독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-29198 게시물

새로고침

PyTorch Official Channel Announces Live Q&A for Version 2.15: Meta Team to Explain NCCL Defaultization and NVFP4/FP8 Optimizations

주제 · NCCL对称内存与GIN중대
2026-10-28 23:12 GMT+8

The official PyTorch channel released a live Q&A video preview for PyTorch version 2.15 on October 28, 2026. According to the publisher's information, PyTorch 2.15 comprises 3,117 commits from 583 contributors since version 2.14. Core changes in this update include making the rewritten NCCL backend the default option, optimizing 4-bit (NVFP4) LLM decoding performance on NVIDIA Blackwell GPUs via NVGEMM, and adding FP8 support and faster decode kernels on Apple Silicon.

In the upcoming Q&A session scheduled for 11 AM PT on November 5, 2026, four core contributors from Meta—Tristan Rice, Natalia Gimelshein, Huy Do, and Jane Xu—will provide an overview of the 2.15 release and answer community questions about PyTorch and its new capabilities. Chris Gottbrath from Gottbrath Tech will moderate the session. This arrangement indicates that the resource is not just a routine introduction but a channel for first-line maintainers to address technical queries directly.

For readers involved in AI infrastructure development, this update involves changes to several critical underlying mechanisms. For instance, the defaultization of the NCCL backend introduces a flight recorder that dumps data on hangs, with TORCH_DIST_USE_NCCL2=0 or nccl-legacy provided as rollback paths; torch.distributed adds support for one-sided transports, including a NIXL backend and CUDA-graph-capturable transfers; and FSDP2 and pipelining implement finer memory and communication control. Additionally, FlyDSL kernels on AMD gfx950, simplified SYCL kernel authoring on Intel GPUs, and torch.compile support on Python 3.15 expand platform compatibility.

Given that only the episode description and topic list have been obtained, without a complete dialogue transcript, this recommendation aims to guide readers to this authoritative technical Q&A resource. Engineers can watch the video to hear the Meta team explain specific changes such as NCCL defaultization, NVFP4 decoding optimization, and Apple Silicon FP8 support, as well as how they handle potential rollback and compatibility issues.

출처:youtube.com

리서치

US Federal Grand Jury Indicts OpenAI Engineer Abdulrahman Alfozan for Bugging Conference Room

주제 · OpenAI工程师窃听案중대
2026-10-28 14:11 GMT+8

According to reporting by Sarah Brager for the San Francisco Chronicle, syndicated by Yahoo News, a US federal grand jury on Tuesday (October 28) formally indicted 35-year-old Abdulrahman "Abdul" Alfozan, a former OpenAI engineer. Alfozan, a Saudi national who joined the company in 2025 to work on ChatGPT-related projects, faces one felony count of interception of oral communications.

Court filings state that Alfozan installed an encrypted listening device in a conference room at OpenAI's San Francisco headquarters, a space frequently used for discussions on confidential research, business plans, and government partnerships. The device was hidden behind a television and wrapped in toilet paper. FBI agents arrested Alfozan in Millbrae on October 9, just over a week after security personnel discovered the device during a sweep of the room.

BBC technology reporter Laura Cress noted that according to the official complaint, Alfozan had used an AI assistant to query meeting room schedules and ask how to build and remotely access recording devices. OpenAI stated on X that it used its own AI models to analyze the device's origin and confirmed there are currently no indications that the individual improperly accessed or exfiltrated model weights, source code, or customer data.

If convicted, Alfozan faces a maximum penalty of five years in prison, a $250,000 fine, and three years of supervised release. He remained in custody without bail following his initial appearance. His attorney, William Pernik, described the allegations as "wholly out of character" and asked the public to withhold judgment until all facts are known.

출처:Yahoo News

리서치

Apple Adds 512GB RAM Option for Mac Studio M5 Ultra, Starting at $22,299

간단한 요점
2026-10-29 01:26 GMT+8

Apple officially launched a 512GB unified memory configuration for the Mac Studio equipped with the M5 Ultra chip on October 28, 2026. Previously capped at 256GB, this new option requires the base model with a 36-core CPU and 80-core GPU (starting at $6,799) plus an additional $15,500 upgrade fee, resulting in a minimum total price of $22,299. With maximum storage of 16TB, the price rises to $29,799.

This update addresses the demand for high-capacity local AI inference. Unlike the M5 Max chip, which remains limited to 128GB, the M5 Ultra's 512GB configuration allows users to run larger open-source models or complex agentic workflows on a single desktop without relying on multi-machine clusters. While the premium pricing targets professional developers and enterprises, the release aligns with Apple's prior commitment to offer this option by late October.

출처:MacRumors

리서치

Instagram Head Adam Mosseri Predicts Public Negativity Toward AI Will Continue to Worsen

주제 · Adam Mosseri간단한 요점
2026-10-29 07:05 GMT+8

Adam Mosseri, the head of Instagram, publicly stated on the Colin and Samir podcast that he believes public negativity toward artificial intelligence is not only worsening but will continue to do so. "It's not only getting worse, it's going to continue to get worse," Mosseri said. He predicted that the next couple of years will see more frustration and anger—some founded, some unfounded—which will make things difficult for many people, not just those in the industry.

Addressing the trend of college students booing tech CEOs at graduation ceremonies, Mosseri acknowledged that many people genuinely dislike AI, noting that this sentiment extends beyond young people and creatives. He attributed this widespread anxiety to a core psychological driver: "Fundamentally what I think it comes down to when you really have, like, a heart-to-heart with someone is a fear of being left behind in some way."

As the leader of one of the world's largest image and short-video platforms, Mosseri's remarks reflect how major AI application companies internally assess public trust crises. Previously, Instagram began requiring AI accounts to label themselves and cracked down on fake AI profiles, though Mosseri admitted it is impractical to ban all AI content from the platform. He also questioned current ranking mechanisms, stating that high-quality AI content is likely not ranked as well as non-AI content, and noted that the term "AI slop" is used too liberally.

출처:The Verge

리서치

WIRED AI Reporter Puts AI Agents in Charge of a Premier League Fantasy Team: Division of Labor Improves Predictions, but AI Still Falls Short of Human Superforecasters

주제 · AI智能体足球预测중대
2026-10-29 02:30 GMT+8

Benching Erling Haaland—that was OpenAI’s Codex’s first recommendation for WIRED Senior AI Reporter Will Knight’s Premier League fantasy team. Knight unleashed a horde of AI agents into the Premier League Fantasy Football game, letting them draft their own teams, set lineups, and adjust rosters based on weekly news, aiming to answer a specific question: Can a group of robots predict better than humans?

This wasn’t Knight’s first attempt. A few years ago, he used a tool from the UK’s Turing Institute that modeled purely on historical data, completely unaware of player injuries or coaching changes, yielding recommendations that were astonishingly absurd. This time, he changed his approach: he fed Codex several relevant research papers first, then asked it to build a harness—a system of instructions and code that guides different agents through various steps—including checking recent news and running predictive models.

Codex quickly assembled an analysis framework that looked information-rich, but the initial recommendations remained foolish. Knight then instructed Codex to create a team of specialized agents: some responsible for reading news and blog posts, others for analyzing recent match results, more for running models, and a batch dedicated to critiquing suggestions. This change significantly improved the output. He subsequently increased the number of agents to over one hundred, hoping the system would become smarter, but aside from inflating his API bill, there was little change.

Haifeng Xu, a computer scientist at the University of Chicago who studies the application of AI in market scenarios, believes that new models perform better in predictions fundamentally because they can digest and analyze massive amounts of real-world information and orchestrate more complex systems to generate guesses. Xu told Knight that AI is simply getting smarter at understanding the world, which is the root cause of their progress. Xu’s team began testing commercial models in prediction markets such as Kalshi and Polymarket, having models guess outcomes like the NBA Championship winner or whom President Trump would meet with. They quickly found that the models performed well, sometimes even surpassing average people without special guidance. In their tests, GPT-5 had the strongest predictive ability, followed by Grok 4, Claude Sonnet, Gemini, and Llama.

Xu’s team subsequently developed Prophet Arena, a benchmark for measuring model predictive capability. Another benchmark, Prediction Arena, developed by researchers at Harvard University and Arcada Labs, gave frontier models $10,000 to wager on prediction markets over 57 days. The models consistently lost money in the experiment, but researchers observed steady improvement over time. Xu told Knight that AI makes predictions most effectively using custom harnesses, and that having agents perform distinct tasks—reading news, collecting data, and analyzing the big picture—seems to be the most effective path.

However, AI still frequently lags behind highly skilled human experts known as superforecasters. Xu explained that although humans cannot consume vast amounts of information like AI, their understanding of the real world—the dynamics behind news stories—is deeper. Yet this advantage is shrinking. A competition called the Metaculus Cup recently showed that AI can sometimes defeat the best predictors. Another benchmark, ForecastBench, is expected to release results soon, potentially providing further evidence that AI can catch up to human levels. Knight believes that major labs will treat predicting the future as a new way to measure progress in machine intelligence.

출처:WIRED

리서치

Google Expands AI Search Ads Globally as Core Ad Growth Slows

주제 · 谷歌AI搜索广告扩张중대
2026-10-29 05:48 GMT+8

Philipp Schindler, Chief Business Officer at US tech giant Alphabet (Google's parent company), announced during Wednesday's third-quarter earnings call that the company is expanding high-quality ads within its "AI Mode" to international markets for the first time. The move aims to leverage generative AI to open new revenue streams as growth in its core search advertising business continues to slow.

Alphabet reported $127.6 billion in total revenue for the third quarter, up 25% year-over-year. However, Google Search revenue grew only 12% year-over-year, down from 17% in the second quarter and 19% in the first. Following these results, Alphabet shares fell 8%. "We are now scaling high-quality ads within AI overviews in AI mode to more users and queries," Schindler said. "This includes taking our monetization footprint global for the first time. Today, we’re announcing the expansion of ads in AI mode into international markets alongside multi-turn ads and AI mode in the U.S."

Google claims early results suggest the formats are working. According to Schindler, click-through rates for sponsored product listings in AI Mode are already comparable to those for organic listings, and user research indicates people find the ads relevant. This builds on Google’s May announcement of new Gemini-powered ad formats, including conversational discovery ads and sponsored recommendations integrated into AI-generated answers.

Despite the slowdown in search ad growth, CEO Sundar Pichai noted that Search queries grew at their fastest year-over-year rate in five years, crediting AI features with changing how people use Search. Pichai stated that queries in AI Mode are three times longer on average than traditional searches, and one in five AI Mode searches globally involves voice, images, or documents. Additionally, Google Cloud revenue surged 103% to $30.7 billion, partially offsetting the softness in the advertising segment.

출처:TechCrunch

리서치

HPE Integrates SASE Solution EdgeConnect into Mist and Aruba Central for Unified AI Network Operations

중대
2026-10-27 21:00 GMT+8

US technology company Hewlett Packard Enterprise (HPE) announced on October 27, 2026, that it will integrate its Secure Access Service Edge (SASE) solution, HPE Networking EdgeConnect, with its two major network management platforms, HPE Mist and HPE Aruba Central. This move aims to break down data silos between wired, wireless, and security tools, optimizing user experiences from edge to cloud through a unified agentic AI framework.

As enterprise IT environments become increasingly distributed, network and security teams often lack a unified view across campus, branch, WAN, and cloud domains, making troubleshooting difficult and increasing security risks. HPE stated that this integration combines its premier SASE, wired access, and wireless LAN solutions into a single operational experience, providing greater visibility, automated protection, and faster issue resolution.

Technically, HPE is expanding the scope of its AI engine, HPE Marvis, to cover the SASE domain. Tasks will be automated via an "agentic mesh," while "user experience twinning" technology will identify and fix problems without users being present. Additionally, a Large Experience Model (LEM) will be used to predict and prevent performance issues in key applications like Zoom and Teams.

For existing customers, those using HPE Mist or Aruba Central will gain access to advanced EdgeConnect capabilities, including application-aware routing, WAN optimization, and integrated security. Customers using HPE's full stack of wired, wireless, and SASE products will achieve end-to-end visibility and control. HPE plans to release the EdgeConnect-Mist integration during calendar year 2026, with the Aruba Central integration following in the first half of 2027.

David Hughes, SVP & GM, SASE and Security for Networking at HPE, noted that enterprise networks have become too complex to manage through isolated tools and domains. By bringing EdgeConnect together with Mist and Aruba Central, HPE intends to combine its best AI, wired, wireless, and SASE innovations, giving customers a clear path to self-driving network operations.

출처:Yahoo Finance

리서치

eBay Reports Q3 Results: Revenue Up 14%, GMV Up 15%

중대
2026-10-29 04:05 GMT+8

US e-commerce platform eBay released its financial results for the third quarter ended September 30, 2026, on October 28. The report shows revenue of $3.2 billion, up 14% year-over-year on an as-reported basis and up 13% on an FX-neutral basis. Gross Merchandise Volume (GMV) reached $23.2 billion, up 15% year-over-year on an as-reported basis and up 16% on an FX-neutral basis.

In terms of profitability, eBay reported GAAP diluted earnings per share of $1.11 and non-GAAP diluted EPS of $1.51. GAAP and non-GAAP operating margins were 19.2% and 26.3%, respectively. The company returned $827 million to stockholders during the quarter, including $689 million in share repurchases and $138 million in cash dividends.

CEO Jamie Iannone stated that the strong growth reflects increased consumer adoption of "recommerce" (secondhand goods trading), driven by expansion in focus categories, the C2C marketplace, and live shopping.

출처:prnewswire.com

리서치

HBS Professor Randy Cohen Discusses Replicating Private Equity in Public Markets on Alpha Exchange Podcast

중대
2026-10-29 01:00 GMT+8

The Alpha Exchange podcast released a conversation with Randy Cohen on October 28, 2026. Cohen is a Professor of Finance and Entrepreneurship at Harvard Business School and co-founder of PEO Partners, a firm focused on bringing private equity strategies to public markets. His academic background covers asset pricing, investment management, and the challenge of identifying genuine investment skill, providing professional support for the views presented in the episode.

According to the show notes, Cohen describes an approach to private equity replication that begins with the economics of a leveraged buyout (LBO). He notes that companies attractive to buyout investors generally exhibit strong profitability, attractive valuations, healthy cash generation, and relatively manageable business risk. These characteristics overlap meaningfully with established equity factors. By adding industry selection and a measured amount of leverage, Cohen argues that a liquid portfolio can capture important elements of traditional private equity exposure.

The episode also explores the differences and challenges facing this replication strategy. Private equity managers can make operational improvements and exercise control, while public portfolios offer liquidity but expose investors to daily market volatility. Additionally, the smoothing of private valuations—what Cliff Asness has called 'volatility laundering'—complicates the comparison. The discussion covers these trade-offs, the substantial role of fees, and the potential for liquid private equity to serve as a complement rather than a replacement for traditional allocations.

출처:alphaexchange.simplecast.com

리서치

Alphabet Q3 Earnings: YouTube Ad Revenue Misses Expectations, Shares Drop Over 10% After-Hours

주제 · Alphabet중대
2026-10-29 04:49 GMT+8

US tech giant Alphabet (Google's parent company) disclosed in its Q3 2026 earnings report that advertising revenue for its video platform YouTube reached $11.2 billion, a 9% increase year-over-year. This figure fell short of the $11.43 billion analyst consensus estimate, and the growth rate slowed from 13% in the previous quarter. Following this news, Alphabet's shares dropped more than 10% in after-hours trading.

Despite the underperformance in advertising, YouTube showed momentum in other areas. The company stated that subscribers for YouTube Music and Premium now exceed 150 million, up significantly from 125 million in March 2025, noting that the subscription business is growing faster than ads. Additionally, according to Nielsen data, YouTube's share of US television viewership was about twice that of Netflix. Google Chief Business Officer Philipp Schindler noted that over a third of YouTube's Q3 ad revenue came from living room devices.

This earnings release comes amid intensifying competition in the streaming industry. Rival Netflix lowered its full-year ad revenue forecast last week from $3 billion to $2.7 billion due to delays in rolling out its ad products. Meanwhile, Instagram announced a major overhaul of its long-form video features this week, designed to directly challenge YouTube and bolster its presence on TV screens. Ross Benes, a senior analyst at EMARKETER, commented that everyone in streaming is chasing YouTube's success, with even Instagram pivoting its strategy based on YouTube's performance.

출처:Business Insider

리서치

Microsoft Q1 Earnings: Azure Grows 48%, Capex Below Guidance, Free Cash Flow Remains Positive

주제 · 微软高管换血중대
2026-10-29 04:42 GMT+8

US tech giant Microsoft released its financial results for the first quarter of fiscal year 2027 (ended September 2026) on October 28, 2026. The report showed revenue reaching $92.3 billion, up 19% year-over-year, and net profit jumping 32% to $36.6 billion, both beating Wall Street expectations.

Azure cloud computing revenue, a key metric for investors, grew 48% to $33.2 billion, significantly outpacing the approximately 45% growth analysts had previously expected. This figure is viewed as the clearest sign that Microsoft's massive artificial intelligence infrastructure investments are translating into actual sales. Although the company spent $45.2 billion on capital expenditures, including data centers, during the quarter—an increase of about 30% from $34.9 billion a year earlier—this amount fell below the "more than $50 billion" guidance provided to investors in July.

Despite the heavy spending, Microsoft maintained healthy cash flow. Free cash flow for the quarter was $18.6 billion, down 28% year-over-year but still positive. This contrasts with recent reports from Meta, which saw its stock fall after raising spending forecasts and reporting negative free cash flow. Microsoft CFO Amy Hood had previously stated that the company expected to remain free cash flow positive throughout fiscal year 2027.

Additionally, Microsoft reported results under new business groupings for the first time, combining Azure with business software like Excel, Teams, and Copilot into an "Agent and Infra" segment, which saw revenue rise 24% to $76.8 billion. Following the news, Microsoft shares rose about 4% in after-hours trading. However, some "neocloud" companies dependent on Microsoft's compute demand, such as CoreWeave and Nebius, saw their stocks drop due to the lower-than-expected capital expenditure figures.

출처:DNYUZ

리서치

Vercel introduces Virtual Models to decouple AI model selection from application code

주제 · Vercel虚拟模型功能간단한 요점
검증됨 2026-10-29 06:35 GMT+8

Cloud development platform Vercel announced on October 28, 2026, the introduction of "Virtual Models" within its AI Gateway service. This feature aims to decouple the model string used in application code from the actual underlying large language model and its configuration. Developers can create a virtual model ID (such as vmc/coding-agent) and centrally configure the corresponding real-world model, priority, fallback strategies, and settings like Zero Data Retention in the AI Gateway dashboard.

Vercel noted that 125 new models from more than 10 labs became available on AI Gateway in the last 90 days, averaging about 10 per week. Previously, switching models typically required changing code and redeploying the application. With Virtual Models, teams only need to edit the configuration at the gateway level, and all applications calling that virtual ID will automatically pick up the new model and settings within minutes, without any code changes or deployments.

Additionally, Virtual Models enable task-specific optimization. For example, high-frequency, cost-sensitive support ticket triage tasks can be assigned to lightweight models, while low-frequency but reasoning-intensive legal contract review tasks can use high-performance models. Platform and security teams can control modification permissions via roles, ensuring governance policies (such as data privacy requirements) are uniformly enforced across all applications, even if requestors attempt to override these settings.

출처:Vercel

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

For now, Knight’s AI team has not surpassed the league average, and he will not use any of its suggestions for his own team selection. But he expects AI to undoubtedly improve, perhaps soon dominating fantasy leagues with ease. He wrote that he somewhat hopes it won’t happen too quickly, as that would ruin the fun of earning bragging rights by winning the competition on one's own merit.