독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-10115 게시물

새로고침

Meta Expands Instagram's AI Age-Estimation Tech Globally to Identify Teens via Visual Cues

중대
2026-11-09 23:32 GMT+8

US technology company Meta announced on November 9 that it is expanding its AI-powered age-estimation technology on Instagram to a global scale. The move aims to more accurately identify users suspected of being teenagers and place them under the strict protections of "Teen Accounts," even if those users provided an adult date of birth during registration.

According to TechCrunch, Meta's age-estimation tools use AI to analyze contextual clues across entire profiles, such as birthday celebration posts, mentions of school grades, and content across various formats including posts, comments, bios, and captions. Additionally, the company examines visual cues from photos and videos, including a person's height or bone structure. Once the system suspects a user is a teen, the account is automatically classified as a "Teen Account," triggering additional safeguards such as restricting direct messages to people already followed or connected, hiding harmful comments, and setting accounts to private by default.

This global rollout comes amid significant legal pressure on Meta. In August, Meta reached settlements worth up to $18 billion with nearly every US state and territory to resolve lawsuits alleging its platforms harm teens. As part of these agreements, Meta committed to improving how it identifies teens and children under 13. "That’s why as part of our agreement with 52 bipartisan attorneys general... we’re investing in even stronger technology to proactively catch accounts that may belong to children under 13," Meta wrote in a blog post.

Beyond Instagram, Meta stated it is continuing to roll out the technology on Facebook, with a similar global rollout planned for Threads later this year. For parents in the US, Instagram will display an in-feed video explaining Meta's approach to ensuring their teens are in age-appropriate protections. This shift marks a transition for social media platforms from relying on self-reported ages to employing complex AI behavioral analysis and biometric inference to meet regulatory demands.

출처:TechCrunch

리서치

Apple Shifts Tax Haven Status: Singapore Overtakes Ireland, Saving ~$2.7B in FY2026

주제 · 苹果税务重心转移중대
2026-11-09 01:29 GMT+8

According to a report by the Financial Times, Apple Inc. has shifted its primary tax haven status from Ireland to Singapore, as revealed in its latest annual filing with the U.S. Securities and Exchange Commission (SEC). This marks the first time Apple has provided such detailed country-level tax disclosures under new US accounting regulations.

The Financial Times estimates that Apple saved approximately $2.7 billion in taxes during fiscal year 2026 by booking profits in Singapore. Although Singapore's implementation of a new global minimum supplementary tax increased Apple's local tax burden by about $2.5 billion, the net savings compared to other jurisdictions remained substantial.

The filing shows Apple achieved a pre-tax income of $159.2 billion and paid $27.1 billion in income taxes for the fiscal year, resulting in an effective tax rate of 17.0%. Previously, Apple relied heavily on Ireland's low corporate tax rates, but international tax reforms have necessitated a restructuring of its global profit allocation.

For investors, this transparency offers a clearer view of how geographic tax strategies impact net income. While the FT estimates that up to $50 billion in profits flowed through Singapore entities, Apple has not independently confirmed this specific figure, only disclosing the tax effects required by the new standards.

출처:Apple Aktie: Singapur überholt Irland als Steuerstandort

리서치

RIAA CEO Mitch Glazier: Using Market Substitution to Prove AI Music Training Is Not Fair Use

주제 · RIAA诉Suno版权案중대
2026-11-09 23:30 GMT+8

In a November 2026 episode of The Verge Decoder podcast, RIAA CEO Mitch Glazier detailed the legal logic behind the Recording Industry Association of America's lawsuit against AI music generation company Suno. His central statement was: 'We are clearly using this precedent-setting litigation to prove at the legal level that training on music does not constitute fair use.' The mechanism supporting this conclusion is the fourth factor of the fair use doctrine, which courts consider the most critical: the effect on the potential market for the copyrighted work.

Glazier explained this mechanism by stating: 'You are essentially creating music that competes side-by-side with human music on the same platform; users can click on either one and pay the same subscription fee. So you are affecting the economic pool, and you are effectively substituting in the marketplace.' In other words, the substitution argument is not separate from the training issue but serves as the legal tool to prove infringement, because the final output of training (AI-generated music) directly replaces the economic position of the trained works in the market.

This legal strategy is embedded within the broader governance philosophy described by Glazier. He stated: 'My global focus is: have we secured the basic rights of creators? If we have, fans can take over and drive the market.' Legal action is not the end goal but a prerequisite to create the conditions for market mechanisms to function. Once rights are established, subsequent steps such as labeling informing consumer choice, platform competition driving standardization, and fan preferences determining economic outcomes can take effect.

Glazier believes that, unlike the Napster era, positions are converging in the current AI controversy. He said: 'The good news is, unlike Napster, I don’t feel like we are facing it at odds.' He characterized the current struggle as 'keeping art human and preserving the connection between artists and fans,' rather than merely a business model dispute. In his view, fans are the ultimate market force: 'If you know a song is AI, your view of it changes,' and this preference difference will drive economic results.

On the legislative front, Glazier introduced the design concept of the NO FAKES Act. This is a proposed United States federal bill (not yet enacted into law) aimed at establishing a new intellectual property framework for voice and likeness rights. He pointed out the current legal gaps: 'United States federal law does not actually recognize an individual’s right to their own voice or likeness; state-level publicity rights typically apply only to advertising and endorsement scenarios.'

The takedown mechanism designed in the bill draws lessons from the Digital Millennium Copyright Act (DMCA). Glazier described a three-layer balance structure: the notice provider faces thresholds (including penalties for false notices), the recipient has counter-notice rights, and for the middle ground involving constitutional issues such as parody or news, a procedure for a judge to quickly rule on a temporary restraining order was designed. 'Nobody wants platforms to decide constitutional questions,' he said.

출처:The Verge

리서치

Medium Launches TK App to Integrate AI Directly into Writing Process with Transparency Focus

중대
2026-11-10 00:12 GMT+8

Online publishing platform Medium launched a new app called TK on Monday, aiming to bring artificial intelligence deeper into the writing process. The core idea is to make it easier to turn ideas into written words by incorporating an AI assistant directly into the editor itself. Users can ask for help by typing “@TK” followed by their request or prompt.

In publishing and journalism, “TK” traditionally serves as a placeholder for missing information (To Come). Medium gives this term a new meaning, using the AI assistant to help users fill in the blanks. According to TechCrunch, TK can answer questions that would normally require switching tabs to look up, such as the number of days in a particular month, and can pull in bits of writing previously saved to the app during ideation. Additionally, TK ships with skills for editorial feedback, headline feedback, and matching drafts to the style guide for Medium publications.

This move comes at a time when the publishing and literary worlds are reckoning with the role AI has played in creating new long-form prose. Multiple authors have lost book deals over suspicions of using AI technology, while other literary awards have come under scrutiny after winners were suspected of writing with AI assistance. While writers have long used automated tools for proofreading and light grammatical edits, there is growing concern across the industry that an AI assistant’s editorial “suggestions” can become the writing itself, instead of reflecting the writer’s own thoughts.

Medium leans into this debate, taking the side that these nuances will ultimately no longer matter. The company celebrates what AI can bring to writing, stating in a blog post that “artificial intelligence makes writing more important, not less,” and proclaiming that “AI is a technology that has the power to amplify human intelligence and it uses the written word as a foundation.” To address concerns about low-quality content and blurring the line between human-written and AI-written text, TK allows authors to publish notes that transparently explain how AI was used in their writing. Medium explained that this is “a start of what we believe is required to publish in the AI era, which is to make the provenance of writing clear, transparent, and inspectable, right down to the prompting and source material.”

Beyond writing assistance, TK will help users remember and organize previous writings and suggest prompts. The app can optionally connect to your own AI agent through TK’s MCP server or Claude skill, so it can better understand your projects, positions, and other ideas stored in your agent’s memory. Meanwhile, the new app incorporates human feedback in a social layer, allowing writers to share early drafts and get feedback. At launch, TK is available on the web, with a desktop app on the roadmap. It is free for those with a Medium account.

출처:TechCrunch

리서치

AWS Launches $100 Million Health Impact Initiative to Provide AI and Cloud Technology to Global Health Organizations

주제 · AWS健康影响计划중대
2026-11-09 22:00 GMT+8

US cloud computing company Amazon Web Services (AWS) announced the launch of its Health Impact Initiative on November 9, 2026, committing $100 million over five years in AI, agentic services, and cloud infrastructure. The initiative aims to support organizations addressing critical health gaps affecting underserved populations worldwide. Nonprofit health organizations, government agencies, research institutions, and technology companies can apply for support to build and scale health solutions ranging from prototypes to production environments.

This new program builds on AWS's previous social impact investments. Since 2021, AWS has provided more than $60 million in AI and cloud technology to over 600 health organizations in more than 80 countries through initiatives like the AWS Health Equity Initiative. The Health Impact Initiative focuses on three key domains: expanding access to quality health services, advancing women's and children's health research, and improving food security and nutrition. Dr. Rowland Illing, Chief Medical Officer at AWS, stated that while generative AI is transforming healthcare, transformative technologies often fail to reach the communities that need them most due to barriers in accessing and deploying these tools.

Awardees will receive AWS credits sized to their estimated needs, starting with up to $10,000, valid for 12 months and extendable to 24 months. Applications require a description of the problem, the proposed solution, and an estimate of the number of lives impacted, with reviews conducted weekly. Previous AWS-supported projects include the Akshaya Patra Foundation in India, which optimized school meal distribution for millions of children, and Eve's Ark, which preserved over one petabyte of women's health data on AWS.

출처:Amazon News

리서치

"Deep Residual Learning for Image Recognition": Solving Deep Network Training Difficulties by Reconstructing Layers to Learn Residual Functions

검증됨 2026-11-10 00:12 GMT+8

장기 리딩 · 《Deep Residual Learning for Image Recognition》(2015)

In the field of computer vision, as the number of neural network layers increases, models often face the challenge of being difficult to train. The paper "Deep Residual Learning for Image Recognition" proposed by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun offers an innovative solution to this core pain point. The study explicitly states that "Deeper neural networks are more difficult to train," so the authors introduced a residual learning framework aimed at making it easier to train networks significantly deeper than those previously used.

The core of this method lies in redefining the role of network layers. While traditional networks attempt to directly learn unreferenced functions, residual networks explicitly reconstruct layers to learn residual functions with reference to the input. This mechanism allows the network to be optimized more easily and achieve accuracy gains from significantly increased depth. Empirical evidence shows that these residual networks are not only easy to optimize but also maintain or improve performance even when depth is substantially increased.

Evaluations on the ImageNet dataset show that the authors tested residual networks up to 152 layers deep, which is 8 times the depth of VGG networks but with lower complexity. An ensemble model composed of these residual networks achieved a 3.57% error rate on the ImageNet test set and won first place in the ILSVRC 2015 classification task. Additionally, due to the advantages brought by extremely deep representations, the authors obtained a 28% relative improvement on the COCO object detection dataset and won multiple competitions. For researchers and engineers seeking to understand the foundations of modern deep learning architectures, this paper, submitted in 2015 and formally published in 2016, provides key insights into how structural innovations can break through depth limitations. It is recommended to start with the abstract and methodology sections, focusing on the definition of residual functions and their impact on the optimization process.

출처:《Deep Residual Learning for Image Recognition》(2015)

리서치

The Superinvestors of Graham-and-Doddsville: The Intellectual Origins of Value Investing and the Margin of Safety

검증됨 2026-11-10 00:12 GMT+8

장기 리딩 · 《The Superinvestors of Graham-and-Doddsville》(1984)

In 1984, Warren Buffett delivered his renowned essay "The Superinvestors of Graham-and-Doddsville," aimed at challenging the efficient market hypothesis. He used a thought experiment involving a "national coin-flipping contest" to illustrate that if 225 million people participated in flipping coins, after 20 rounds, 215 individuals would inevitably remain who had guessed correctly every time. These individuals might boast of exceptional skill, but if one discovered that 40 of them came from the same zoo or the same "intellectual village," such success could not be explained by luck alone—it must instead be attributed to a shared intellectual origin. Buffett argued that these long-term outperformers of the market all originated from the value investing framework established by Benjamin Graham and David Dodd, known as "Graham-and-Doddsville."

The core principle shared by these investors is identifying the discrepancy between a company's intrinsic value and its market price. They disregard academic concepts such as beta coefficients, capital asset pricing models, or covariance of security returns, focusing instead on just two variables: price and value. For example, Walter Schloss never attended college, yet he adhered to Graham’s principles by reviewing handbooks and annual reports to find securities trading significantly below the valuation a private owner would assign. He did not worry about whether he bought on a Monday or Thursday, nor did he consider seasonal factors—only whether a business was worth one dollar while he could buy it for 40 cents. This exploitation of the gap between price and value forms the foundational mechanism of value investing.

Another key principle is the "margin of safety." Buffett cited the Washington Post Company, noting that in 1973 its market capitalization was only $80 million, yet its assets could be sold to any of ten potential buyers for no less than $400 million. If the stock price fell further to a $40 million valuation, traditional risk measures (such as beta) would suggest increased risk—but in reality, the lower the purchase price, the higher the potential return, and the lower the actual risk. This is akin to building a bridge designed to carry 30,000 pounds, but only allowing trucks weighing 10,000 pounds to cross. That substantial buffer is the margin of safety, ensuring that even if estimates are off, investors will not suffer permanent loss.

For today’s readers, overwhelmed by information overload and market noise, maintaining rationality remains a universal challenge. Suppose you are evaluating a new technology company whose market sentiment is extremely optimistic, driving its price-to-earnings ratio to 100 times, while your fundamental analysis suggests a fair P/E ratio of only 20 times. Although the stock may continue rising in the short term, the lack of a margin of safety implies massive downside risk if expectations fail. Conversely, if a mature company’s stock price drops to 50% of its liquidation value, and management is trustworthy, even if the market temporarily ignores its worth, this significant divergence between price and value provides a protective margin of safety similar to that described in the text. This scenario is hypothetical, but it reflects the core logic in the original article regarding the relationship between price and value.

출처:《The Superinvestors of Graham-and-Doddsville》(1984)

리서치

이 뷰에서 모든 콘텐츠를 확인했습니다

로그인 / 회원가입
독서리서치레이더투자 프레임워크

However, Glazier also acknowledged the boundaries of this strategy. He mentioned: 'I believe Warner partnered with Suno; they trained models on Warner’s music so they wouldn’t run into these issues.' Note that this statement carries uncertainty ('I believe'). If this licensing model becomes widespread and generated content no longer competes for the same subscription fees on the same platforms as human music, the conditions for the market substitution argument may no longer be met. His response in this scenario would shift to supply-side control: restricting the distribution scope of AI-generated content through licensing agreements to prevent it from diluting the economic share of human music on streaming platforms.

Regarding labeling, Glazier noted that California’s transparency law has required Suno to add machine-readable AI-generated watermarks since August 2026, and competition between Apple and Spotify will drive broader labeling practices. However, he admitted that for DIY creators uploading via distributors like DistroKid, the integrity of the technical labeling chain remains an unsolved problem: 'Your entry point is DistroKid, TuneCore, the distributors. Are they willing to use AI detection technology to determine if this is fully generated?'

Overall, Glazier presented a layered logic: legally, establishing through litigation that training does not constitute fair use (using market substitution as the argumentative mechanism); politically, establishing a new framework for voice and likeness rights through legislation; and commercially, guiding consumer choice through labeling and platform competition. The three layers are mutually conditional—markets cannot function without established rights, and consumers cannot make informed choices without labeling. The actual effectiveness of this approach depends on whether AI companies accept the legal outcomes, whether platforms cooperate with labeling, and whether licensing models alter the premises of the substitution argument.

All readers seeking a deeper understanding of the essence of value investing are recommended to read "The Superinvestors of Graham-and-Doddsville." Published in 1984 and authored by Warren Buffett, the essay not only clarifies that value investors do not rely on luck or mimicry of leaders, but are grounded in independent intellectual judgment, but also vividly explains why seeking discrepancies between price and value remains viable even in an era dominated by efficient market theory. It is advised to begin with the coin-flipping analogy at the essay’s start, gradually grasp the concept of intellectual origin, then explore practical applications through specific cases to understand the use of the margin of safety.