LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-1298 posts

Atualizar

Alibaba Announces HKEX Approval for T-Head Semiconductor Spin-off IPO

Tópico · 平头哥分拆上市Material
2026-11-11 11:08 GMT+8

Chinese technology giant Alibaba Group announced late Tuesday that the Hong Kong Stock Exchange (HKEX) has allowed the company to spin off its semiconductor subsidiary, T-Head. Alibaba stated it plans a public offering for T-Head shares in Hong Kong, subject to final approval from the exchange and the China Securities Regulatory Commission (CSRC).

According to disclosures, Alibaba had already filed confidentially on September 11 to list T-Head on the HKEX. The announcement did not disclose a specific date or proposed size for the listing. Following the news, Alibaba's Hong Kong-listed shares traded approximately 1% lower on Wednesday morning, down about 29% year-to-date.

T-Head is Alibaba's subsidiary focused on developing artificial intelligence chips, positioned as a potential rival to Nvidia in the Chinese market. In late September, Alibaba unveiled its next-generation AI chip, the Zhenwu V900, claiming it delivers three times the performance of its predecessor, the Zhenwu M890 released in May. This move aims to counter U.S. restrictions on China's access to advanced chips by promoting domestic alternatives.

In recent months, Chinese financial regulators have increased scrutiny on tech IPOs and the rush of capital into the sector. Meanwhile, other Chinese chip companies, such as Baidu's Kunlunxin and memory chip maker YMTC, have reportedly delayed their own listing plans.

Fontes:CNBC

Pesquisa

German University Releases Open-Source AI Model AnySleep for Second-Level Sleep Analysis

Tópico · AnySleepVisão rápida
2026-11-12 01:00 GMT+8

A research team at the University of Applied Sciences Aachen (FH Aachen) in Germany has developed AnySleep, an open-source artificial intelligence model designed to analyze human sleep data. The findings were recently published in npj Digital Medicine, a peer-reviewed journal belonging to the Nature Portfolio.

Traditional sleep analysis typically records brain activity via electroencephalography (EEG) and divides nightly data into 30-second segments to assign sleep stages. This coarse granularity often misses brief wake events or subtle changes in brain activity that last only a few seconds. AnySleep processes nocturnal measurement data to map shorter intervals on a second-by-second basis, creating a more detailed overall picture of sleep phases. Study author Stephan Bialonski noted that if the evaluation interval is too long, these changes are not adequately considered, leading to inaccurate results.

The model was trained on sleep data from approximately 20,000 subjects, varying widely in age, living conditions, and health status. While AnySleep has already succeeded in identifying differences in sleep structures between healthy individuals and those with diseases, providing clues to certain clinical patterns, Bialonski emphasized that the technology is not yet suitable for making medical diagnoses. The project continues in cooperation with several hospitals.

The source code for AnySleep is publicly accessible and can theoretically be installed on any computer. Additionally, FH Aachen offers a browser interface called SomnoBot, allowing users to utilize the AI tool without coding knowledge. According to the press release, sensitive data never leaves the local computer, making the work data protection-compliant even in strict clinical research environments.

Fontes:t3n.de

Pesquisa

Semantic Diffing of Formal Policies with Anti-Unification

Material
2026-11-12 07:33 GMT+8

When an automated policy generation process translates the same natural language requirement into two formal policies, the first question facing reviewers is not which one is correct, but where they actually differ. Textual diff tools fail systematically here: swapping the order of two conjuncts, rewriting an external destination as non-internal, or replacing one predicate with an equivalent one will flag many lines as changed, while the real difference in authorization behavior—such as changing And to Or—may be lost in the noise.

The AWS Machine Learning blog introduces semi-persistent, an open-source engine that computes semantic diffs of formal policies using anti-unification. Anti-unification stems from Plotkin's classic 1970 work, aiming to find the most specific common pattern between two expressions. The engine runs over an e-graph, which tracks expressions known to be equal; users inform the engine which syntactic changes do not alter meaning by declaring algebraic laws (such as associativity, commutativity, and idempotence of conjunction) and domain rules (such as equating an external destination with a non-internal one).

The engine outputs a single term where shared structure appears only once, and each divergence is marked as a Variants node containing the complete subtrees of both candidates. Selecting the first child of each Variants node reconstructs candidate A, while selecting the second child reconstructs candidate B. This means the anti-unification result is a lossless merge of the two inputs: it does not discard information but folds identical parts together and retains and locates differing parts.

The blog demonstrates how this works with a concrete example. The requirement is: sensitive data sent to an external destination must be encrypted. Candidate A is written as sensitive(data) ∧ external(dst) → encrypted; Candidate B is written as ¬internal(dst) ∧ sensitive(data) → audited. A textual diff would flag condition rearrangement, predicate rewriting, and different conclusions. The engine initially reports two divergences: one in the conditions (external vs. ¬internal) and one in the conclusions (encrypted vs. audited). The user declares the domain fact as a rewrite rule—equating external with ¬internal—and after saturation, re-runs the diff. The condition divergence disappears, leaving only the conclusion divergence: Candidate B audits where the requirement demands encryption. A single Variants node precisely points to the location requiring review.

In a more complex example, two encodings of a cross-Region replication policy exhibit five textual differences. Three are purely syntactic: reordering of conjuncts, swapping equality arguments, and expanding approvedRegion into two specific Regions. After declaring And as associative, commutative, and idempotent, and Eq as commutative, the first two are absorbed; adding a domain rewrite equating approvedRegion to the disjunction of the two Regions absorbs the third. The remaining two divergences are genuine behavioral differences: one encoding uses Or instead of And in the confidential object clause (allowing either encryption or same-account status, rather than requiring both), and the other uses strict less-than instead of less-than-or-equal at the size threshold (objects exactly at the threshold take a different path). Semantic diffing compresses five textual changes into two reviewable semantic divergences, reducing the compression ratio from 0.59 to 0.40.

Fontes:Amazon Web Services

Pesquisa

AWS Proposes Three Code-Level Checks to Detect GPU Silent Data Corruption Bypassing Hardware Monitoring

Tópico · AWSMaterial
2026-11-12 07:30 GMT+8

The fundamental reason silent data corruption (SDC) is difficult to detect is that it occurs internally within the compute path. A GPU's multiply-accumulate unit returns a wrong but plausible value—for example, 4.0001 instead of 4.0000—without triggering any error signal. Every existing integrity check guards a different boundary: ECC protects memory, Xid captures driver and hardware faults, DCGM deep health checks run on idle nodes, and NVLink and PCIe checksums verify whether transmission bytes arrived intact. Checkpoint checksums confirm storage round-trip consistency. None of these checks whether the output of the arithmetic logic path is correct, because there is no ECC or parity on that path.

This means detection must come from the training task itself. AWS adds three checks to the PyTorch DistributedDataParallel training loop, arranged by where faults are likely to land. The first is a non-finite gradient check: before gradient aggregation, GPUs are flagged if their local gradient norm is NaN or Inf; after aggregation, the reduced gradients are checked for NaN or Inf. The second is a ring replay audit: before the optimizer step, each GPU recomputes a micro-batch from another node's GPU, comparing loss values bit-for-bit and, under deterministic kernels, comparing gradient fingerprints. This can catch finite but incorrect computed values. The third is weight fingerprint comparison: after the optimizer step, fingerprints of all replicas' weights are compared, catching corruption in the update step or in stored weights.

There is strict logic governing the execution order and timing of the three checks. Faults in the compute path must be caught before the optimizer step, because once erroneous gradients propagate through aggregation to all replicas, all weights remain consistent, making it impossible to distinguish afterwards. Conversely, faults in the optimizer step can only be seen after it occurs. Ring replay is feasible because between the backward pass and the optimizer step, all replicas hold identical weights, allowing any GPU to recompute another GPU's micro-batch and expect identical bits.

In tests training a 953-million-parameter model on 32 H100 GPUs, running only the gradient hooks and NaN/Inf checks (excluding audits) increased step time by approximately 7.4%; running deterministic audits every 100 steps added about 10.1%; running audits every 10 steps added about 34.7%. Across 15,320 fault-free audit steps and 3,080 weight checks, no alarms were triggered, yielding an upper bound on the false-positive rate of approximately one in five thousand (at 95% confidence) using the rule of three.

This solution has clear boundaries. It was tested only on DDP layouts; tensor parallelism, pipeline parallelism, and FSDP sharding require separate adaptation. One-time finite faults that happen to fall on non-audit steps will not be captured by any check. Deterministic audits use math attention kernels, whose cost grows with the square of sequence length, meaning overhead in long-sequence training may far exceed the measurements in this article. AWS explicitly states that synthetic tests do not establish precision in production environments and recommends retaining manual approval before replacing nodes.

Fontes:Amazon Web Services

Pesquisa

AWS Enterprise Guide: Connecting OpenAI Codex to Amazon Bedrock via the Simplest Path, Where Configuration Boundaries Are More Critical Than Authentication

Material
2026-11-12 07:35 GMT+8

Platform teams face a specific contradiction: developers are already using Codex to explore code, make changes, and handle repetitive engineering work, but if each team configures model access independently, it creates a second stack for identity, routing, and operations within the organization. The solution offered by the AWS Machine Learning blog is not to unify tools, but to establish unified decision rules for access paths.

The core argument is explicitly stated in the article's conclusion: start with the simplest access path that meets the requirement, prove identity and model connectivity with a small pilot, and only add gateways or client controls when addressing specific needs. The article categorizes paths into three layers. The first layer is Native AWS Access, suitable for scenarios where IAM and identity controls are sufficient; developers obtain temporary credentials through AWS IAM Identity Center, and Codex calls Amazon Bedrock Runtime directly. The second layer is AgentCore Gateway, suitable for teams requiring AWS-managed routing infrastructure or Amazon Bedrock Guardrails. The third layer is the LLM Gateway, suitable for organizations already operating a gateway or needing to enforce controls in the request path. Custom integrations are advanced variants of the third layer, not quick-start options.

The decision rules themselves are not complex, but the validation sequence is more easily skipped than the path selection. The article requires pilots to include three checks: first, use the AWS CLI to call Bedrock Runtime directly to confirm IAM permissions and regional availability; second, configure Codex to point to approved models; and finally, run a secure read-only task. Two users are required to test identity isolation—one person can only prove successful login, while a second person verifies whether attribution remains independent.

The section most noteworthy for platform engineering teams is 'Configuration Boundaries.' The article explicitly states that model providers, credentials, and region settings belong to user-level or managed configuration because they represent endpoints and organizational choices, not repository defaults. Project-local .codex/config.toml is repository-scoped configuration; it should be subject to code review but cannot serve as an enterprise enforcement boundary. AGENTS.md is suitable for carrying persistent repository guidance, such as test commands and coding standards, but it too is not an enforcement boundary. This distinction means that if an organization's security policy relies on repository files to constrain model access, any developer could bypass it by modifying repository configurations. True boundaries exist in user-level configurations, IAM policies, and requirements.toml.

At the governance level, the article treats telemetry, billing, and identity records as evidence sources answering different questions. Codex OpenTelemetry describes client activity (token usage, turns, API requests, tool calls), while Amazon Bedrock service metrics describe server-side activity (latency, input/output tokens, error rates). AWS identity records and the Cost and Usage Report provide attribution and billing evidence. The article specifically warns against treating token volume as a cost report and emphasizes that telemetry does not establish correctness or safety—it only records what happened, not whether it should have happened.

Fontes:Amazon Web Services

Pesquisa

Databricks Security Team: Key to Detecting Credential Attacks Lies Not in the Login Itself, but in Cross-Identity Behavioral Correlation

Tópico · DatabricksMaterial
2026-11-12 07:00 GMT+8

Modern cyberattacks increasingly begin with a single legitimate login. In a recently published detection guide, the Databricks security team describes a typical scenario: a data engineer's personal access token leaks from continuous integration logs, and within minutes an attacker uses that token from a new IP address to list secret scopes, catalogs, and repositories, then reads a secret holding a service principal credential. An hour later, that service principal—not the engineer itself—creates functions, adds itself to an admin group, and exports several init scripts. Each step alone might look like normal platform usage, but together they constitute a complete intrusion.

This article, written by Solomon Boyd, Alexis Merritt, Arun Pamulapati, Sam Pezzino, and David Veuve, proposes a six-stage model to organize the detection of credential-driven attacks: gaining access, mapping the environment, harvesting more credentials, finding executable paths, establishing persistence, collecting data, and cleaning up traces. However, the article's truly transferable insight is not these six stages themselves—they are no different in essence from generic attack chain models—but rather the judgment regarding event serialization: attackers rarely move through these six stages in a straight line; they rely on non-linear sequencing, using one set of credentials to validate the environment, then retrieving another set of credentials to return through a new identity or workspace that has no associated reconnaissance activity.

This implies a specific detection trap: if you only track the identity corresponding to the stolen credential, you will lose the trail the moment the attacker switches to a new identity. In the above scenario, the engineer's token was used to read secrets, but creating functions, joining admin groups, and exporting scripts were all completed by the service principal. The article explicitly states that searching for activity scoped only to the engineer's identity would completely miss these events. The correct approach is to correlate secret reads with the subsequent first activity of the new identity, IP changes, and user-agent changes, treating events scattered across different identities as fragments of the same intrusion.

Regarding specific detection patterns, the article provides some actionable threshold suggestions. For example, it is unlikely that a person operating via a UI interface would call five or more different discovery-class APIs (listing secret scopes, catalogs, schemas, external storage locations, storage credentials, repositories, etc.) within one minute. A typical sequence for credential harvesting is the rapid, repeated appearance of listScopes, listSecrets, and getSecret across multiple scopes. Privilege escalation focuses on whether events such as putAcl, addPrincipalToGroup, or setAdmin originate from unfamiliar identities or execution contexts—particularly permission changes made from inside a job runtime, which deserve more scrutiny than similar operations performed by an administrator in the console.

Fontes:Databricks

Pesquisa

NVIDIA Principal Cloud Architect Mark Chmarny: AICR Validates Configuration, k8s-aibom Tracks Workloads; The Boundary Between Them Is the AI Governance Gap

Material
2026-11-12 01:00 GMT+8

When teams deploy AI workloads using NVIDIA Dynamo, Kubeflow, or Slurm, a reproducible deployment bundle does not tell you what is actually running on the cluster. Application teams define and update workloads independently of platform recipes, and mutable image tags like :latest add further complexity, meaning two nodes under the same tag may be running different code. NVIDIA Principal Cloud Architect Mark Chmarny explains in an NVIDIA technical blog how this gap is closed by integrating Google's open-source project k8s-aibom.

Chmarny describes the two systems as opposite ends of the same problem. AICR operates at the runtime deployment layer—drivers, operators, kernels, Kubernetes components, and deployment configuration—and asks: Is the GPU infrastructure built from a known and validated configuration? k8s-aibom sits higher up the stack and later in the cluster lifecycle; it observes Kubernetes workload and Pod APIs and generates a CycloneDX 1.6 AIBOM record for each identified AI workload, asking: What AI assets are actually running on that infrastructure?

Chmarny points out that neither system alone is sufficient to support AI governance, and the boundary between them is exactly where accountability is currently missing. AICR records qualified deployment intent and compares it against the cluster at a point in time; k8s-aibom observes supported workloads and records what it finds, including the confidence level with which each attribute was determined. The former provides control and reproducibility, while the latter provides workload discovery and traceability.

Specifically, AIBOM helps teams discover unregistered or shadow AI workloads in selected namespaces, identify models and inference runtimes, track external AI APIs and vector databases, detect changes or drift in AI dependencies, and support security reviews, incident response, and compliance audits.

The integration works by making k8s-aibom a qualified component within an AICR recipe. Two stock recipes select it by default: h100-gke-cos-inference and h100-gke-cos-inference-dynamo; other GKE recipes can opt in during generation via the --runtime-inventory flag. Chmarny emphasizes that qualification is not a rubber stamp: AICR binds upstream source, chart, CRD, and digest-pinned controller images into one tested set, verifies provenance against Google’s expected build identity, reviews rendered RBAC and security defaults, and exercises installation, upgrade, health checks, and uninstallation on the tested deployment paths.

Using AICR v1.2.0 pinning k8s-aibom 1.6.2 as an example, the workflow proceeds from recipe generation through packaging, deployment, and validation without manual assembly or tuning. After deployment, the k8s-aibom controller runs in the k8s-aibom-system namespace. During validation, the controller must be available, the cluster-scoped AIBOMControllerConfig must report Ready=True, and both CRDs must report the storage version expected by the pinned chart.

Namespaces opt in via the label aibom.k8saibom.dev/enabled=true. Chmarny specifically notes that this label controls which namespaces emit AIBOM records but does not narrow the controller’s read scope—the informers still cache cluster-wide workload and Pod specs, including container arguments and inline environment variables. No sink or secret access is configured by default, so no data leaves the cluster without an explicit decision.

Fontes:NVIDIA Technical Blog

Pesquisa

American Century Investments Co-Chief Investment Officer: AI Capital Expenditure Resembles Cloud Computing Investment from Two Decades Ago, Not the Dot-Com Bubble

Material
Verificado 2026-11-12 07:19 GMT+8

In an analysis released in October 2026, Keith Lee, Co-Chief Investment Officer of Global Growth Equity at American Century Investments, and Ryan Walker, Client Portfolio Manager, attempted to answer a question confusing the market: Why have growth stocks with strong earnings growth experienced severe volatility?

Their explanation begins with the duration mechanism. The value of growth stocks comes primarily from future earnings; when long-term interest rates rise, investors increase the discount applied to distant cash flows, making these long-duration assets the first to suffer. The article points out that yields on U.S. long-term Treasury bonds recently reached their highest level since 2002, driven by three factors: U.S. federal debt surpassed $40 trillion in mid-August; inflation has remained above the Federal Reserve's target for five consecutive years; and in September 2026, the Federal Reserve raised interest rates under the leadership of new Fed Chair Kevin Warsh.

The impact of rising interest rates on AI investment is twofold. On one hand, higher discount rates directly depress growth stock valuations. On the other hand, a large amount of AI infrastructure investment relies on debt financing—estimated that the four largest AI cloud providers and Meta doubled their debt scale to $350 billion over the past five years, issuing approximately $121 billion in bonds alone last year. Rising interest rates mean increased financing costs for these projects, making it more difficult to achieve returns.

However, the core argument of the two authors is not pessimistic. They explicitly state that they do not believe current AI investment is comparable to the excessive spending of the dot-com bubble period. They cite a historical analogy: twenty years ago, when Amazon, Microsoft, and Alphabet invested heavily in building cloud infrastructure, investors similarly worried whether these expenditures would generate returns. The authors point out that while these investments were costly and put pressure on current earnings through channels such as depreciation, they ultimately became key drivers of growth and earnings. The authors believe that current AI capital expenditure may create similar long-term benefits, because the demand for computing power required to run AI applications far exceeds the supply capacity of existing infrastructure.

But they also retain important qualifying conditions. Past success does not guarantee future results, and the returns on current AI investments remain uncertain. Not all companies will benefit equally from AI spending, and some investments may prove more successful than others. In the authors' view, companies with broad exposure to AI models, semiconductors, and cloud infrastructure may be in a more favorable position as AI applications continue to expand.

A notable market detail is that earnings growth itself has already improved valuations. S&P 500 earnings grew by more than 50% over the twelve months ending in June 2026, with Big Tech profits growing over 100% year-over-year, and the semiconductor industry nearing 140%. Earnings growth increased the denominator in the P/E ratio, and combined with weaker stock price performance, made growth stock valuations more attractive relative to their own history and even to value stocks. But the authors also acknowledge that the stock market focuses more on future earnings than past performance; analysts expect earnings growth of nearly 30% next year for large-cap S&P 500 stocks—whether this expectation can be realized is precisely the core tension behind the current volatility.

Fontes:americancentury.com

Pesquisa

Oxford Professor Philip Stier: AI Models Face Fundamental Difficulties in Predicting Out-of-Distribution Extreme Events

Tópico · AI气候模型极端预测Material
2026-10-22 06:00 GMT+8

At the University of Oxford, Professor of Atmospheric Physics and Director of the Intelligent Earth Centre, Philip Stier, discussed with Jason Mitchell, Chief Investment Officer for Responsible Investment at Man Group, how artificial intelligence is transforming climate modeling during an episode of the firm's "A Sustainable Future" podcast. Stier’s core assessment is that purely data-driven AI models face fundamental difficulties when predicting out-of-distribution extreme events, a limitation that cannot be overcome simply by increasing the volume of training data or running more simulations.

Stier defines the challenge of predicting climate change as an "out-of-distribution problem": we are required to forecast states that have never occurred in the training dataset. Traditional physical models encode physical laws through systems of equations; these equations remain valid even under unprecedented conditions. In contrast, AI models learn patterns from historical observations. When future states exceed the coverage of the training data, the model lacks a basis for extrapolation. He cites the current El Niño event as an example: because observational records are too short, such extreme ENSO (El Niño-Southern Oscillation) events may not be present in the training data, making pure AI-driven climate predictions unreliable in such scenarios.

A deeper issue lies in the biases inherent in the training data itself. Stier points out that most AI weather models currently rely on reanalysis data for training—a combination of historical observations and outputs from traditional physical models. This means AI inherits not only the flaws of observational data (such as insufficient coverage in developing countries and limited funding for public observation systems) but also the parametrization biases of the physical models themselves. For instance, reanalysis data does not idealize the diurnal cycle of precipitation, which is related to the parametrization of convective clouds; AI models will learn this same bias. While acceptable for short-term weather forecasting provided predictive skill improves, this inherited bias can distort judgments on key processes like cloud feedback in climate prediction.

Stier’s warnings regarding the misleading effects of large-sample ensembles are particularly specific. He notes that running "very, very large" ensembles creates a false sense of certainty—results are often reported to high decimal precision—but this does not mean true uncertainty has been reduced. If the model itself contains biases affecting the tail distribution, then even perfect sampling of the model’s tail yields incorrect estimates of extreme events. He mentions that academia is researching methods to guide generative models to efficiently sample rare tails, but emphasizes that these approaches are only meaningful if the underlying model is reliable and well-calibrated.

For practitioners using AI-assisted investment research, Stier’s argument provides a clear boundary for validation: when the prediction target involves out-of-distribution events, the model’s output should not be assigned the same confidence level as in-distribution predictions. Similar situations in financial markets include unprecedented liquidity crises, tail risks of new asset classes, or market behavior following structural breaks. Precise probabilities generated by large-sample Monte Carlo simulations or backtesting are themselves misleading if the underlying model fails to capture critical mechanisms. Stier believes the pragmatic path to resolving this issue is to combine physical constraints with data-driven methods, rather than using AI to completely replace equation-based modeling.

Fontes:man.com

Pesquisa

TSMC Board Approves NT$1.34 Trillion Capex Budget, Plans Up to $1 Billion Investment in OpenAI and Anthropic

Tópico · 台积电Material
2026-11-12 05:30 GMT+8

Taiwan Semiconductor Manufacturing Company (TSMC) held a board meeting on November 10, 2026, approving a capital budget of approximately $40.927 billion (equivalent to NT$1.34 trillion). This budget covers advanced process nodes, advanced packaging, upgrades to mature and special process capacities, factory construction, and R&D for 2027. Simultaneously, the board passed a strategically significant resolution: to subscribe to Anthropic common stock and OpenAI preferred stock with a maximum limit of $500 million each, totaling an investment ceiling of $1 billion. Additionally, up to $500 million was approved for establishing a semiconductor ecosystem strategic investment company. It is crucial to note that these are approved limits, not yet actual expenditures or completed subscriptions.

This move marks an extension of TSMC's participation in the AI industry. Previously, TSMC primarily served as a foundry for chip designers like Nvidia and AMD; now, it plans to hold direct equity stakes in AI model developers. TSMC expects capital expenditures for 2026-2028 to increase significantly compared to the previous three years, reflecting the continuous push of AI computing demand on the simultaneous expansion of high-end processes and packaging technologies. Through equity investments, TSMC may aim to more closely bind top-tier AI customers and ensure visibility of future high-end capacity demand.

However, the market should interpret this signal with caution. The long-term capital budget approved by the board cannot be directly viewed as new spending in 2026, and the two equity investments have not been settled. This transition from 'manufacturing partner' to 'shareholder' reflects TSMC's intent to seek deeper strategic binding in the AI wave, but its actual execution effect and specific impact on the industry landscape remain to be observed.

Fontes:TechOrange 科技報橘

Pesquisa

First Eagle Global Value Team: How the AI Capex Cycle Is Squeezing the Market and Creating Bottom-Up Opportunities

Material
2026-10-21

In its Q3 2026 Market Overview, the Global Value Team at First Eagle Investments presented a core thesis: the artificial intelligence (AI) capital expenditure cycle is "sucking the oxygen out of the market," pushing up real interest rates, and weighing on other assets within the traditional 60/40 portfolio. The team believes this dominance should influence expectations for market performance going forward.

The mechanism supporting this argument lies in the fact that massive capital expenditures related to AI construction not only drive up input costs but also lead to a severe narrowing of market leadership. Data shows that while the NYSE FANG+ Index surged 13.1% in the third quarter, the S&P 500 Index rose only 2.3%, the MSCI EAFE Index advanced 0.8%, and the Bloomberg Global Aggregate Index fell 3.5%. This divergence indicates that behind the positive surface performance of benchmark indices, losers outnumber winners.

The team points out that currently record corporate profit margins are supported by US fiscal deficits (accounting for 6% of US GDP), low US household savings rates, and US private nonresidential investment reaching dot-com bubble levels. However, they warn that historically, the emergence of transformative technologies has often been followed by sharp corrections after periods of overinvestment; frictions are already emerging in the data-center narrative, with rising financing costs and slower permitting squeezing marginal economic benefits.

Regarding macro risks, First Eagle specifically highlighted the yield on 10-year US Treasury bonds rising to a 24-year high of nearly 5.3%, and sovereign debt concerns reflected by France's borrowing rate approaching 5%. They believe that although the market appears expensive from a top-down perspective, selective opportunities may exist from a bottom-up view in sectors that have de-rated due to rising capital costs—such as healthcare, housing, and software—as well as international stocks that remain cheap relative to US equities.

This viewpoint provides investors with a framework for understanding current changes in market structure: namely, that rising real interest rates and fiscal vulnerability are reshaping asset pricing beneath the AI boom. For investors seeking value outside the crowded AI trade, focusing on sectors being "squeezed" out of the market by high capital costs may be a reasonable strategy.

Fontes:First Eagle Investments

Pesquisa

Google Research Proposes a Personalized Coaching Framework Using World Models to Simulate Counterfactual Scenarios for Healthy Habits

Tópico · SensorWMVisão rápida
2026-11-12 03:54 GMT+8

Knowing your health age only tells you where you stand today. To turn that into an actionable plan requires answering a counterfactual question: What will happen to my body over the next 12 weeks if I sleep 30 minutes longer or add a brisk walk to my daily routine? Simply plugging new step or sleep targets into a static calculator misses a key fact—the body is an interconnected system that adapts as a whole. When you consistently move more and sleep better for several weeks, your resting heart rate and heart rate variability gradually improve as well, creating additional downstream health benefits.

In research published on November 11 by Google Research, Zeinab Esmaeilpour and Ahmed A. Metwally developed an action-conditioned world model (SensorWM) to address this issue, pre-training it on de-identified data from approximately 2.28 million person-weeks collected from roughly 100,000 Fitbit users. Given two weeks of baseline wearable data, SensorWM predicts how 20 physiological indicators will evolve over 12 weeks under different step, sleep, and aerobic exercise regimens. It does not simply extrapolate trends; instead, it simulates a causal chain: altering input behaviors, observing physiological adaptation, and tracking cascading effects.

However, simulation alone does not equal good advice. The most aggressive workout plan is meaningless if it leads to burnout by the second week. The team designed a feasibility-constrained optimizer that balances projected health-age improvements against each individual's historical habit flexibility and clinical safety boundaries: individuals with high habit flexibility may receive a more ambitious target (reducing health age by 2.3 years), while those with more constrained routines receive a gentler ramp-up plan (reducing health age by 1.3 years). A language model then translates the top-ranked pathways into a supportive, progressive 12-week coaching plan.

In blinded evaluation experiments, five certified lifestyle coaches selected the counterfactual-simulation-based plan as the best overall plan in 71.4% of assessments, while five board-certified primary care physicians judged that 99.2% of the plans fell within safe exercise intensity limits. However, the authors explicitly emphasized that these evaluations were offline retrospective studies, the model is a research prototype rather than a clinical diagnostic tool, and it has not yet been validated through prospective trials to demonstrate actual health improvement outcomes.

Fontes:research.google

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

The method has clear boundaries. It does not prove any candidate is correct: both translations can share the same error, and shared errors will not appear as divergences. It relies on user-declared algebras and rewrites; if domain knowledge is incomplete, true equivalences may remain as Variants nodes. The search itself may also be incomplete—when using default Monte Carlo search rather than exact solving, results are labeled budget rather than exact, meaning a smaller common pattern may not have been found. The engine does not evaluate policies, query SMT solvers, or judge which translation matches the requirement; it only organizes information, directing reviewer attention to locations that might genuinely change behavior.

For teams using large language models to automatically generate formal policies, the practical value of this method lies in the fact that, after generating multiple candidates, there is no need for manual line-by-line comparison or reliance on the model's own judgment. Semantic diffing provides a deterministic, repeatable starting point for review: examine the positions marked by Variants nodes, and resolve them one by one using schema validation, replay, example input/output pairs, or human judgment.

The engine is open-sourced under the Apache-2.0 license and requires a Rust toolchain to build, taking S-expression files as input. The blog provides several runnable policy examples covering comparative candidate translations across four policy formats: Cedar, Dogwood, OpenShell, and CEL.

From a practical deployment perspective, this scheme links detection, localization, and recovery into a closed loop. Once a check triggers an alarm in the training loop, the alarm carries a suspect_window recording the range of steps since that check last passed, indicating the fault was active during one of those steps. On SageMaker HyperPod clusters orchestrated by Amazon Elastic Kubernetes Service (Amazon EKS), Kubernetes node names contain instance IDs, so mapping a suspect GPU's rank to a specific instance requires only a single lookup. Operations personnel can use this to stop the task, isolate the node, check DCGM metrics, and upon confirmation, complete hardware replacement via node health labels or the BatchReplaceClusterNodes interface. The training script then resumes from a checkpoint saved before the suspect_window.

The article also distinguishes between detection and response boundaries. Once an intrusion is confirmed, the response must cover all paths created by the attacker, not just the original credential. In the example scenario, revoking the engineer's token would leave the service principal's credential, its admin group membership, and the function it created intact. A complete response includes: deactivating the identity at the account level, revoking personal access tokens, exporting and stopping related compute, rotating every key the attacker read, removing persistence paths created by the attacker, dropping functions and revoking grants, and restoring context-based ingress policies. The article specifically mentions an easily overlooked time window: OAuth tokens already held by the service principal remain valid for up to an hour after deactivation.

At the prevention level, the article lists context-based ingress control as the strongest control against credential reuse, because it evaluates the requester's identity, source, and purpose rather than relying on a single network perimeter. For organizations moving away from VPNs, this provides a mechanism to limit the available scope of stolen tokens even without corporate network anchors.

It should be noted that the detection thresholds proposed in the text are heuristic suggestions based on the Databricks security team's operational experience with its platform, not measurement results validated by public benchmarks. The article itself advises readers to treat these guidelines as a starting point and to tune them according to the normal automation patterns of their own environments. Furthermore, the specific implementation of this method is tightly bound to Databricks' audit log architecture and the Genie detection application; migrating to other platforms requires re-identifying corresponding event fields and query patterns.

Regarding resource impact, upstream measured k8s-aibom v1.6.1 on GKE using NVIDIA Dynamo and AICR configurations. The test used CPU-only nodes with workload Pods remaining in pending state, so these figures describe controller inventory overhead: approximately four API server requests per minute, average controller CPU usage of 2.2 millicores, and a controller memory working set of 28 to 34 MiB.

Chmarny also clarifies where the integration deliberately stops. Runtime AIBOM writes are Kubernetes resources, but the observations themselves are unsigned; signing would make captured observations tamper-evident but could not prove every workload declaration true or establish successful execution. k8s-aibom can verify Sigstore signatures, enforce signer-identity constraints, check Rekor inclusion, and assign qualifying claims verified confidence, but AICR disables this verification by default. Correlating platform sources with runtime AI inventory remains future integration work.

A notable boundary: AIBOM is built from workload specs, so it can exist before any Pod is running. An AIBOM report may show Ready=True while the Pod is still pulling its image. It tells you what was declared and what was resolved once the Pod exists, but it does not prove successful scheduling or execution.