026-11-13 13:50 GMT+8

According to a Bloomberg report, Elon Musk's two major companies, SpaceX and Tesla, are engaged in "intensive" negotiations regarding a potential merger. Sources indicate that board members and advisors from both firms are actively working on defining the conditions for the combination and the leadership structure of the merged entity. The process has accelerated significantly compared to earlier expectations: vague considerations in October suggested a deal might be announced in the first half of 2027, but it could now be revealed as early as December 2026 or early 2027.

The report describes this as a "blockbuster" for financial markets. The resulting conglomerate would encompass electric vehicles, energy storage, rockets, and satellites. The stated goal is to allow artificial intelligence (AI) to permeate all areas of the business. Currently, SpaceX has a market value of approximately $1.9 trillion, while Tesla is valued at around $1.6 trillion.

However, the merger faces significant hurdles, with national security cited as the biggest obstacle. SpaceX conducts extensive business with the US government, whereas Tesla operates large production capacities in China. These two aspects must be cleanly separated. Additionally, the changing political landscape in the US is a factor, with Democrats having announced they will scrutinize Musk's business dealings.

ソース:heise online

リサーチ

AI enthusiast Muhammad Zulali uses Claude and Codex to hack 1994 Jeep Cherokee engine control unit, passing US California smog tests

トピック · AI逆向工程汽车ECUクイックテイク
2026-11-13 18:30 GMT+8

According to Tom's Hardware, car owner Muhammad Zulali used Anthropic's Claude and OpenAI's Codex to fix the emissions issues of a 1994 Jeep Cherokee, allowing it to pass California smog tests. The project utilized hardware worth no more than $15, including an Arduino Nano microcontroller.

Zulali faced the challenge of communicating between a modern laptop and the vehicle's primitive Engine Control Unit (ECU). Claude guided him to identify that the car used a Chrysler SCI bus at approximately 7,800 baud, with electrical signals inverted compared to a typical serial port. By using the Arduino Nano to handle signal inversion and connect to the Cherokee's 6-pin diagnostic plug, the system was able to read data such as coolant temperature and battery voltage.

After obtaining initial data, Claude assisted in collecting engine data and fault codes during test drives at 25 mph and 50 mph to diagnose the cause of excessive NOx emissions. Since the ECU firmware was not reflashable, Zulali instead used Codex to write code that injected custom logic into a nearby EEPROM, adjusting fuel pressure and other parameters to meet emission standards.

"Claude did most of the diagnostic work and Codex did pretty much all of the memory work," admitted Zulali. "I just did the wiring and the driving." The project has been open-sourced on GitHub as xj-scanner under the GPLv3 license. This case demonstrates the practical application capabilities of large language models in reverse-engineering legacy hardware protocols and guiding embedded system debugging.

ソース:Tom's Hardware

リサーチ
次へ →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

For interactive scenarios, even a 2-second session startup is considered too slow. The article suggests initiating warm-up requests when the user opens a chat window or application, utilizing the time window before the user sends their first message to have the session ready. Running sessions respond in approximately 200 milliseconds, while new session startups take about 2 seconds—pre-warming hides the latter within user operation latency.

It should be noted that the aforementioned performance figures and mechanism descriptions come from Amazon Web Services' own explanations of its platform, representing vendor-reported data without independent third-party replication.

For market participants, understanding the logic behind the evolution of ROA assumptions is key to predicting the flow of US pension funds. The ROA assumption not only determines pension expenses on corporate income statements (as a credit item for expected asset returns) but also profoundly influences pension fund trading behavior in public markets. When ROA assumptions stabilize and interest rates remain high, pension fund demand in the fixed-income market may shift from purely 'hedging liabilities' to 'seeking yields,' which will have a substantive impact on the supply and demand dynamics of long-term bonds. Wellington advises that plan sponsors should develop investment policies based on their own risk-return objectives and investment horizons, rather than forcibly binding specific long-term ROA assumptions to manage pension expenses or reported liabilities.