ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-2154 posts

Refresh

Mixup: Beyond Empirical Risk Minimization

Verified 2026-11-21 00:02 GMT+8

Long-term reading · 《mixup: Beyond Empirical Risk Minimization》(2017)

Although large deep neural networks are powerful, they often exhibit undesirable behaviors such as memorizing training data and sensitivity to adversarial samples.

Traditionally, these networks are trained by minimizing the average error on the training data, following the principle of Empirical Risk Minimization (ERM). However, when the number of model parameters is comparable to the number of training samples, ERM can lead models to "memorize" rather than "generalize" the data, resulting in unstable predictions on out-of-distribution samples.

To address this issue, the paper proposes the mixup method. This is a simple, data-agnostic data augmentation technique. Its core idea is to construct virtual training samples via convex combinations of two randomly selected training samples and their labels. Specifically, for input vectors $x_i$ and $x_j$, and corresponding one-hot encoded labels $y_i$ and $y_j$, new samples are generated as $\tilde{x} = \lambda x_i + (1-\lambda) x_j$ and $\tilde{y} = \lambda y_i + (1-\lambda) y_j$, where $\lambda$ is a mixing coefficient sampled from a Beta distribution. This approach implicitly encodes the prior knowledge that linear interpolation of feature vectors should result in linear interpolation of target values.

Experiments show that mixup significantly improves the generalization ability of state-of-the-art neural networks on datasets such as ImageNet-2012 and CIFAR-10. Additionally, it reduces the model's memorization of mislabeled data, increases robustness to adversarial samples, and stabilizes the training of Generative Adversarial Networks (GANs). The implementation of Mixup requires only a few lines of code with minimal computational overhead, making it an efficient and universal regularization technique. This 2017 paper, "Mixup: Beyond Empirical Risk Minimization," is suitable for researchers and engineers seeking to understand how simple data augmentation strategies can enhance the generalization and robustness of deep learning models. It is recommended to start reading from the introduction and methodology sections, focusing on how it replaces complex domain-specific data augmentations at extremely low computational cost.

Sources:《mixup: Beyond Empirical Risk Minimization》(2017)

Research

Berkshire Hathaway Owner's Manual: Think Like a Partner, Measure Value with Look-Through Earnings

Verified 2026-11-21 00:02 GMT+8

Long-term reading · 《Berkshire Hathaway Owner's Manual》(1996)

In the Berkshire Hathaway Owner's Manual released in 1996, Warren Buffett articulated the core philosophy of the company's operations to its shareholders. This document is not a detailed financial disclosure but is intended to help new shareholders understand management's operating principles, objectives, and limitations. For readers seeking to establish a long-term investment framework, its discussions on ownership concepts, value measurement standards, and accounting limitations offer enduring guidance.

First, the manual establishes an "owner-partner" mindset. Buffett noted that although Berkshire operates as a corporation, its attitude is partnership-based. He and Charlie Munger view shareholders as owner-partners and themselves as managing partners. They do not consider the corporation to be the ultimate owner of business assets but rather a vehicle through which shareholders own those assets. This concept requires investors not merely to hold pieces of paper whose prices fluctuate daily, but to imagine themselves as partial owners of businesses—much like owning a farm or an apartment building—with expectations of long-term holding. This shift in mindset forms the foundation for resisting short-term market noise.

Second, the company's long-term economic objective is to maximize the average annual growth rate of per-share intrinsic value, rather than simply pursuing scale. Buffett explicitly stated that they do not measure economic significance or performance by company size, but by per-share progress. He acknowledged that as the capital base expands, the pace of per-share progress will inevitably slow in the future, potentially falling well below the previous 25% compound annual growth rate; however, the goal remains to exceed the average of large U.S. corporations. This principle underscores the efficiency of compounding under limited capital, reminding investors to focus on the quality of return per unit of capital rather than the mere expansion of totals.

Third, to overcome the limitations of traditional accounting, Berkshire introduced the concept of "look-through earnings." Because Berkshire directly owns diversified businesses and also holds significant publicly traded common stocks through its insurance subsidiaries, reported consolidated earnings often fail to reflect true economic performance. Consequently, Buffett and Munger largely disregard these consolidated figures, instead reporting their share of undistributed earnings from investee companies (net of taxes that would have been paid if distributed as dividends). They believe that as long as investee companies effectively use retained earnings to create more value, these unreported earnings will ultimately be fully reflected in Berkshire's intrinsic value through capital appreciation. This method reveals the gap between book profits and economic reality, encouraging investors to focus on the actual reinvestment capacity of cash flows.

Suppose a modern investor faces sharp fluctuations in the stock price of a high-growth technology company, feels anxious, and considers selling. Applying the principles of the Owner's Manual, this investor should not fixate solely on daily quotes but should assess whether they truly understand the company's competitive moat and the efficiency of its use of retained earnings. Just as Buffett treated Coca-Cola or Gillette shares, if long-term prospects are favorable, short-term price changes are meaningless except as opportunities to buy at lower prices. The investor should ask: Do I view this investment as a partner would? Am I focused on the growth of per-share intrinsic value rather than the instantaneous fluctuations of market capitalization?

Sources:《Berkshire Hathaway Owner's Manual》(1996)

Research

2026-11-20117 posts

Google Gemini adds agentic computer control to Mac app

Material
2026-11-20

Google has introduced a new "@Computer use" command in the Gemini app for Mac, enabling the AI assistant to perform agentic tasks such as sorting through files, editing documents, running applications, and accessing the internet. This feature transforms Gemini from a passive sidebar tool into an active agent that can execute complex workflows on the user's local machine.

To mitigate risks associated with autonomous file modification or deletion, Google has implemented granular access controls and automatic temporary backups. Every file Gemini interacts with is backed up; these copies are deleted when a new task begins (if the user approves the outcome) or automatically removed after 24 hours. The agent operates in a background window, typing and clicking without taking over the user's mouse and keyboard, allowing continued device usage during tasks.

The feature is currently restricted to paid subscribers (Google AI Plus, Pro, or Ultra) using Apple Silicon Macs running macOS Sequoia 15.0 or later. It is available only to users aged 18 or older in the US and supports English exclusively. This launch follows similar agentic capabilities already offered by competitors like Meta, OpenAI, and Anthropic.

Sources:TechRadar

Research

Paul Christiano: We May Be Six Months From AI That Could Overthrow Governments, and the Right Move Is to Pause Algorithmic Progress

Material
2026-11-20 23:24 GMT+8

Paul Christiano's central claim in his November 2026 Dwarkesh Patel podcast appearance is that the window for safe AI development is closing faster than anyone expected, and that the correct response is to pause the algorithmic improvements that make AI systems more capable per unit of training effort. He estimates a 1-in-3 chance that superintelligence powerful enough to overthrow governments emerges within a year on the current trajectory, and a 10% risk of losing control over AI systems within the next twelve months if development continues at full speed.

The mechanism Christiano describes runs through automation of AI research itself. He points to reporting from Anthropic and METR suggesting the pace of AI progress has roughly doubled over the last six to twelve months, even accounting for diminishing returns and compute bottlenecks. If AI systems increasingly write the code and design the experiments that build the next generation, the doubling accelerates: eight months, then four, then two, approaching physical limits within about sixteen months. Christiano says this is the most obvious top priority right now — understanding whether we are three, six, or twelve months from systems that could overthrow governments.

His policy proposal follows from a specific asymmetry he identifies. Pausing now is more feasible because the most dangerous capabilities have not yet emerged, and because internal models are at least somewhat protectable in a way that deployed systems are not. He states it seems impossible to protect the model weights of a system deployed across major cloud providers against a state actor, whereas for an internal model there is at least some chance. Pausing later, after those capabilities exist and are deployed, is far more dangerous because the weights could be stolen, and once that happens the race becomes forced and the security posture of whoever holds the weights is worse than that of the original developer. He also notes that much of China's current AI progress depends on leaked American algorithmic advances, so if the US stops now, China cannot fully catch up for a long time. He describes the theft scenario as the most likely doomed path.

What Christiano proposes labs do right now is stop accumulating further algorithmic innovations in pre-training efficiency, reinforcement learning, and architecture design until three conditions are met: security posture adequate to defend those innovations if they must be withheld from deployment, alignment methods sufficient to control the systems already built, and governance processes mature enough to notice when things start going wrong. He explicitly notes he is not speaking as an OpenAI board member, though he confirms he is one.

The pessimism in his framing is structural rather than rhetorical. He identifies interpretability of superintelligent minds as the most promising available plan, but assigns it only a 5% to 10% chance of working, and expects it to function for some number of months before breaking down unpredictably. His optimistic scenario involves two sequential 50/50 saving throws: a 50% chance of finding something that works, and a 50% chance that if nothing works, we find a way to stop. That yields a 75% chance of things being broadly okay. Zvi Mowshowitz, who calls the situation grim, says he would very much take a 75% chance of things turning out okay, while noting the odds are against us.

Sources:thezvi.substack.com

Research

Clouded Judgement's Jamin Ball: AI Makes Exploits Instant, Forcing Rushed Patches That Break Systems

Material
2026-11-20 22:04 GMT+8

Jamin Ball, who writes the Clouded Judgement newsletter tracking cloud software companies, proposes a specific mechanism to explain the recent wave of IT outages: artificial intelligence is compressing the time between vulnerability discovery and exploit availability to near zero, forcing security teams into rushed patching cycles that break production systems.

Ball's argument rests on a historical baseline. Traditionally, when a vulnerability was discovered in software, it took significant time—sometimes an indefinite amount—for an attacker to develop a working exploit. That gap gave security teams breathing room to craft thoughtful patches and deploy them carefully across complex environments. The PHP security issue he references is a concrete example: the vulnerability was known for a long time, but researchers believed no practical exploit was possible.

What changes with AI, according to Ball, is the timeline. AI systems can identify vulnerabilities faster than human researchers, and can convert those vulnerabilities into workable exploits almost instantly. This eliminates the window that previously allowed careful patch management.

The forced consequence is a "move fast and break things" posture. Companies can no longer afford to be thoughtful about patch deployment because every second an unpatched vulnerability remains exposed, an exploit may already exist. But patching is not a simple swap. Ball notes that updating a library or component introduces cascading issues: downstream systems break, duplicate versions of old software persist, and new bugs emerge from the interaction between patched and unpatched components. For large organizations with complex IT estates, patch management is inherently difficult to execute without introducing instability.

Ball's conclusion is that what we are observing is not primarily a wave of cyberattacks but a wave of self-inflicted operational damage: every organization, large and small, is rushing to patch known vulnerabilities and, in the process, breaking their systems more frequently. He explicitly labels this as a hunch—"I could be wrong"—and notes that many of the current outages are being described as non-cyber related.

The prediction that follows is more concerning. Ball argues the current outages are merely the first phase of what he calls "security Armageddon." The rushed patches themselves will introduce new weak spots and vulnerabilities. Attackers, armed with AI tools that make exploit development trivial, will find and weaponize those new weaknesses. We are in the outage portion now; he expects the attack portion to follow.

The strongest objection to this framework is that it assumes a behavioral change—security teams shifting from careful to rushed patching—without direct evidence that this shift is actually occurring at scale. The outages could have mundane causes: configuration errors, hardware failures, software bugs unrelated to patching urgency. Ball himself acknowledges this ambiguity. The argument works as an explanatory hypothesis for why outage frequency has increased, but it does not yet establish that AI-driven exploit acceleration is the proximate cause rather than one contributing factor among many.

Sources:cloudedjudgement.substack.com

Research

UK Investigatory Powers Tribunal Rules Home Office Must Confirm Apple Encryption Backdoor Order

Material
2026-11-20

The UK Investigatory Powers Tribunal, an independent court overseeing surveillance orders under the Investigatory Powers Act (IPA), ruled on Thursday that the UK Home Office must confirm it served Apple with a Technical Capability Notice (TCN) requesting weakened encryption. The judge rejected the government's argument that disclosing the existence of the order would damage national security, although the ruling noted that not every detail of the order needs to be made public.

The TCN prompted Apple to disable its iCloud Advanced Data Protection (ADP) feature for users in the UK last February. ADP provides end-to-end encryption for stored data, meaning even Apple cannot access the files. While the existence of the notice was already known, the Home Office had repeatedly refused to officially confirm or deny it. This ruling forces the government to admit the action but does not compel it to reverse the request, meaning UK users will continue to lack access to ADP for the immediate future.

Liberty, a UK privacy group that initiated the legal challenge on behalf of two British citizens, called the decision a "victory for transparency." Akiko Hart, Director of Liberty, stated, "This case is about privacy and how we keep our personal data safe. In a digital age, end-to-end encryption gives us control over our personal information and the freedom to communicate securely." Jurgita Miseviciute, Head of Public Policy at Proton, welcomed the ruling as a win for transparency but criticized the IPA itself, saying, "There’s no such thing as a backdoor to encryption that only lets the good guys in."

Apple has previously stated that it "never built a backdoor or master key" to any of its products or services and "never will." The company confirmed that iMessage and FaceTime remain end-to-end encrypted globally, including in the UK. Privacy groups including Liberty and Privacy International announced they will continue to challenge the UK government's power to issue these orders in the first place. A final hearing on this broader constitutional question is expected in 2027.

Sources:TechRadar

Research

Nvidia Nears $500M Acquisition of Inception to Bolster Nemotron and Robotics

Material
2026-11-20

Nvidia is nearing a deal to acquire the AI startup Inception for at least $500 million, according to three people familiar with the talks told The Information. The proposed acquisition also includes performance bonuses for the more than 30 employees at the two-year-old San Francisco Bay Area startup.

Inception builds language models using the diffusion process known from image generation. Rather than generating responses token by token like conventional models such as ChatGPT and Claude, it refines a full response step by step. The startup most recently released Mercury 3.

The acquisition aims to bolster Nvidia's open-source Nemotron models and physical models for robots, helping it compete with OpenAI and Anthropic as both build their own inference chips. Palantir and US government agencies already use Nemotron as an alternative to Chinese models.

Talks began before Nvidia CEO Jensen Huang's resignation and after Microsoft expressed interest. Inception co-founders and professors Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov are expected to join Nvidia full time. Neither company initially commented.

Sources:The Decoder

Research

Pydantic Becomes CVE Numbering Authority to Accelerate Security Disclosures

Material
2026-11-19

On November 17, the CVE Program officially added Pydantic as a CVE Numbering Authority (CNA). This designation allows the organization to independently assign Common Vulnerabilities and Exposures (CVE) identifiers and publish official security records for vulnerabilities found in the open-source projects it maintains. The scope of this authority includes core libraries such as Pydantic, Pydantic AI, Pydantic Monty, and httpx2.

The move addresses a growing bottleneck in the open-source security ecosystem. As the volume of valid security reports has increased significantly this year, the traditional workflow required every new vulnerability ID to be issued by GitHub, which acts as the CNA for advisories published through its Security tab. GitHub reviews these requests for the entire open-source community, leading to delays between a fix being released and its corresponding CVE record becoming public. By assuming the role of CNA, Pydantic removes this dependency on an external queue.

For developers and security teams, the practical impact is speed and clarity. Previously, a fixed release might appear days before the associated CVE ID was assigned, creating a gap where scanners like pip-audit or Dependabot could not immediately correlate the patch with a standardized identifier. Now, Pydantic can publish the CVE record at the exact moment the fixed version is released. Each record will detail affected and fixed versions, the nature of the impact, available mitigations, and credit for reporters who wish to be acknowledged.

The reporting process for users remains unchanged. Vulnerabilities should still be disclosed privately via the affected project's GitHub Security tab or by emailing security@pydantic.dev if no private channel exists. Public issue trackers are explicitly discouraged for initial disclosure. The Red Hat Root and the broader CVE Program facilitated this partnership, acknowledging the need for distributed authority in high-volume maintenance environments.

Sources:pydantic.dev

Research

NYT: Former Nvidia CEO Jensen Huang's management style reportedly built on fear and humiliation

Material
2026-11-20 21:25 GMT+8

According to a report by The New York Times, former Nvidia CEO Jensen Huang ruled through praise and fear, and also by humiliating people. This account provides new internal context for the leadership crisis at Nvidia, which saw Huang ousted earlier this month amid scandals involving alleged ties with China and personal relationships.

The Times states that Huang told subordinates that Nvidia’s wealth meant it wasn’t possible to motivate them with money, so humiliation was the only lever left. He reportedly told one executive his mother should have aborted him, and told another employee he would rip his arms off and beat him to death with them. While thousands of employees have signed a letter asking for Huang’s return, these allegations suggest a divided workforce.

The reporting indicates these incidents were not isolated. In 2005, Michael Winston, a former HR chief, threatened to sue and called Huang’s conduct the most abusive he’d seen; Nvidia paid him more than $1 million. Huang’s spokesman calls the allegations fabrications. This detailed account of Huang's internal management practices offers a specific explanation for the board's decision and employee reactions that goes beyond the external financial and regulatory scandal coverage.

Sources:TechCrunch

Research

Infineon and Megmeet Collaborate on 800 VDC Power Architecture for AI Data Centers

Quick take
2026-11-20

German semiconductor manufacturer Infineon Technologies has announced a collaboration with Shenzhen Megmeet Electrical to develop 800 VDC power architectures for next-generation AI data centers. The core of the partnership involves integrating Infineon’s second-generation CoolSiC MOSFET technology into Megmeet’s three-phase AI power shelf. This integration aims to reduce switching and conduction losses, thereby improving energy efficiency, power density, and reliability in high-load server environments.

According to Megmeet, the resulting platform delivers system efficiency well above the peak requirements of the 80 PLUS Ruby certification standard. The devices also incorporate Infineon’s .XT interconnection technology, which is designed to lower thermal resistance and enhance heat dissipation. This supports the rapid dynamic load changes typical of AI processors, addressing the growing demand for scalable and reliable power systems as data center consumption rises.

Beyond the initial power shelf, the companies plan to expand their cooperation to server power supply units (PSUs), intermediate bus converters (IBCs), and solid-state transformers (SSTs). These components are critical building blocks for emerging 800 VDC distribution systems, which seek to support higher power densities across AI facilities. The agreement builds on an existing relationship where Megmeet previously adopted Infineon’s CoolMOS and OptiMOS products for single-phase platforms.

Sources:eeNews Europe

Research

Lawyer's $9,450 private AI server with 12 used Nvidia V100s runs DeepSeek V4.1 Flash at 184 tokens/s

Quick take
2026-11-20

Nick Mermiges, a lawyer who builds private AI infrastructure, has upgraded his server from eight to twelve used Nvidia V100 GPUs for a total cost of $9,450. The system now runs every expert in the DeepSeek V4.1 Flash model using 4-bit NVFP4 quantization, achieving up to 184 tokens per second on code generation tasks.

The upgrade addresses a memory constraint: the original eight cards provided 254 GiB, but the model's experts require 278.9 GiB when loaded in NVFP4 format. Previously, Mermiges had to compress most experts to 2-bit precision. With the addition of a third board of four V100s, all 15,360 experts can now reside on the cards in their intended 4-bit precision without further compression.

A key technical limitation is that the Nvidia V100 architecture, released in 2018, lacks native FP4 or FP8 tensor cores. Consequently, every NVFP4 value must be unpacked into half-precision floating-point numbers via software before computation. The performance gain comes from reduced memory bandwidth usage—reading fewer bytes—rather than faster arithmetic operations.

Mermiges reports that the first token latency for a 256K context window dropped to 58 seconds, nearly twice as fast as his previous setup. Prompt reading speeds exceeded 4,500 tokens per second. He claims this configuration outperforms published stacks of three Nvidia DGX Spark units in single-request decoding while costing roughly half as much.

The system uses a specific topology where layers are divided across three boards, each containing four cards connected by NVLink. Hand-offs between the first two boards occur within a shared PCIe switch, but communication with the third board crosses host PCIe bridges, adding latency. Power caps were raised to 175 W per card, though pushing them to 200 W caused system instability during heavy multi-chat loads.

Sources:Tom's Hardware

Research

German Publisher Associations Issue Formal Legal Warnings to Meta Over AI Training Data Access

Material
2026-11-20 21:27 GMT+8

The Federal Association of Digital Publishers and Newspaper Publishers (BDZV) and the Media Association of the Free Press (MVFP), representing German media houses, have issued formal legal warnings (Abmahnung) to US technology giant Meta. They are demanding that Meta sign a cease-and-desist declaration regarding its updated Commercial Terms for business users.

The dispute centers on Meta's planned changes to its service terms, which would grant the company the right to automatically scrape, index, cache, and analyze domains, websites, and URLs associated with commercial Meta accounts, such as those of publishers. Although Meta initially included the word 'train' in the list of permitted uses for these scraped contents, it removed this specific term following massive protests from the publishing industry. However, the revised terms still do not explicitly exclude the use of this data for generative AI services, leading publishers to fear that their journalistic content could be appropriated without compensation for training Meta's own AI systems.

While technical measures like robots.txt can instruct crawlers to avoid certain areas, they are not absolute access barriers. The publisher associations argue that by securing contractual consent through these new terms, Meta could override any technical objections raised by website operators.

In response to the escalating conflict, some German publishers are preparing concrete withdrawal scenarios. Ruth Betz, Managing Director of Zeitungsverlag Waiblingen, stated that her company has played out all scenarios and is prepared to withdraw from all Meta products to protect its content rights. Similarly, Funke Mediengruppe announced it is reviewing its engagement on Meta platforms, citing the one-sided reduction of value exchange between the social network and media companies.

Sources:heise online

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

The Berkshire Hathaway Owner's Manual was written by Warren Buffett in 1996. It is suitable for individual investors who wish to transition from speculation to investing and seek to build a robust long-term thinking framework. Readers are advised to begin with the section on "Business Principles Related to Ownership," focusing on chapters regarding the partner attitude, intrinsic value objectives, and the definition of look-through earnings, as these constitute the bedrock of Berkshire's investment philosophy.

Christiano's argument is attributed, not independently verified. His probability estimates rest on noisy benchmark extrapolation and his own judgment about how quickly automated AI R&D compounds. The METR report he cites was characterized by Zvi as confirming existing expectations rather than delivering a new measurement. The theft scenario depends on assumptions about state-level cyber capabilities and the difficulty of protecting weights across cloud deployments. None of these are established facts; they are the premises of one researcher's case for a specific policy action, presented with unusual concreteness about timelines and probabilities.