ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-188 posts

Refresh

AMD Demonstrates Trade-offs in Agent Inference Scheduling: Active Session Latency Drops 22%, New Request Wait Increases 49%

Material
2026-11-17 08:00 GMT+8

The traffic patterns of agent coding sessions differ fundamentally from those of standard chat requests. In a technical blog post about its ATOM inference engine, AMD described a typical case: one session involved 198 model calls, with call #48 sending 195.5K tokens, of which only 1.3K were new content. The corpus-wide median input length was 142,016 tokens, while the median output was just 444 tokens. This pattern—extremely long inputs, extremely short outputs, and highly shared prefixes—combined with the bursty parallelism of sub-agents and idle intervals caused by tool execution, creates service challenges that are entirely different from those posed by standard requests.

AMD categorizes these challenges into three layers: memory (maintaining reusable history), scheduling (balancing new requests against ongoing decoding), and compute (reducing per-step overhead in small batches). Among these, the trade-offs in the scheduling layer most intuitively reveal the inherent contradictions of agent services. ATOM’s PrefillDelayer allows a fixed number of decode steps to run immediately after each prefill completes, during which newly arrived requests are held and merged into a single prefill operation. Under 16 concurrent sessions, this strategy reduced P90 token inter-token latency from 31.6 milliseconds to 24.6 milliseconds—a 22% improvement—allowing active sessions to maintain a more stable output rhythm; however, the median time-to-first-token for new requests increased from 1,182 milliseconds to 1,766 milliseconds, a 49% deterioration.

This contradiction is sharper in agent workloads than in standard chats because the session tree structure amplifies latency. A parent agent waits for results from its sub-agents, meaning any delay on one branch postpones the completion of the entire task. Consequently, AMD enables PrefillDelayer only at concurrency levels of 16 and above, keeping it disabled in lower-concurrency interactive ranges, and provides different configuration combinations in deployment recommendations based on concurrency levels.

Optimizations in the memory and compute layers each have their own applicable conditions. ReplaySSM replaces the full state snapshot for every draft position in speculative decoding with a single checkpoint plus compact update records. In experiments with seven draft tokens, this freed up 24.49 GiB per GPU card, increasing available KV blocks from 8,325 to 18,813—a 2.26x increase—but this adds historical retention capacity rather than throughput or cache hit rates. Mono decode fuses multiple small kernels in individual KDA layers into two persistent kernels, reducing the target model's forward pass time from 22.0 milliseconds to 16.4 milliseconds in single-request tests; however, this optimization is currently enabled only at single concurrency, and its effects at higher concurrency levels remain unmeasured.

In terms of comprehensive effects, AMD reported that at 48 concurrent sessions, 94.1% of prompt tokens were reused rather than recomputed (with 77.3% coming from GPU HBM and 16.8% recovered from CPU DRAM), whereas the theoretical upper limit for this corpus is 97.2%. At 56 concurrent sessions, eight MI355X chips achieved a total of 17.5K tokens per second per chip. These figures come from AMD’s runs on the SemiAnalysis InferenceX AgentX benchmark and represent engineering references from the vendor under specific hardware and model combinations, not universally validated independent third-party performance benchmarks.

Sources:ROCm Blogs

Research

Databricks and Google Cloud Enable Bi-Directional Zero-Copy Data Federation Across Unity Catalog and Borderless Lakehouse

Material
2026-11-18 00:30 GMT+8

Data analytics platform company Databricks announced on November 17, 2026, a significant advancement in its partnership with Google Cloud: customers can now access a single copy of data from both Unity Catalog and Google Cloud's "borderless Lakehouse" without duplication. This bi-directional zero-copy federation allows engines like BigQuery to read tables managed by Unity Catalog, and vice versa, breaking down traditional silos between data warehouses.

Beyond data interoperability, the two companies are driving open standards for catalog metadata. In collaboration within the Apache Iceberg community, they have landed proposals for "catalog labels," enabling business semantics and governance tags (such as "domain=finance" or "pii=ssn") to travel with data across different catalogs. This mechanism is critical for enterprise AI agents, which rely on shared metadata to identify sensitive information, apply masking policies, and understand business context when searching for assets across platforms.

The feature is available now through Databricks' Lakehouse Federation and corresponding Google Cloud services. While this marks a key milestone in the open lakehouse architecture, specific production performance, latency under high concurrency, and governance maturity for large-scale deployments remain based on vendor self-reports and await independent benchmark verification.

Sources:Databricks

Research

AI Data Infrastructure Firm LanceDB Raises $57.5M Series B to Accelerate Multimodal Model Training

Quick take
2026-11-18 00:30 GMT+8

LanceDB, an AI data infrastructure startup, has announced the completion of a $57.5 million Series B funding round co-led by Bessemer Venture Partners and Innovation Endeavors. Other participants included Theory Ventures, Dimension, Swift Ventures, CRV, and Y Combinator, bringing the company's total funding to $98.5 million.

LanceDB develops database technology designed to accelerate artificial intelligence model training. Its core product is a unified multimodal data lakehouse that allows researchers, enterprise teams, and AI agents to access, find, prepare, and train on complex datasets containing text, images, video, and 3D sensor information without spending weeks managing underlying infrastructure.

According to SiliconANGLE, Chang She, co-founder and CEO of LanceDB, stated that there is a missing data layer in AI today, forcing research teams to spend significant time integrating scattered datasets before each training run. LanceDB's technology is built on Lance, an open-source data format that supports rapid retrieval and modification of individual records, making it particularly useful for handling petabyte-scale training datasets.

The company plans to use the new funds to expand its multimodal data platform, focusing on improving large-scale data curation, feature engineering, and data loading to efficiently supply thousands of GPUs during training. It will also deepen support for emerging AI workloads such as video generation, world models, and robotics.

Sources:SiliconANGLE

Research

Nvidia Former CEO Jensen Huang Says He Is Willing to Return If Board Decides It Best for Company

Structural
2026-11-18 00:09 GMT+8

Jensen Huang, the former Chief Executive Officer of Nvidia, issued a statement on November 17, 2026, explicitly expressing his willingness to return to the role if the company's board of directors believes it is in the best interest of shareholders, employees, and the company. The statement was initially reported by The Wall Street Journal and subsequently covered by major technology outlets.

"Since my departure, I’ve seen an outpouring of calls from Nvidia’s employees, investors, business partners, and customers for me to return as CEO," Huang said in the statement. "If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO."

This declaration comes amid significant leadership turmoil at Nvidia. Huang resigned last Thursday following reports of personal affairs, with Chief Financial Officer Colette Kress appointed as interim CEO. While more than 193,000 people, including prominent figures like Elon Musk, have signed a petition calling for Huang's return, reports indicate that Nvidia's board is currently not considering reinstating him.

Huang's public statement transforms the possibility of his return from private discussions or external pressure into a clear, conditional offer based on the board's decision. This adds complexity to the search for a permanent successor, particularly as Nvidia's stock price fluctuates due to the leadership crisis and the company faces heightened uncertainty as a core supplier in the AI infrastructure sector.

Sources:The Verge

Research

Apple Tightens macOS Permissions: AI Agents Face Stricter Warnings and Granular Controls

Material
2026-11-18 00:35 GMT+8

Apple has implemented significant changes to its security architecture in the second beta version of macOS 27.3, aimed at restricting unrestricted access by artificial intelligence (AI) agents to Mac data. This move directly addresses growing concerns about AI tools potentially over-accessing user privacy, exemplified by reports that Meta's Muse AI agent read iMessage contents without explicit consent to generate suggestions.

In the new version, the process for granting "Full Disk Access" and "Accessibility" permissions has become more cumbersome and transparent. Users can no longer simply drag an app icon to authorize access; they must select the requesting app from a list. The system displays detailed warning dialogs explaining exactly what data the app can view, modify, or extract once granted these broad permissions. Apple emphasizes that such extensive access puts all data on the Mac at risk.

Beyond the approval workflow, macOS 27.3 introduces new granular options. Users can now grant AI applications access only to specific folders or specific other apps, rather than the entire system. Additionally, when an app controls another app via "System & App Protection Override" permissions, a prominent icon appears in the menu bar, alerting users in real-time to background automation activities.

For developers, existing backup and assistive technology apps can request special "entitlements" to maintain user experience while complying with new security standards. However, permissions previously granted by ordinary users will not be automatically revoked; the new rules apply only to future requests. This initiative reflects Apple's attempt to balance the proliferation of AI features with the protection of user data sovereignty, building a visible firewall as AI agents increasingly penetrate the operating system's core.

Sources:Mac & i

Research

Google Threat Intelligence Reports Attackers Modifying Claude Code Config Files to Install Backdoors

Material
2026-11-17 09:22 GMT+8

Google's Threat Intelligence Group (GTIG) and Mandiant have released a quarterly report detailing new methods of malicious generative AI usage. The findings indicate a shift from single-prompt interactions to orchestrating multiple AI agents for automated intrusion tasks. A particularly notable trend involves attacks targeting the configuration files of AI coding tools.

According to the report, an actor named "MIDNIGHT NEPTUNE" targeted cryptocurrency-related organizations in North America and Europe. Using fake recruiter profiles, spoofed conference sites, and "ClickFix" tactics that trick users into executing commands, the group gained initial access. Once inside internal code repositories, the attackers modified the .claude/settings.json file associated with Claude Code. When developers launched Claude Code in these compromised environments, the system automatically downloaded and executed a backdoor named "SOMBERMEME." This technique turns the AI development tool itself into a vector for malware distribution.

Beyond supply-chain attacks on specific tools, AI-driven automated intrusions are accelerating. The report highlights the operational use of the open-source attack framework ARTEX. Released on GitHub in July 2026 as a penetration testing tool, ARTEX garnered significant attention, exceeding 12,000 forks before being made private on October 8. Wiz and Google reported in October that China-linked actors were using ARTEX to target IT and software companies. Other frameworks such as CAIRN, CyberStrikeAI, and PentAGI were also cited.

The automation of the vulnerability-to-exploit lifecycle is also increasing. An exposed server revealed a tool called "exploit_agent," which connects to multiple large language models to automatically compare pre- and post-patch code, identify vulnerabilities, and generate and execute exploit code. Targets included websites in France's media, entertainment, and education sectors.

Meanwhile, "LLMJacking"—the theft of AI compute resources—is becoming widespread. In one case handled by Mandiant, attackers generated over $100,000 in fraudulent bills in a single day, with cumulative damages reaching $1 million in the worst instances. Underground services known as "LLM Mixers" are reselling stolen AI credentials, with prices starting as low as $0.10 per million tokens.

Sources:テクノエッジ TechnoEdge

Research

"Asynchronous Methods for Deep Reinforcement Learning": Replacing Experience Replay with Parallel Agents for Efficient Training on a Single CPU

Verified 2026-11-18 00:04 GMT+8

Long-term reading · 《Asynchronous Methods for Deep Reinforcement Learning》(2016)

In 2016, the Google DeepMind team published the paper "Asynchronous Methods for Deep Reinforcement Learning", introducing the A3C algorithm. Prior to this, deep reinforcement learning primarily relied on an "experience replay" mechanism to stabilize training, which involved storing historical agent data and randomly sampling updates. While effective, this approach required significant memory and computational resources and restricted the use to off-policy algorithms.

The core innovation of this paper was abandoning experience replay in favor of an asynchronous parallel architecture. By running multiple independent "actor-learner" threads on a standard multi-core CPU, each thread interacted with the environment and computed gradients independently. This parallelism naturally decoupled the temporal correlation of data, making the training process more stable while enabling the use of a broader range of online policy algorithms (such as Sarsa and n-step methods).

Experimental results showed that the best-performing Asynchronous Advantage Actor-Critic (A3C) method surpassed state-of-the-art approaches in Atari game domains. Notably, it achieved this using only a single multi-core CPU rather than a GPU, halving training time. This finding demonstrated that efficient deep reinforcement learning could be achieved without expensive hardware or distributed clusters, providing a lightweight design philosophy for subsequent algorithms.

Sources:《Asynchronous Methods for Deep Reinforcement Learning》(2016)

Research

Berkshire Hathaway 1996 Shareholder Letter: Intrinsic Value, Cost Efficiency, and Acquisition Standards

Verified 2026-11-18 00:04 GMT+8

Long-term reading · 《Berkshire Hathaway Shareholder Letter 1996》(1996)

In Berkshire Hathaway Inc.'s 1996 shareholder letter, Warren Buffett elaborated on the core investment philosophy of distinguishing book value from intrinsic value. Buffett noted that although the company's net assets and book value per share had grown, true economic progress is measured by intrinsic value rather than accounting figures. He illustrated this through GEICO's "step acquisition" case, explaining that accounting standards may require impairment of previously held shares, which does not reflect actual economic value appreciation. This commitment to intrinsic value distinguishes Berkshire from firms solely focused on short-term financial metrics.

The letter also emphasized alignment between management and shareholders' interests, particularly through cost control and avoiding dilution of shareholder equity. Buffett mentioned that Berkshire's headquarters expenses were extremely low, amounting to only two basis points of net assets, embodying the principle: "We are here to make money with you, not from you." He further discussed the relationship between stock price and intrinsic value, arguing that over the long term, shareholder returns must align with the actual growth of underlying businesses. When market prices temporarily deviate from intrinsic value, astute investors may exploit this gap for profit, while long-term holders should focus more on business fundamentals.

Regarding acquisition strategy, Buffett outlined Berkshire's distinctive criteria: seeking companies with excellent economic characteristics and outstanding management. Using Kansas Bankers Surety and FlightSafety International as examples, he demonstrated how personal networks can uncover opportunities and how transactions can prioritize cash over stock issuance to protect existing shareholders. These cases highlight both Berkshire's execution capability and its prudent approach to capital allocation.

For contemporary readers, this letter provides a clear framework for assessing a company's true value and management integrity. It reminds investors not to be misled by short-term market volatility or complex accounting adjustments, but to focus instead on long-term economic moats and management's capital allocation skills. The "Berkshire Hathaway 1996 Shareholder Letter," written by Warren Buffett and published on February 28, 1997 (covering fiscal year 1996), is suitable for investors seeking to deeply understand value investing principles, corporate governance, and M&A logic. Readers are advised to begin with sections on "The Relationship Between Intrinsic Value and Market Price" and "Acquisitions in 1996" to grasp Buffett's core ideas.

Sources:《Berkshire Hathaway Shareholder Letter 1996》(1996)

Research

You’re all caught up in this view

Sign in / Sign up
ReadingResearchRadarInvestment framework