ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-1298 posts

Refresh

The EU's 'Simplified' Legislative Package: When Deregulation Rhetoric Meets the Enforcement Capacity Gap

Topic · 欧盟AI去监管化争议Material
2026-11-11 20:07 GMT+8

At the fifth AI Week hosted by the Baden-Württemberg State Data Protection Authority (LfDI Baden-Württemberg), Ralf Bendrath, advisor to the Greens/European Free Alliance group in the European Parliament, described the evolution of the European Commission's legislative discourse over recent years: from 'competitiveness' to 'simplification,' and now to open talk of 'deregulation.' He pointed out that this legal modification tool, known as 'Omnibus laws,' changes multiple regulations at once. While ostensibly aimed at reducing administrative burdens, it actually touches upon the rule system the EU established over the past decade to protect fundamental rights and the environment.

Specifically regarding the 'Digital Omnibus Legislative Package' proposed in November 2025, Bendrath listed several modifications he believes attack the core of the General Data Protection Regulation (GDPR): redefining personal data, limiting the information rights of data subjects, and issuing passes for artificial intelligence training. Additionally, the Commission plans to decide via implementing acts when pseudonymized data should be considered anonymous, and to expand the definition of 'scientific research' to cover the application of existing knowledge in 'novel ways'—which Bendrath considers primarily a backdoor for AI training. A survey by noyb showed that only 7% of Meta users surveyed in Germany wanted their data used for AI training.

However, not all simplifications are opposed. Bendrath welcomed the one-click rejection function for cookie banners, binding machine-readable browser signals (similar to the previously failed 'Do Not Track' standard), and a unified list across Europe triggering privacy impact assessments. He cited former European Central Bank President Mario Draghi to acknowledge that the GDPR is indeed subject to serious inconsistencies in execution across member states.

Max von Grafenstein, Professor at the University of the Arts Berlin, approached the issue from a practical perspective. He participated in writing a review by innovation economist Nicholas Martin, which showed that venture capital received by European startups has decreased since 2018, but von Grafenstein argued that the root cause lies in legal uncertainty and complexity, rather than the level of protection itself. He drew a key distinction: lowering compliance costs does not have to come at the expense of protection levels.

Von Grafenstein places his greatest leverage on the concept of 'state of the art' in Article 25(1) of the GDPR. In environmental law, this term means 'the most effective available technology' and has operated for decades; in data protection, it has hardly been activated. He suggested that regulators require companies to submit proof of effectiveness, such as demonstrating that cookie banners effectively reject tracking rather than merely complying formally. This would allow the market to spontaneously generate better solutions, just as diesel particulate filters once spurred an effective market.

An instance derived from his research project YourPrivacyLawyer is Consenter, a consent agent: after users set preset preferences, the agent automatically executes choices. Research shows that users made more targeted decisions, and services with identifiable high protection levels actually received more consents—turning data protection into a competitive advantage.

Sources:heise online

Research

SambaNova Deploys Zhipu's GLM-5.3 on SambaCloud, Claims Inference Speed Exceeds 200 Tokens/s

Quick take
2026-11-12 02:16 GMT+8

AI chip and cloud service provider SambaNova announced on November 11, 2026, that Zhipu AI's flagship open-weight model, GLM-5.3, is now running on its inference cloud platform, SambaCloud. Designed for complex software engineering and long-horizon agent tasks, the model achieves state-of-the-art results among open-weight models on benchmarks like Terminal-Bench 3.0 and DeepSWE v1.1, while using fewer output tokens per task than several closed frontier models, according to SambaNova.

To demonstrate real-world performance, SambaNova provided a trace from a typical agent run: over 17 minutes and 30 seconds, the model made 41 calls, processing 5.4 million input tokens but generating only about 42,000 output tokens. Because each step in an agent loop re-sends all previously seen context, input tokens grew from 38K on the first call to over 114K by the eighth. SambaNova noted that more than 91% of this input was served from cache, keeping per-call latency low enough for developers to complete entire tasks in real-time interaction.

Currently, GLM-5.3 runs at over 200 tokens/s on SambaCloud. This high throughput makes it suitable for production agent workflows that require frequent tool calls and fast feedback.

Sources:sambanova.ai

Research

Cloudflare Engineers Explain: Why Raft Stalls on the Global Internet, and How QuePaxa Bypasses Fault Detection with Optional Leaders

Topic · QuePaxa共识算法Material
2026-11-11 22:24 GMT+8

The default choice for distributed consensus systems is Raft or Paxos; they are mature, understandable, and work well in local area networks. However, Cloudflare’s Meerkat service needs to operate across continents, where connection links often exhibit high and volatile latency. The three authors, João Pedro Leite, Bob Halley, and James Larisch, explain that this environment led them to abandon all mainstream algorithms in favor of QuePaxa, published in 2023 by Tennage, Băsescu, et al.

The root of the problem lies in two structural flaws of leader-based consensus. The first is destructive interference: when a leader fails or slows down, multiple replicas simultaneously initiate elections, blocking each other from obtaining a majority vote, causing elections to repeatedly fail. Even if a network fault lasts only two seconds, an election storm can persist far longer than the fault itself. The authors term this metastable failure—a failure pattern that sustains itself even after the initial cause has disappeared.

The second flaw is fault detection itself. In asynchronous networks, a process cannot directly observe whether a remote process is dead; it can only observe whether expected messages arrive on time. Latency spikes and connection interruptions make healthy leaders appear failed. Raft typically deploys an election timeout at five to ten times the network round-trip time; on transoceanic links with round-trip times in the hundreds of milliseconds, detecting a failed coordinator may take nearly a second—during which time the system is completely stalled. Cloudflare experienced this firsthand in 2020: a partial failure of a core network switch caused an internal consensus system to degrade, slowing dashboard operations to one-eightieth of normal speed for six hours. No data was lost, but writes were entirely unavailable.

QuePaxa’s response consists of three design choices. First, the leader becomes optional. Any proposer can submit proposals; the leader is simply the one with the highest priority and the most likely to complete a decision in a single round-trip. When the leader goes down, clients continue sending requests to other proposers, and the system does not stop. Second, constructive interference replaces destructive interference. Proposers do not compete against each other’s proposals but instead collaborate to converge on the known highest-priority proposal—when a proposer receives a higher-priority proposal, it adopts it rather than fighting it. Third, hedging timeouts are not constrained by network latency. Clients send requests to all replicas, and non-leader replicas wait for a short local timeout before proposing their own. This timeout is not for fault detection and does not need to exceed network round-trip time, allowing it to be set very small without freezing the system.

Execution is divided into four phases. In Phase Zero, proposers randomly select priorities and submit proposals, adopting the highest-priority proposal received from a majority of recorders; if the leader confirms its proposal reached a majority of recorders first, it can decide immediately. Phase One spreads the adopted proposal to a majority of recorders. Phase Two allows proposers to query a majority of recorders for the highest-priority proposal seen in Phase One; if it matches their own, they decide. Phase Three prepares proposers who failed to decide for the next round by adopting the best proposal from a majority of recorders. The majority intersection property guarantees safety: any two majorities share at least one recorder, ensuring that decided proposals cannot be overturned by subsequent rounds.

Sources:Cloudflare Blog

Research

IBM Report: AI Is Top Stress Test for Digital Sovereignty, Confidence Outpaces Proof

Topic · IBM数字主权指数Material
2026-11-11 21:00 GMT+8

US technology company IBM published the "Proof, Not Promises: The Digital Sovereignty Index" report on November 11, 2026, revealing an "overconfidence gap" among large enterprises in the AI era. While 82% of surveyed executives expressed high or extreme confidence in their ability to provide evidence of sovereign control, only 26% of enterprises reached the index's "Leading Digital Sovereignty" tier. This discrepancy suggests that perceived technological control often fails to withstand formal scrutiny as AI adoption accelerates.

The report identifies AI as the primary stress test for digital sovereignty. Over the past three years, 94% of surveyed enterprises faced at least one sovereignty challenge, with 44% of these issues directly related to AI, such as provider failures, governance problems, or unexplainable model outputs. This proportion exceeded concerns about data access issues (33%), government restrictions (30%), and formal audits (38%). Among those affected, 81% were required to provide formal evidence of control to regulators, customers, or auditors.

IBM's index measures an organization's ability to retain and prove control across four pillars—AI, technology, operations, and data—evaluating six dimensions: responsibility, traceability, constraint, intervention, legitimacy, and exit. The study found that only 33% of respondents regularly test their ability to demonstrate sovereignty. Most enterprises rely on manual or time-consuming evidence collection (56%), with fragmented records across teams. Enterprises in the "leading" tier were more than twice as likely to resolve their most testing sovereignty challenges without material disruption (52%, compared to 22% for establishing-tier enterprises).

Jonathan Adashek, Senior Vice President and Chief Global Affairs Officer at IBM, stated that when business issues turn into sovereignty risks, boards need to demonstrate real control, including how they can adapt strategies or change critical providers. He noted that the core problem is not usually a lack of control, but that sovereignty efforts are not built to answer these verification questions smoothly under pressure.

Sources:IBM Newsroom

Research

Shopify Engineering Team: Why Search and Recommendation Share a Single Generative Model

Topic · ShopifyMaterial
2026-11-12 02:27 GMT+8

In industrial systems, search and recommendation typically follow separate pipelines: search handles explicit queries, while recommendation relies on implicit behavioral signals. The Shopify engineering team proposes a more fundamental observation—that both share the same set of inputs and objectives. A search query provides an explicit signal of intent, whereas recommendation infers intent from browsing and purchasing behavior; yet both require product knowledge and buyer context to identify relevant items. This common foundation makes joint training a natural choice rather than an engineering compromise.

To serve both search and recommendation with a single language model, a shared product vocabulary is essential first. Shopify’s catalog contains billions of products; the raw numeric product ID itself carries no product information, and treating every product as an independent output would create a vocabulary of billions of entries, resulting in extremely sparse training samples. Their solution is the semantic ID: first, an embedding model encodes product content into vector representations, then RQ-KMeans converts these embeddings into discrete three-token sequence codes, with each codebook level containing 8,192 entries. The key effect is that related products share encoding prefixes—for example, products starting with [6001] cover various vests and gilets, while [6001, 5561] further narrows down to padded gilets in different colors. This means that training data for one product can update the shared representation of its neighbors, providing a learning pathway for products with sparse interactions.

Does joint training truly benefit both tasks? Shopify provided ablation experiments to answer this. Removing search data caused recommendation Recall@50 to drop from a baseline of 21.0 to 20.9; removing recommendation data caused search OneDCG@10 to drop from 28.8 to 28.5, and search purchase Recall@10 to drop from 33.4 to 32.6. Both directions showed declines, indicating that shared training indeed produced cross-task gains. The team also found that while supervised fine-tuning teaches the model to reproduce historical samples, accurately predicting the next token does not fully equate to retrieval quality—multiple products can satisfy the same intent, yet training samples only record one observed purchase or click. To address this, they introduced reinforcement learning, using an LLM relevance judge to provide reward signals, which yielded an additional 1.5-point lift in OneDCG@10 for the search task.

Production deployment constraints dictate model scale and inference methods. Shopify positions the generative model as a candidate generator, sharing the latency budget of a live request with downstream ranking. They selected Liquid AI’s LFM2-2.6B, achieving a p95 latency of 25.5 milliseconds (with beam width 100), whereas the larger Qwen3-14B, despite higher quality, incurred a 108.4-millisecond latency that left insufficient time for subsequent ranking stages. For constrained decoding, they built upon the STATIC method to represent the catalog prefix tree as a sparse index on the GPU, enforcing constraints via a CUDA/C++ plugin within TurboMind, reducing constraint overhead from over 50 milliseconds to approximately 1 millisecond. Multi-token prediction further compressed the autoregressive generation of three codewords into a single forward pass, saving time to generate a larger candidate set for refinement by a lightweight ranker.

Sources:Shopify

Research

Amazon Launches Open-Source AWS Emulator Built by AI Agents to Solve Code Validation Bottleneck

Topic · AWS EmulatorMaterial
2026-11-12 01:30 GMT+8

Amazon Web Services (AWS) launched a new service called "AWS Emulator" on November 11, 2026. This free, locally hosted software suite replicates hundreds of the company's cloud services on developers' workstations. According to Matt Goldberg, Senior Principal Engineer at AWS, the entire service was designed and written exclusively by autonomous artificial intelligence agents, with not even a single line of code written or reviewed by a human engineer.

The tool aims to address a major bottleneck in application development arising from the adoption of AI coding agents like AWS Kiro and Claude Code. These agents generate code faster than humans can verify it, but validating all AI-generated code against real AWS cloud infrastructure has become impractical due to latency and cost considerations. Existing third-party emulators cover only a small fraction of AWS services and quickly drift out of sync.

Inspired by the manual simulation efforts required for AWS's early distributed locking service "Alf," Goldberg used AWS's public Smithy API models as specifications and the live AWS platform as ground truth. He prompted AWS Kiro to rebuild Amazon Simple Queue Service, which generated a functional replica within 15 minutes. Subsequently, dozens of autonomous agents running on Amazon Bedrock AgentCore were organized into three teams: one generating end-to-end fidelity test suites, another finding bugs and filing GitHub issues, and a third implementing fixes. An automated merge gate supervised by adversarial security agents checks for vulnerabilities before code enters the repository.

Built in Rust, AWS Emulator runs as a local HTTP server accessible via command-line interface on Linux, Windows, macOS, or Docker. Developers simply redirect their AWS CLI or SDK calls to localhost for rapid, offline validation. The project is open-sourced under a Creative Commons CC0 license, including its raw commit history and a public backlog of known discrepancies. Swami Sivasubramanian, VP of Agentic AI and Emerging Technologies at AWS, stated that this illustrates the evolution of the software engineer's role into a "systems orchestrator," defining boundaries and guardrails rather than writing grunt-work code.

Sources:SiliconANGLE

Research

Google DeepMind's AlphaProof Code Verified Signal's Cryptographic Core in Four Weeks

Topic · AlphaProofMaterial
2026-11-12 02:18 GMT+8

Google DeepMind has unveiled AlphaProof Code, an AI system that claims to have mathematically proven the cryptographic core code of the Signal messaging app meets security specifications in just four weeks, work that typically takes years.

According to DeepMind researchers Pushmeet Kohli and Swarat Chaudhuri, the system verified the actual cryptographic core code running on users' devices, not just theoretical designs. After experts defined formal security requirements, AI agents built on Gemini Carbon wrote over six million lines of Lean proof language code, flagging roughly one percent of the codebase for human review. Two deliberately planted bugs were caught immediately during testing.

Signal CTO Ehren Kret stated that this approach strengthens trust in data privacy. DeepMind plans to extend the method to power grids and medical devices through systems like CodeMender.

Sources:The Decoder

Research

FreeBSD Establishes AI Working Group to Draft Code Contribution Policy

Topic · FreeBSD AI政策Quick take
2026-11-11 19:30 GMT+8

The FreeBSD Project announced in its Q3 2026 status report that it has established an AI working group to draft a formal policy on artificial intelligence usage. This initiative follows an internal survey indicating strong developer support for clearer guidelines on AI's role in code contributions.

The poll revealed that 88% of respondents want the FreeBSD Project to adopt an AI policy. Regarding implementation details, 35% requested disclosure of AI-assisted code, while 22% believed such contributions should be treated like any other code submission. Only 16% advocated for outright rejection of AI-generated code. Additionally, 65% supported using AI-generated documentation provided it undergoes human review.

The working group is currently developing the policy draft, with final reports expected in Q4. This move highlights the open-source community's effort to establish transparent and actionable governance frameworks amidst the growing prevalence of large language models (LLMs).

Sources:phoronix.com

Research

Man Group Credit Team: This Credit Cycle Defined by Divergence, with AI Creating New Concentration Risks in the Investment-Grade Market

Topic · 曼集团信用周期分化论Structural
2026-10-28 07:00 GMT+8

In its October 2026 credit cycle outlook, the Man Group credit team put forward a central judgment: this cycle is characterized by divergence rather than broad-based credit deterioration. This assessment is not based on optimistic assumptions but stems from a data-driven review of private credit, high yield, leveraged loans, emerging market corporate bonds, and the investment-grade market—each segment showing resilience, albeit distributed extremely unevenly.

AI acts as the most concrete amplifier of this divergence. In the private credit sector, high-profile defaults in the software and SaaS industries in 2026 sparked questions about the asset class's overall resilience. However, data from the Cliffwater Direct Lending Index shows that realized credit losses were only 0.47%, below the long-term annual average of approximately 0.98%. Nevertheless, the non-accrual rate rose from 1.27% at the end of 2025 to 1.85% in Q1 2026, implying a drop in recovery rates from roughly 53% to 49%. Man Group points out that the key lies in AI's disparate impact on different borrowers: mission-critical enterprise software deeply embedded in client operations and difficult to replace represents a fundamentally different credit proposition compared to commoditized, point-solution SaaS products. Lenders must determine whether AI adoption is supporting a borrower's pricing power and customer retention or threatening both simultaneously.

In the investment-grade market, AI creates risk in a different way—not through defaults, but through concentration. Hyperscale technology companies are estimated to spend up to $800 billion on AI infrastructure this year, financing these expenditures through corporate bonds and asset-backed structures, thereby generating significant supply of long-dated bonds. The spread on the JPMorgan USD Investment Grade Hyperscaler Index reached approximately 140 basis points by late August 2026, roughly double its level at the start of 2025. Man Group warns that if these investments fail to generate expected cash flows, hyperscalers could become a major source of pressure on investment-grade returns—particularly given their projected share of the investment-grade index is expected to reach 10% within four years. Notably, this exposure is not limited to USD portfolios: Alphabet has borrowed approximately $60 billion across six currencies, demonstrating how the same issuer can appear in portfolios diversified by currency or geography, while previously independent funding pools increasingly share the same underlying credit risks.

Man Group simultaneously notes that the current environment increases the cost of misjudgment. Spreads per unit of leverage in the high-yield market have compressed to among the tightest levels on record, with the CCC-to-BB spread ratio sitting at the 98th percentile of its range over the past decade. Leveraged loan borrowers, due to floating-rate debt passing benchmark rate changes more rapidly into interest expenses, may face cash flow pressures earlier than fixed-rate high-yield bond issuers. PIK (Payment-in-Kind) interest income accounts for 6.92% of total direct lending index income, nearing post-pandemic lows; however, the real issue is whether its usage is concentrated among already stressed borrowers.

Sources:man.com

Research

Manufacturing Company Bright Machines Pushes Data Out of Air-Gapped Factories: Edge Capture and Federated Query Replace Nine-Service Pipeline

Material
2026-11-12 07:42 GMT+8

When customers visit Bright Machines' factories, they stand before assembly dashboards watching production lines run, then request specific data: yield by station, cycle time trends, downtime by cause, and batch genealogy. Bright Machines has this data internally; operations teams view it daily. However, customers cannot access it directly over the network because the factory's security design is air-gapped: there is no inbound access, very limited outbound paths, and zero tolerance for risk. Previously, the only way to share data was to rerun reports locally and send them via email. Each new report required the data engineering team to reshape data through multiple ETL jobs, taking hours to days.

Bright Machines builds software-defined automated production lines for manufacturing clients, running inside client factories. Each line runs a full stack of local services—including proprietary MES, integrations with client MES, operator screens, robot workflow orchestrators, and assembly dashboards—all persisted to a local PostgreSQL cluster on-site. The data drives KPIs for yield, finance, operations, quality, vendor component quality, NCR tracking, batch genealogy, OEE, and shift-level downtime and cycle-time analysis.

The team initially attempted to rebuild a cloud copy of each factory database based on application events, spanning nine loosely coupled services: a message broker, custom shoveler, queue, object storage, several serverless functions, a cloud relational database, a scheduler, an ETL pipeline, and a popular open-source BI tool with basic RBAC support. While this architecture worked fine on the factory floor, it failed off-site: pipelines silently dropped data when schemas changed frequently; dead-letter queues were difficult to monitor, so missed or duplicated events went unnoticed; once data landed in a bucket, no one could confirm its accuracy or source; and there was a lack of demonstrable governance and role-based access control.

The requirements for the new architecture were prioritized as follows: access control must match the existing role restrictions within the factory; there must be a single auditable outbound path explainable to any security team; freshness must be measured in minutes; and the edge must use open formats. The core rule of the threat model is: data can be pushed out, but nothing external can reach in.

This ruled out two obvious pull-based solutions: Lakeflow Connect ingesting directly from PostgreSQL, and PostgreSQL exports picked up by Auto Loader. Pull-based ingestion requires a network path from the cloud to the factory database and running ingestion compute in an environment outside the account where the data resides. The question became: how do we achieve the same governed result while keeping every component touching raw factory data within our own environments?

The answer was to flip the direction: do not pull from the cloud, but push from the edge. Bright Machines' architecture consists of five steps. First, Debezium Server tails the PostgreSQL replication slot at the factory edge, writing changes as an append-only Apache Iceberg change log to local object storage. If the WAN drops, the backlog becomes local object files rather than write-ahead logs piling up on the production database. Second, an edge component copies each object to a bucket in the site's cloud account—this is the factory's sole outbound data path, with credentials restricted to writing only to this bucket. Cross-account replication moves the objects to the analytics account. Third, Iceberg tables are registered in AWS Glue, and Unity Catalog federates Glue as a read-only catalog, allowing over a hundred tables from the factory's PostgreSQL databases to be queried directly in Databricks without copying. Fourth, in Databricks SQL, Bronze views collapse the change log to the current state, and a Lakeflow job rebuilds Silver models every ten minutes. Fifth, AI/BI dashboards read data from the Silver layer with role-based permissions, and OpenSharing shares the same governed data zero-copy to customers.

Sources:Databricks

Research

ElevenLabs Launches Enterprise AI Voice Agent Marketplace with Salesforce, IBM, Zendesk

Quick take
2026-11-11 20:00 GMT+8

US-based voice AI company ElevenLabs launched the ElevenLabs Marketplace on November 11, 2026, a curated platform helping enterprises discover industry-specific AI voice agents built by technology partners. Launch partners include Salesforce, IBM, NiCE, Twilio, and Zendesk, covering financial services, insurance, healthcare, telecommunications, and retail.

Mati Staniszewski, CEO of ElevenLabs, stated that enterprises can find agents that already fit their workflows without building from scratch. The company plans to support purchasing eligible partner solutions through existing ElevenLabs commitments starting next year.

Sources:ElevenLabs

Research

Man Group: Macro Dual-Perspective Country Allocation Model Supplements Systematic EM Strategy with Low-Correlation Alpha

Material
2026-10-21 06:00 GMT+8

In systematic emerging markets (EM) equity strategies, stock selection is the primary source of alpha; company fundamentals—earnings dynamics, balance sheet quality, and valuation signals—do indeed have predictive power at the security level. But country-level returns are a different matter: macro forces such as monetary policy cycles, terms of trade, external demand, and financial conditions often dominate. Man Numeric researchers Mickael Nouvellon and Dawei Zhou point out a key but long-ignored fact: risk contribution at the country level in EM is significantly higher than in developed markets (DM), meaning the opportunity set for country allocation is structurally larger, while stock-selection models are typically not designed to exploit it.

The most intuitive approach is to standardize relative valuation and 12-month price momentum, then combine them into a cross-sectional long/short country signal. However, Man Numeric finds structural flaws in this method: both valuation and momentum are slow-moving variables, resulting in country exposures that can remain static for years rather than rotating with the macro cycle. Take South Korea as an example: due to its chronic valuation discount relative to EM peers (driven by corporate governance concerns and a concentrated, export-oriented industrial structure), the model has maintained a long position in Korean equities for almost the entire past two decades, largely independent of the prevailing macro environment. More critically, this signal overlaps heavily with the information used by stock-selection models; its country tilts are often positively correlated with a fund manager's existing stock-picking alpha, offering limited diversification benefits.

Man Numeric’s alternative is to combine basic economic logic with machine learning techniques, utilizing high-quality, point-in-time macroeconomic data to build a country allocation model. This framework operates through two complementary perspectives: the time-series perspective compares each country against its own history, capturing idiosyncratic dynamics (e.g., whether Brazil’s current macro environment is improving or deteriorating relative to its historical norm); the cross-sectional perspective compares all countries simultaneously, simulating a global allocator’s relative value judgment (e.g., whether Brazil is more attractive right now than Indonesia or South Africa). These two signals are fused into a single view that considers both country-specific dynamics and relative positioning, enabling the identification of how oil price increases affect net importers like South Korea differently than net exporters like Saudi Arabia, and modeling the spillover effects of US interest rates or Chinese industrial output on specific EM economies through trade linkages.

The case of South Korea from 2021 to 2026 demonstrates the model’s dynamic adaptability. In 2022, the model’s conviction shifted from sovereign credit concerns to external trade vulnerabilities—the energy price volatility triggered by the war in Ukraine made the trade balance the largest drag, with the score hitting bottom in August. In early 2023, a sector rotation signal indicated that domestic momentum was turning before the global backdrop caught up; thereafter, the model remained bullish during periods of stabilizing trade balances and softening core inflation, until a new sector rotation in September 2024 pulled the score back to cycle lows. In early 2026, a sharp drop in silver and gold prices, combined with strong semiconductor exports and softening oil prices, reinvigorated Korea’s export engine, leading to the most significant improvement in the score. This stands in stark contrast to the static, multi-year tilts generated by fundamental models.

Sources:man.com

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

However, a disturbing contradiction emerges. A Bitkom survey shows that 90% of enterprises believe targeted reform of the GDPR is necessary, but looking closely at their demands: 88% complain that specific provisions are unclear, 61% complain about a lack of guidance from regulators, and 76% hope for a centralized database of regulatory decisions. In other words, enterprises are asking for more regulatory guidance and more reliable precedents, not less regulation.

Ironically, Baden-Württemberg is doing the opposite. According to the budget draft of the Green-Black coalition government, Tobias Keber's data protection authority will gradually cut 40% of its positions by 2029, despite receiving additional tasks. A broad alliance comprising the Network for Data Protection Expertise, the Chaos Computer Club Stuttgart branch, Digitalcourage, the German Association for Data Protection, FIfF, and the Humanistic Union Baden-Württemberg has sent an open letter opposing these cuts, citing Article 8(3) of the Charter of Fundamental Rights of the European Union and Article 52(4) of the GDPR, demanding that regulators must be equipped with sufficient resources to 'effectively perform their duties.'

Von Grafenstein calls this the 'scaling problem' of data protection: regulators can no longer review all cases. He points out that legislators can 'lower' data protection costs simply by cutting regulatory resources—'this is location policy.' His proposed way out is to let the market work for regulators: requiring companies to submit proof of protection effectiveness, so that companies themselves have the incentive to develop better solutions. But he also admits that someone still needs to review these proofs.

The core tension of this argument lies in the fact that the political drive for deregulation comes from the digital industry's annual lobbying expenditure of 151 million euros (with Meta alone spending 10 million euros), while actual compliance needs point toward more, not less, regulatory interaction. When both happen simultaneously—rules are loosened and enforcement resources are cut—the actual efficacy of protection levels faces systemic risk beyond paper promises. Bendrath offers a pragmatic bottom-line judgment: unlike enacting new laws, there are already rules in effect here; if parties cannot agree, everything remains as it is.

Cloudflare reported internal evaluation results: in an environment where servers span nearly all continents, Meerkat continuously processes requests and never fully stalls. Under normal conditions, more than 95% of decisions are completed in Phase Zero by the leader in a single round-trip; when the leader fails, the decision rate drops to approximately 65%, recovering once a replacement leader is installed after one minute. When decisions are not completed in Phase Zero, throughput briefly drops to one-third of normal because three round-trips are required instead of one. The authors emphasize this is a trade-off, not a free lunch: more active proposers mean more broadcasting and a lower probability of deciding in a single round, but they believe trading bandwidth for uninterrupted operation is worthwhile.

Boundaries to retain: The QuePaxa paper predicts a 50% decision probability per round, though actual performance is higher due to convergence effects, with no guaranteed upper bound on the number of rounds. Cloudflare’s performance data is self-reported internally, and the algorithm itself has not undergone independent real-world testing on the same scale as Raft or Paxos. The authors also note they are extending the original crash-stop design to support crash-recovery, which increases engineering complexity and requirements for correctness verification.

Practical implications for readers: If your consensus system runs in a local area network with predictable latency, Raft remains the correct choice. But if your replicas span continents, network latency is unpredictable, and the cost of false fault detection is system-wide stall, QuePaxa’s optional leaders and constructive interference mechanisms offer an alternative path with specific design rationale and preliminary production evidence. The core insight is not that QuePaxa is faster than Raft, but that it does not stop completely when the network deteriorates—it is fast when the network is normal, and it continues to advance when the network is not.

It must be clarified that all the aforementioned performance figures are self-reported production metrics from Shopify, without independent verification. Online gains (orders +7.95%, conversion rate +9.85%) come from Shop home feed recommendation experiments, while search quality improvements stem from internal OneDCG@10 evaluations. The applicability conditions of this architecture are tightly coupled with Shopify’s product catalog structure, request latency budgets, and participating merchant scale; systems with different product distributions, latency constraints, or interaction densities may not be able to directly reuse the same decisions. The team itself notes that the next goal is to integrate intent understanding, candidate generation, and ranking into a single model, introducing online continuous training—meaning the current architectural positioning (as a candidate generator) is itself a phased engineering choice rather than a final form.

Man Group’s conclusion is that the opportunity lies in distinguishing borrowers and structures capable of carrying their debt through a more demanding phase of the cycle from those priced as though nothing will change. This argument presupposes that divergence does not evolve into broad-based deterioration—if growth slows or financing costs remain elevated for an extended period, the impact is unlikely to be distributed evenly across all issuers. This article was drafted with AI assistance, but the arguments and data attribution originate from the named author team.

A key design decision was why register Iceberg tables in AWS Glue rather than pointing Databricks directly at the storage bucket. A change log on object storage is a stream of newly added metadata files. A small function advances the pointer for each table in Glue only after the new metadata file has fully replicated, and only forward. Glue holds the pointer, Unity Catalog holds the permissions, and Databricks reads only the validated version. This gives Bright Machines precise control over when each table moves forward—a catalog hop owned by them, with guards on the pointer.

Another easily overlooked operational issue is distinguishing between an idle production line and a broken pipeline. When a line stops for a shift change, no new rows arrive, which looks exactly like a pipeline break. Row counts cannot distinguish these two extremes. Debezium updates a one-row heartbeat table in each source database every five minutes, regardless of whether the line is running. This heartbeat update travels the same path as production data. The first task of the Lakeflow job checks the heartbeat; if any heartbeat is older than thirty minutes, the task fails, turning the entire job red. Dashboards continue to display the last good data rather than presenting stale data as current.

Actual results: A row committed on the production line is queryable in Databricks within one to two minutes at steady state, with Silver models and dashboards refreshing every ten minutes. Heartbeat checks detect pipeline faults within thirty minutes, compared to days previously. The first customer-visible dashboard went live about two months after kickoff. The proof of concept was completed by two engineers in approximately six weeks.

The migration was phased into four stages: piloting on a PoC line; going live on the highest-value production line; running parallel against the old snapshot path for row-by-row validation; and finally performing a hard cutover for customer dashboards.

An unexpected finding was that scale appeared in the wrong dimension: moving hundreds of gigabytes per day, but in the form of hundreds of thousands of small objects, meaning every design decision ultimately concerned object rate rather than byte volume. The team stated that if they were to start over, they would build a reconciliation path on day one—a sync driven by create events never goes back to retrieve missed objects, and downstream systems won't notice until a query touches them.

Bright Machines summarized five lessons useful for other teams needing to share data from restricted sites: choose a platform starting from your existing security model rather than approximating it; design the unique outbound path first and buffer behind it; maintain control over when tables move forward; build data once and let permissions handle the rest; and assert freshness on heartbeats, not on row counts.

It should be noted that this is a self-reported case study by the Bright Machines team on the official Databricks blog; all performance figures are self-reported and not independently verified. The architecture is deeply tied to the AWS, PostgreSQL, and Databricks stacks, but the authors note that the Databricks-side design is identical across any cloud, so migration only requires changing the edge sync target, cross-account replication, and catalog hop resources.

Next, Bright Machines plans to build natural language analytics (Genie), custom AI agents for operations, and predictive analytics on top of the governed data, enabling production line operators, quality engineers, and maintenance technicians—who have domain expertise but are not SQL users—to get answers grounded in the same governed Silver data by asking questions.

In backtests from 2008 to 2026, this macro model produced a realized Sharpe ratio of 0.85, a substantial improvement over the fundamental model’s level of under 0.5, while turnover was 28% lower and maximum peak-to-trough drawdown was 4.6% compared to the fundamental model’s 8.7%. Because it is built on macroeconomic data rather than valuation or price information, its daily returns correlate with those of the fundamental model at only 0.08, constituting a largely uncorrelated source of alpha.

However, Man Numeric explicitly notes several important boundaries: all results are gross returns before transaction costs; both models rebalance daily, implying considerable annualized turnover, which means translating them into actual tradable strategies requires addressing cost issues through position smoothing, reduced rebalancing frequency, or explicit trading constraints; the single-country study on Korea is merely an illustration of the model’s mechanics, and its results should not be generalized to other markets; the fundamental model generated over 40% of its value added since 2008 in the last 12 months alone, and the model was particularly vulnerable when Russia became uninvestable in 2022. These are self-reported backtest results by Man Numeric as a strategy provider and have not yet been independently verified.

Man Numeric concludes that this is not a replacement for stock selection, but a complement to it—a systematic overlay that transforms country risk from an unmanaged residual into an intentional, data-driven source of alpha. As macro differentiation across emerging markets continues to deepen, they believe the rationale for integrating this capability into institutional investment processes will only strengthen.