LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-18133 posts

Atualizar

How Slurm HRES Plugins Integrate Quantum Processors into Supercomputing Scheduling: Why Traditional Resource Models Fail to Adapt to QPUs

Tópico · Slurm HRESMaterial
2026-11-18 01:00 GMT+8

Quantum processors are being installed in supercomputing centers worldwide, running alongside high-performance computing systems. For scientists in fields such as chemistry, weather modeling, and engineering simulation, the QPU is simply another accelerator—like CPUs and GPUs—that they wish to request, reserve, and execute using familiar methods. However, prior to Slurm 26.11, the scheduler lacked a native way to model QPUs as schedulable resources. Integration solutions at each site were repeatedly customized, users faced different interfaces depending on the site, administrators struggled to maintain proprietary code that was difficult to upgrade, and QPU vendors had to rebuild integrations for every customer.

The root of the problem lies in the structural mismatch between the physical attributes of QPUs and Slurm's existing resource models. QPUs are typically cluster-wide accelerators and do not belong to any single compute node. Slurm's Generic Resource (GRES) mechanism, used for GPUs, binds devices to specific nodes—if that node fails, the QPU becomes unreachable until an administrator reconfigures the cluster. Another common approach combines licenses with SPANK plugins: while licenses represent cluster-wide resources, SPANK only intervenes after a node has been allocated, meaning a job can hold compute nodes idle while waiting for the QPU. Furthermore, multiple SPANK plugins for different QPUs compete for the same command-line options, and many rely on Slurm's internal APIs, creating a risk of breakage with every upgrade. Crucially, Slurm itself cannot detect whether a QPU is undergoing calibration or is offline, leading it to launch jobs that cannot utilize the device.

The Hierarchical Resources (HRES) plugin interface introduced in Slurm 26.11 addresses these gaps. HRES was added to Slurm 25.05 as a test feature and represents an evolution of licenses. Like a license, an HRES is a count of a cluster-level resource that does not belong to any single node; unlike a license, this count is associated with a specific set of nodes—the ones permitted to use the resource. Sites describe the QPU as a cluster-level HRES in resources.yaml, specify which nodes can access the device, and point to a short Lua script for the resource logic. This division of labor, with resources.yaml holding site-specific endpoints and policies and the Lua plugin handling device logic, allows the plugin to be written once and reused across sites.

The core capabilities of the plugin manifest during key callbacks in the job lifecycle. The maintenance callback queries middleware for QPU status at set intervals, setting or clearing disable_layer flags to stop or resume allocation. The controller job prerun callback performs admission checks after a job is scheduled but before nodes are allocated, ensuring that jobs do not hold idle nodes while waiting for the QPU. Because the prerun holds the controller's job lock, time-consuming reservation operations are placed in an optional asynchronous prerun, with the payload sent by Slurm to the job nodes. If a job fails to start, a revoke callback cancels the reservation, and a controller epilog callback releases the resources upon job completion.

Fontes:NVIDIA Technical Blog

Pesquisa

Former Nvidia CEO Jensen Huang Publicly States Willingness to Return, Citing Board Approval Condition

Tópico · NVIDIA Huang Resignation & ProbeMaterial
2026-11-18 05:22 GMT+8

Former Nvidia Corp. Chief Executive Officer Jensen Huang issued a public statement on social media platform X on November 17, 2026, expressing his willingness to return to the company. This marks his first public comment since resigning last week following revelations of an affair with Alice Huang, a former cloud executive suspected of being a Chinese spy.

"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 wrote. "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."

The statement follows reports by the Financial Times that Huang had privately discussed a comeback with confidants. Additionally, an online petition demanding his reinstatement has garnered support from prominent tech figures including Elon Musk, Y Combinator CEO Garry Tan, and Epic Games CEO Tim Sweeney, although some signatures appear questionable.

Despite Huang's expressed willingness, his return remains uncertain. CNBC reported that Nvidia's board is not considering bringing him back. During today's earnings call, Nvidia Chair Harvey Jones reiterated that the board does not intend to reinstate Huang. Sources previously told Reuters that Huang misled the company's general counsel and chief marketing officer about his relationship with Alice Huang.

Nvidia Chief Financial Officer Colette Kress currently serves as interim CEO. The Wall Street Journal reported that Nvidia has hired an executive search firm to find a new permanent CEO, evaluating both internal and external candidates. Meanwhile, Nvidia established a new national security and export control oversight committee today, supervised by Kress, to strengthen governance around trade compliance and risk management.

Fontes:SiliconANGLE

Pesquisa

NVIDIA Reports Record $115.8B Q3 Revenue Amid Jensen Huang Resignation; Board Establishes Export Control Oversight Committee

Tópico · Jensen Huang ResignationEstrutural
2026-11-18 07:46 GMT+8

US chipmaker NVIDIA reported its third quarter fiscal 2027 earnings on November 17, 2026, with revenue hitting a record $115.8 billion, up 103% year-on-year. This surge was driven by the first quarter of production shipments for the Vera Rubin platform. However, the report came just five days after CEO Jensen Huang resigned, with CFO Colette Kress leading the call as interim CEO.

Regarding the leadership transition, new Board Chairman Harvey Jones stated that Huang stepped down due to "disclosure of personal behavior inconsistent with the company's code of conduct, standards, and cultural norms." The board unanimously accepted his resignation and has no intention of reinstating him. Jones noted that the board is evaluating both internal and external candidates for a permanent CEO role. Additionally, NVIDIA established a National Security and Export Control Oversight Committee, composed of independent directors, senior external advisors and legal counsel, to manage engagement with US government agencies.

Despite the turmoil, Kress emphasized business continuity, stating employee commitment remains high and noting recent visits from TSMC leadership. Financially, GAAP net income grew 9% sequentially, lagging behind 22% operating income growth due to equity holding variances: Q2 saw $7.8 billion in net income from equity securities versus $1.8 billion in Q3, which included impairments. Networking revenue exceeded $21 billion.

On demand, Kress cited a cloud industry backlog of $2.5 trillion, up from $2 trillion three months ago. Frontier labs are expected to account for around 15% of data center revenue this year. Hopper GPUs have recontracted at up to twice their original price, and B300 rates have risen over 60%, despite token prices falling nearly 60% over the last six months.

For China, Q3 revenue from licensed Hopper products was $2.4 billion. Kress confirmed that future guidance excludes all licensed Hopper data center products, as shipments to China have ceased. Gross margins slipped to 74.9% in Q3, with Q4 guided lower at 71% to 72% due to supply costs and legacy pricing commitments.

Fontes:morethanmoore.substack.com

Pesquisa

AWS Announces NVIDIA Dynamo Support in SageMaker HyperPod Inference

Material
2026-11-18 07:33 GMT+8

Amazon Web Services (AWS) has announced that Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo. This integration aims to address bottlenecks in efficient Large Language Model (LLM) serving, such as GPU memory coordination and long prompt processing, by combining Dynamo's distributed inference capabilities with HyperPod's infrastructure management to simplify multi-node deployments.

NVIDIA Dynamo is an open-source distributed inference framework supporting multi-node deployment, disaggregated prefill and decode, and KV-aware routing for engines like vLLM, SGLang, and TensorRT-LLM. Within HyperPod, the integration is implemented through custom resource definitions and operators, including the HyperPod Inference Operator, Dynamo operator, Grove (for coordinating pod groups), and the Kubernetes AI (KAI) Scheduler (for GPU resource scheduling).

Users can configure deployments via two paths: "Bring-your-own DynamoGraphDeployment (DGD)" for teams wanting direct control over the serving graph or reusing existing DGDs; and "InferenceEndpointConfig (IEC) with dynamoSpec" for teams deploying new endpoints using a unified configuration format. The latter allows the operator to automatically manage model weight loading, endpoint registration, and autoscaling.

AWS provided benchmark results based on the Qwen3-235B-A22B-FP8 model. On two p5en.48xlarge instances (16 H200 GPUs total) using the TensorRT-LLM engine, disaggregated serving achieved a 2.05x improvement in highest throughput compared to aggregated serving (increasing output tokens per second per GPU from 101 to 207), while meeting performance targets of P90 time-to-first-token under one second and median generation speed of at least 50 tokens per second per user.

The feature is currently available in Amazon SageMaker HyperPod Inference, supporting single-instance and multi-node deployments with integrated node health monitoring and fault recovery mechanisms.

Fontes:Amazon Web Services

Pesquisa

How Salesforce's Internal Agent Maintains Four Million Production Forecasts: Diagnosis, Testing, and Pull Request Submission Only

Material
2026-11-18 06:02 GMT+8

Salesforce's internal forecasting platform generates nearly four million time-series forecasts every night, determining cloud capacity reservations, workload migrations, and infrastructure budget trajectories. The configuration for each forecast instance is remarkably simple—specifying which data to use, which model to run, and the required accuracy threshold. This simplicity allowed the platform to scale from five use cases to over a hundred without a proportional increase in data scientists. However, launching a forecast is just the beginning; maintaining its accuracy is an ongoing burden. As the metrics being forecast constantly change, models drift, and each forecast requires a data scientist to identify the cause of changes and test remediation plans.

To address the scale of model tuning, Salesforce built an internal self-evolving data science agent. Its responsibilities are clearly defined: investigate forecast drift, test candidate fixes against recent historical data, and then submit a pull request—a proposed code change that must be approved by a data scientist. It cannot deploy models or modify production forecasts directly. The term 'evolving' refers to the agent reusing lessons learned from past investigations and reviewer feedback, rather than autonomously deploying models into production.

The entire process consists of six steps. Step 1 is Localization: identifying which forecasts have deviated from their targets. Step 2 is Diagnosis: examining data changes and searching Slack for relevant context. Step 3 is Remediation: attempting new models and parameter settings, including Salesforce's proprietary time-series foundation model, Moirai 2.0. Step 4 is Re-evaluation: backtesting the new models against past data. Step 5 is Decision-making: submitting a pull request only if the new model outperforms the current model by at least 5% on unseen data. Step 6 is Evolution: updating the investigation playbook with successful fixes for reuse in future investigations.

Architecturally, the agent does not replace existing ML technology stacks but sits on top of them. Experiments still run on standard tools: Kubernetes and Spark handle computation, while MLflow tracks and ranks results. The agent's new role is to act as a conversational interface between Slack and these engines—data scientists ask questions in Slack, and the agent determines why a forecast missed and which fixes to attempt. A single investigation may involve dozens of backtests, far exceeding the agent's own compute capacity, so it dispatches each task to the appropriate engine: small jobs run on Kubernetes (one pod per variant), while large jobs run on Spark (testing one configuration across multiple sequences simultaneously). MLflow scores every candidate, sends the top-ranked models back to the agent, which selects the optimal solution and automatically submits a pull request.

Salesforce reports that, to date, the agent has proposed over 100 forecast model fixes, including swapping models from Prophet to Moirai, adjusting change-point sensitivity, and modifying training windows. Backtests observed a 14% reduction in sMAPE error, with some forecasts showing significant improvements. A typical agent investigation takes 27 minutes, compared to hours required for manual debugging in the past. The article cites a specific case: a storage forecast for an AWS availability zone showed stable growth in usage during the first half of 2026, but growth flattened in June as new workloads began landing in other regions. The old forecast continued extrapolating the original trend; without intervention, this would have led to reserving approximately 100 TB of storage that would go unused. The agent's pull request swapped the model for Moirai, which successfully tracked the slowing growth rate.

Fontes:Salesforce

Pesquisa

Microsoft Unveils Copilot Workspaces, Autopilot Agent, and AI Cost Tracking Tools at Ignite 2026

Material
2026-11-18 07:14 GMT+8

US tech giant Microsoft announced a suite of new enterprise AI features at its Ignite conference on November 17, 2026, aimed at addressing workflow collaboration and cost control. Key announcements include Copilot Workspaces, a "multiplayer" collaborative surface for human-AI interaction, and Autopilot, a persistent background agent with its own identity.

According to ZDNET, CEO Satya Nadella introduced Copilot Workspaces during the keynote, describing it as an evolution of Copilot Notebooks where people and agents build on each other's work. The feature is currently available in the Frontier program.

The Autopilot agent represents a shift from prompt-driven interactions to autonomous execution. It can monitor Teams channels, follow up on emails, and manage projects without waiting for user input. Microsoft emphasized that each Autopilot instance has its own Microsoft Entra Agent ID and a named manager, with all sign-ins and actions logged for audit trails. This design integrates AI agents into existing IT security and compliance frameworks.

To tackle growing concerns over AI investment returns, Microsoft also released Agent 365 Business Analytics and Cost Management tools. Business Analytics provides insights into AI spending and value, measuring ROI based on "work completed" rather than just token usage. Cost Management allows IT leaders to set spending policies, limits, and alerts, and to control which models are available to teams.

These updates reflect a broader industry trend: as AI adoption deepens in enterprises, managing multi-agent workflows and precisely quantifying high AI costs have become critical barriers to effective deployment.

Fontes:ZDNET

Pesquisa

US AI Voice Firm ElevenLabs Releases Scribe v3 Live Medical for Real-Time Clinical Transcription

Material
2026-11-17 20:00 GMT+8

US-based AI voice technology company ElevenLabs released Scribe v3 Live Medical on November 17, 2026. This specialized model is designed for real-time speech-to-text transcription in healthcare settings, built upon its latest base model, Scribe v3 Live, and optimized for clinical language. It is now available on the ElevenAgents platform.

General-purpose speech models often struggle with medical conversations because drug names and dosages rarely appear in everyday speech, and similarly sounding medications (such as sitagliptin and citalopram) can be easily misheard. Scribe v3 Live Medical aims to address this by ensuring voice agents accurately capture critical clinical information during patient calls for scheduling, eligibility checks, or claim follow-ups.

According to ElevenLabs' internal testing, the model achieved a Word Error Rate (WER) of 12.8% on PriMock57, a public dataset of 57 mock primary care consultations, marking the lowest WER among the medical models tested and 17% fewer errors than the next closest competitor. On the Fareez OSCE dataset comprising 272 patient interviews, it recorded a semantic WER of just 0.6%. Compared to the general Scribe v3 Live model, it reduced semantic errors by 4.7% on the Eka Medical ASR dataset.

To handle complex medical scenarios, the model introduces key features: support for loading up to 10,000 custom terms, allowing systems to precisely recognize specific hospital, clinic, or brand names; optimized 'turn detection' that identifies natural pauses when patients read out ID numbers or dates of birth, preventing agent interruptions; and support for switching between more than 90 languages in real time to accommodate multilingual patients.

Regarding compliance, Scribe v3 Live Medical offers HIPAA-eligible options for enterprise customers, requiring a Business Associate Agreement (BAA) and Zero Retention Mode to ensure Protected Health Information (PHI) is not stored. Idin Jafari, Director of Digital Strategy & Innovation at Pfizer, noted that combined with Eleven v4 Turbo, the model significantly improves accuracy in recognizing brand and generic drug names, delivering a realistic interaction experience.

Fontes:ElevenLabs

Pesquisa

AWS Announces General Availability of AI Data Preparation Capabilities in SageMaker to Accelerate Model Development

Tópico · AWS Spend LimitsMaterial
2026-11-18 03:40 GMT+8

Amazon Web Services (AWS) announced on November 17, 2026, that new "AI Data Preparation" capabilities are now Generally Available in Amazon SageMaker AI. These features are designed to help teams build high-quality training and evaluation datasets more efficiently, addressing the common challenge where data wrangling becomes the slowest part of generative AI projects.

The new capabilities include five core modules: auto-labeling, human review, synthetic data generation, data quality evaluation, and data transformation. For auto-labeling, the system uses a "jury" of multiple models to classify data or perform entity recognition, using their agreement level as a confidence score. Low-confidence samples can be routed to a human review queue in Amazon SageMaker Studio. The synthetic data generation feature allows users to create task-specific examples grounded in specifications and seed data, producing a code-based pipeline that can be reviewed, modified, and reused. Data quality evaluation utilizes built-in metrics and LLM-as-a-judge techniques to detect duplicates, label noise, and answer leakage.

These functions are accessible via Python SDK APIs or through an AWS AI-ML agent skill, supporting workflows in JupyterLab, Kiro, or local IDEs. AWS emphasizes that all generated data and code remain in user-controlled Amazon S3 buckets, with execution roles managed by user IAM permissions, ensuring data privacy and control.

Yash Sheth, Director of Engineering at Splunk (a Cisco company), stated that the capabilities helped his team rapidly generate statistically representative synthetic datasets for their Agent Observability product while preserving underlying patterns, thereby reducing the time data scientists spend on data wrangling. The features are currently available in the US East (N. Virginia) and US West (Oregon) regions.

Fontes:Amazon Web Services

Pesquisa

AWS Team Self-Test: Three Strategies Reduce Token Consumption in AgentCore Large JSON Workloads by Up to 71%

Material
2026-11-18 06:44 GMT+8

When production agents possess dozens of tools and return full structured data with each invocation, the context window fills within a few conversation turns. The AWS engineering team encountered this issue while building a crew scheduling optimization agent for the airline operations platform Flightscape: more than 40 tools, megabyte-scale JSON inputs, average request latencies exceeding 15 seconds, and some single tool responses exceeding 80,000 tokens. Every tool call injects the full payload into the model context, compounding pressure across multi-turn conversations, driving up costs, and degrading response quality.

Their first finding was that the bottleneck initially lay in the persistence layer. In the original design, heavy payloads such as optimization request JSON, response JSON, and validation reports were stored directly in session state, requiring serialization of these objects every turn. The team moved heavy payloads to Amazon S3, while AgentCore memory persisted only lightweight references (S3 key names). References are generated via content addressing—normalizing the JSON serialization and taking the SHA-256 digest—making writes idempotent: identical payloads map to the same key, allowing unchanged requests in subsequent turns to reuse existing objects without re-uploading.

After resolving the persistence layer, the bottleneck shifted to tool responses. Even with a lightweight session state, complete lists returned by tools were still injected verbatim into the model context. The team introduced on-demand paging: when tool results exceed a sampling threshold, the complete list is stored in S3, and the model receives only a compact summary (total count, initial samples, result handle). When more detail is needed, the model calls the paging tool to retrieve the required slice using offset and limit.

However, not all oversized outputs are lists. Some tools return deeply nested dictionaries containing numerous scalar values and nested structures; retaining them verbatim is too costly, but discarding them entirely deprives the model of reasoning cues. For dictionary outputs, the team designed descriptor-based compression: list fields retain only type, count, and the first five samples; nested dictionaries retain only key counts; scalar values are retained inline when useful. The full payload remains in S3, retrievable by the model on demand via the handle.

The three strategies are composable: Strategy 1 runs at the start and end of each turn, while Strategies 2 and 3 run each time a tool returns a result. In benchmark tests over five conversation turns, the team reported total token usage dropping from approximately 289,500 to approximately 82,900 (a reduction of about 71%), and average trace latency decreasing from 15.4 seconds to 3.1 seconds (a reduction of nearly 80%). Across 100 sample questions, total tokens fell from 23.6 million to 8.5 million (64%), and average latency dropped from 7,930 milliseconds to 2,461 milliseconds (69%). For the heaviest single query (listing valid crews across all validation categories), input tokens dropped from approximately 32,000 to approximately 8,000, and latency decreased from approximately 165 seconds to 9.9 seconds.

Fontes:Amazon Web Services

Pesquisa

LoanPro Multi-Agent Loan Servicing Assistant: How Compositional Routing and the MCP Gateway Isolate Tool Access in Regulated Credit Scenarios

Tópico · LoanPro Agentic AI AssistantVisão rápida
2026-11-18 01:01 GMT+8

When answering borrower calls, loan servicing agents must synthesize approximately 1,400 account fields from eight API endpoints—including payment history, loan terms, open tickets, and available workflow actions—within dozens of seconds to provide meaningful responses. The multi-agent assistant developed by LoanPro in partnership with AWS Generative AI Innovation Center and partner AllCloud is designed around a core decision: rather than having a single large model handle all tasks, it uses a compositional agent to route requests to specialized agents, then isolates tool access through an MCP gateway.

The architecture consists of three layers of collaboration. A composer agent receives requests and determines which specialist agent should handle them: one focused on data retrieval and account summarization, and another on action selection and workflow recommendations. Specialist agents invoke LoanPro’s internal APIs via the AgentCore Gateway using the Model Context Protocol (MCP); the gateway locates tools in a catalog, handles authentication, and manages connections. The rationale for this separation over a monolithic model is evaluability: each specialist agent can be independently tested, tuned, and scaled without requiring a full system redeployment when modifying a single function.

Security boundaries are a critical constraint in regulated scenarios. The AgentCore Gateway isolates LoanPro’s APIs behind a managed boundary, while AWS Lambda functions validate caller signatures against consistency with the single sign-on (SSO) service, and role-based permissions are verified with every request. All AI operations are written to audit logs and attributed to named individuals. The team chose a gateway plus Lambda approach over direct integration for three reasons: authentication is centralized at a managed entry point, internal APIs are isolated behind a managed boundary, and adding new tools does not require redeploying the agents.

Regarding the evaluation process, the team established a baseline using 150 accounts annotated by subject-matter experts to measure agent performance in data retrieval, summary generation, and action selection, setting and meeting a 99% accuracy target. This framework extends into the production environment as a continuous measurement mechanism.

Vendor-reported results include: among more than 800 servicing agents at early partner lenders, agents previously spent 18%–20% (1–2.5 minutes across 5–7 clicks) per call orienting to accounts; the assistant compresses this into a single click. Account summaries are generated in under 15 seconds, compared to the previous 3–5 minutes required for manual navigation. Internal ticket volumes have decreased by nearly 8%. These data points come from AllCloud case studies and LoanPro internal benchmarks and remain unverified by independent third parties.

The applicable conditions for this architecture are: internal APIs already exist and can be toolized, compliance requirements demand audit trails and permission isolation, and multiple specialist tasks require independent evolution. It does not solve problems inherent to model capabilities—if the underlying model cannot accurately synthesize 1,400 fields, neither routing nor the gateway will change that fact.

Fontes:Amazon Web Services

Pesquisa

Microsoft Unveils In-House AI Coding Models, Claiming 96% of Claude Opus 5.5 Performance at 5% of Cost

Tópico · Microsoft MAI Code 2Material
2026-11-18 06:42 GMT+8

At its Ignite 2026 keynote on November 17, Microsoft announced two new frontier AI coding models: MAI Code 2 and MAI Code 2 Flash. According to ZDNET, the company claims that MAI Code 2 delivers 96% of the performance of Anthropic's Claude Opus 5.5 while costing only 5% as much per API usage. A second model, MAI Code 2 Flash, is positioned to compete with lower-end offerings like GPT-6 Luna and Claude Haiku 5.5, also emphasizing significant cost reductions.

This move signals a strategic shift in Microsoft's AI infrastructure approach. Historically, Microsoft has maintained deep ties with OpenAI through multi-billion dollar investments and integrated external frontier models into products like Copilot. By benchmarking its internally built models against industry leaders and offering them as optional choices for GitHub Copilot, Microsoft 365, and Fabric, Microsoft aims to decrease its dependence on external providers. If these claims hold, Microsoft could save hundreds of millions of dollars annually in inference costs.

It is important to note that all performance and cost metrics are self-reported by Microsoft and have not been independently verified. Microsoft clarified that these models are designed for "everyday" coding tasks rather than the most complex agentic or research workloads. The projected savings rely on hypothetical scenarios of high-volume internal usage, meaning the actual financial impact will depend on the rate of adoption across Microsoft's vast software ecosystem.

Fontes:ZDNET

Pesquisa

Netflix Agrees to Higher Residuals and Pension Contributions for German Voice Actors, But AI Training Dispute Remains Unresolved

Visão rápida
2026-11-17 23:59 GMT+8

According to heise online, streaming giant Netflix and the German actors' union Bundesverband Schauspiel (BFFS) have agreed on new common compensation rules. Under the agreement, German voice actors participating in successful Netflix productions will receive higher residual payments (Folgevergütung) starting January 1, 2027. Additionally, Netflix will reimburse employer contributions to the Pensionskasse Rundfunk (Broadcasting Pension Fund) for German-language dubbing productions it commissions. Netflix states it is the first streaming service to support retirement provisions for dubbing work in this manner.

While the official announcement did not disclose specific amounts, industry media DWDL reports that the annual pot for residuals will increase from €475,000 to €750,000. Approximately 82.5% of this sum will be allocated to series, with the remainder going to films. The funds are distributed among the 15 most successful series and films of each year. For the most successful series, participating voice actors can share around €71,000.

However, this new agreement does not address the core dispute that has divided the German dubbing industry since the beginning of the year: AI training rights. BFFS and Netflix merely referred to protection rules agreed upon in 2025, which remain in effect: digital clones and synthetic voices may only be used with the separate and explicit consent of the actors. The announcement says nothing about the controversial use of voice recordings for training AI models, which Netflix requires via its "Assignment of Rights" (AOR) agreement according to the Association of German Voice Actors (VDS).

The VDS claims that voice actors who refuse to sign the AOR agreement are no longer commissioned. Consequently, the pay raise primarily benefits actors who already accept Netflix's terms, while those opposing AI training rights continue to be excluded from these improved conditions. Currently, the VDS has filed a complaint against Netflix with the German Federal Cartel Office, parallel privacy complaints are underway with the Berlin Data Protection Commissioner, and voice actor Matti Klemm has announced plans to sue Netflix.

Fontes:heise online

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

At the architectural boundary, neither Slurm nor the Lua plugin communicates directly with the QPU. Third-party access middleware sits between the HPC system and the QPU, providing a unified interface to various quantum hardware APIs. Slurm treats the QPU as a single schedulable unit, deciding its availability, holder, duration, and source nodes; calibration, error correction, and device control remain within the QPU's own software runtime. This design applies to both current Noisy Intermediate-Scale Quantum (NISQ) devices and future fault-tolerant systems.

Two open-source implementations already exist within the community. The QRMI plugin reserves QPU access using the user's own credentials before the job runs, completing this process in a separate thread outside the scheduler's critical path, and releases it after the job ends; existing applications run without modification. The QDMI plugin, conversely, does not reserve the device but only admits the job; each job submits its work via QDMI, where the provider's queue schedules it on the device. From the user's perspective, both integrations present a consistent interface: job scripts request the QPU via SBATCH --resources, and Slurm keeps the job pending until the plugin reports the device is available.

Notable limitations include the fact that the article is published by the NVIDIA Technical Blog, constituting vendor-explanatory material regarding its own product features. The performance improvements cited (such as eliminating idle nodes) are logical deductions based on architectural design and lack actual pre- and post-deployment comparative data. Furthermore, Slurm 26.11 itself remains in an early adoption phase, and the reliability of community plugins in production environments awaits verification.

It is crucial to emphasize that all figures cited above—the 14% error reduction, 27 minutes, and over 100 fixes—are self-reported results from Salesforce, with no public code, third-party replication, or independent benchmark testing available. Each agent investigation operates within a fixed testing budget, and when evidence is inconclusive, the case is handed off to a data scientist. These constraints are part of the design, meaning the scheme's effectiveness is highly dependent on its specific platform scale and metric characteristics; these numbers cannot be directly applied when migrating to other environments.

Salesforce states it is extending the agent beyond model tuning: monitoring the health of data pipelines behind forecasts, automating the onboarding process for new forecasts, and incorporating textual context (such as product launch plans and migration announcements) to adjust forecasts before changes occur. The team believes the goal is not merely to recover forecasts after they miss, but to anticipate them proactively.

It should be noted that the above figures are self-test results from the AWS/Flightscape team on specific workloads: 23 MB of airline crew scheduling JSON, more than 40 tools, LangGraph orchestration, and Amazon Bedrock model inference. The article does not disclose the specific model versions and inference parameters of the baseline configuration, nor does it quantify the impact of increased pagination tool invocations on end-to-end latency. For agents with fewer tools, different payload shapes, or different model inference configurations, actual benefits require individual verification. The core transferable insight is that context pressure from large JSON loads is not resolved in one go; new bottlenecks are exposed with each layer of optimization, requiring the sequential stacking of corresponding compression or offloading strategies.