LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-10-23129 publicaciones

Actualizar

Open-source AI inference framework llm-d releases v1.0: tests claim 5-10x cost reduction vs commercial APIs under specific loads

Tema · llm-d推理框架Material
2026-10-23 16:00 GMT+8

The open-source project llm-d, founded by Red Hat, IBM, Google Cloud, CoreWeave, and NVIDIA, released version 1.0 on October 23, 2026, aiming to tackle resource contention in large-scale AI inference. Now governed by the Cloud Native Computing Foundation (CNCF), the project introduces techniques such as cache-aware routing, memory offloading, and separating input prompt processing from response generation to improve GPU utilization and reduce running costs.

According to eeNews Europe, in a benchmark test targeting GLM 5.2, an open-weight model with approximately 753 billion parameters, researchers used llm-d on NVIDIA GPUs to achieve peak throughput of more than 6.6 million output tokens per minute while supporting up to 3,000 concurrent coding agents without preemptions. Based on current cloud rental rates, the researchers estimated that inference costs per token were five to ten times lower than comparable commercial API pricing for the tested workloads. Another study at IBM Research reported that using llm-d for quantum error correction hardware investigation resulted in inference costs 11.5 times lower than an entirely proprietary configuration.

It is important to note that these cost savings figures come from specific workloads and configurations rather than general performance guarantees. Version 1.0 also introduces multi-tenant quality of service, session-aware routing, and support for updates without planned downtime, with over 500 contributors from more than 25 organizations currently involved in development.

Fuentes:eeNews Europe

Investigación

Disprz Co-founder Kuljeet Chadha: How Voice AI Scales One-on-One Corporate Coaching

Tema · Disprz企业学习平台Resumen rápido
2026-10-23 21:30 GMT+8

This resource is recommended for readers interested in the application of AI in B2B services and HR technology. It records a talk at the ElevenLabs Summit Bengaluru by Kuljeet Chadha, co-founder of the enterprise learning platform Disprz.

Disprz serves around 500 customers globally with millions of users. In the talk, Chadha explains why the company is moving from traditional passive online courses to a 'voice-first' coaching model. He highlights a key pain point: most corporate training budgets go into formal courses with completion rates under 15%, while actual learning often happens on the job and through coaching.

The presentation focuses on how Voice AI enables scalable 'one-on-one' coaching. Disprz's approach combines each customer's knowledge base, culture, and roles with live business signals, such as regulation changes or dips in sales performance. AI agents identify learning needs, draft modules, obtain human approval, and deliver content directly within tools where employees already work, like WhatsApp or CRM systems.

The latter part of the video showcases a specific example: one of the world's largest logistics companies used Disprz's age-verified delivery simulation to onboard around 100,000 new employees, now running in 13 European languages. According to Disprz, this customer cut its onboarding budget by 50% and reduced onboarding time from 4.84 to 4.2 days.

It is important to note that these efficiency figures are self-reported by Disprz and have not been independently audited or verified with public methodology details. This video serves as a quick reference for understanding potential application models of Voice AI in corporate training, rather than proof of performance benchmarks.

Fuentes:youtube.com

Investigación

oneDNN 3.14 Released: First Integration of AMD Zen Optimizations and ACE Support

Material
2026-10-23 22:12 GMT+8

The UXL Foundation released oneDNN 3.14 on October 23, 2026. As a library providing optimized deep learning operations for frameworks like PyTorch, TensorFlow, and OpenVINO, oneDNN was historically driven primarily by Intel. A significant change in this update is that CPU optimizations for the AMD Zen architecture are now landing in the mainline oneDNN repository, signaling deeper software-level collaboration between the two chip giants.

Regarding AMD Zen optimizations, version 3.14 improves the performance of matrix multiplication (MATMUL) and inner-product primitives. These optimizations currently require users to manually enable the ONEDNN_X64_USE_ZEN=ON build option and may involve fetching the ZenDNN codebase. Although not yet enabled by default, this breaks the previous pattern where AMD relied on a separate fork (ZenDNN), paving the way for unified cross-platform optimization.

Additionally, the release includes initial support for the "AI Compute Extensions" (ACE) instruction set. This is a hardware feature intended to be supported by future Intel and AMD processors to accelerate specific AI workloads. Performance improvements were also made for upcoming Intel Xeon Diamond Rapids processors (supporting AVX10.2 and AMX) and Intel GPU D-Series (Crescent Island).

Fuentes:phoronix.com

Investigación

Paolo Sabatinelli, Chief Product Officer at Immobiliare.it: Innovation in Agentic Networks Lies in Orchestration, Not Individual Agents

Material
2026-10-23 20:00 GMT+8

The deployment path of Immobiliare.it, Italy’s largest real estate platform, offers a concrete reference: validate in a narrow scenario first, then expand to a network. The company launched a voice-enabled conversational agent for each property listing, responsible for answering calls around the clock, responding to questions, collecting lead information, and summarizing interactions for real estate agents. More than 70% of sellers opted to enable this feature, and early results validated the direction.

The core judgment of Chief Product Officer Paolo Sabatinelli is: 'A lot of the innovation is in the orchestration: deciding how agents should work together, how to keep their behavior predictable, and how to improve the experience as a whole.' The actual architecture he describes is the sub-agent pattern—each sub-agent handles only one narrow task, coordinated behind the scenes so the user always perceives a single coherent conversation rather than a patchwork of multiple independent systems.

Starting from the first inbound lead agent, the network gradually expanded to include private seller inbound agents, listing quality outbound agents, cross-channel (voice and text) customer service, and conversational search. Sabatinelli emphasized that API-driven design allows each capability to integrate with external platforms or databases, reducing integration costs for expansion. The team learns together within a single platform, sharing lessons learned across various projects.

It should be noted that this is a customer story published by ElevenLabs. All performance data (a 3.3x increase in lead qualification rate, a 74% increase in the likelihood of users sharing phone numbers, and approximately 80% positive user ratings) are self-reported via vendor channels. Baseline definitions, measurement time windows, and sample methodologies have not been made public. While the observation that buyers are more willing to share personal information via voice than text is reasonable in direction, the specific magnitude cannot be independently verified.

Fuentes:ElevenLabs

Investigación

SemiAnalysis Ep. 040: ASML's Hybrid Bonding Patents Reveal New Dynamics in the AI Supply Chain

Tema · ASML混合键合设备Material
2026-10-23 18:30 GMT+8

Episode 040 of the SemiAnalysis podcast focuses on a potential major technological breakthrough by semiconductor equipment giant ASML. Host Jordan Nanos, along with guests Sherry Zhang and Howie, interprets recent patent filings regarding ASML's die-to-wafer (D2W) hybrid bonding tool. Although ASML has not officially announced the product, patent details reveal a design approach distinct from traditional methods: instead of bringing the die to the wafer, it uses a track-based architecture to bring the wafer to the die. This design aims to leverage ASML's accumulated expertise in alignment precision and particle control from lithography to solve yield bottlenecks in hybrid bonding.

The episode analyzes the profound impact of this technology on the AI hardware supply chain. Currently, Dutch company Besi dominates the D2W hybrid bonding equipment market, with business heavily reliant on advanced logic customers like TSMC and Intel. The SemiAnalysis team suggests that if ASML successfully launches a bonding tool based on the SuperTrak platform, it could gain an advantage in advanced logic packaging through higher station utilization and heterogeneous integration capabilities, posing a direct competitive threat to Besi. Additionally, the guests discuss why High Bandwidth Memory (HBM) is expected to wait until 2029 for large-scale D2W adoption, and how Chinese semiconductor manufacturers are using hybrid bonding to continue chip scaling despite lithography restrictions.

Fuentes:podcasters.spotify.com

Investigación

AGIBOT Open-Sources Embodied AI Dataset: Video Explains How 'Physical Reasoning' Helps Robots Decide Before Acting

Tema · 智元物理推理数据集Material
2026-10-23 18:53 GMT+8

AGIBOT has published a video on its official YouTube channel announcing the open-sourcing of the 'Physical Reasoning' dataset within AGIBOT WORLD 2026 Release 4. This dataset aims to help robots reason more effectively before executing actions.

According to the video description, the dataset captures different actions taken for the same task under different states. It records how robots decide what to do as object quantity, occlusion relationships, spatial occupancy, and task progress change, linking these decisions to real-robot execution.

For developers and observers working in embodied intelligence, this video offers a direct entry point to understand how AGIBOT addresses the connection between perception and decision-making in robotics through data-driven methods.

Fuentes:youtube.com

Investigación

Insilico Medicine Releases Video Introducing Its AI Clinical Trial Prediction System inClinico Agent

Tema · 英矽智能Resumen rápido
2026-10-23 20:41 GMT+8

Insilico Medicine released a video titled "InClinico Agent is Here!" on its official YouTube channel on October 23, 2026. According to the publisher's description, the video introduces the company's newly launched inClinico Agent. This is an AI agentic system designed to accelerate clinical trial intelligence and predict trial outcomes.

For readers focused on specific applications of AI in drug discovery and clinical trials, this video provides a direct official statement regarding the product's functional positioning. While current evidence consists only of the publisher's brief description and does not include independently verified performance data or detailed technical mechanisms, it confirms the company's progress in this specific direction. Professionals in related fields are advised to watch the video for more intuitive product demonstrations and background information.

Fuentes:youtube.com

Investigación

Western Digital Forecasts Up to 72% Revenue Growth for Fiscal Q2 Amid AI Storage Demand

Material
2026-10-23 20:44 GMT+8

US data storage company Western Digital Corp (NASDAQ: WDC) issued financial guidance on October 23, 2026, forecasting that revenue for its fiscal second quarter will increase by 66% to 72% year-on-year, reaching approximately $5.1 billion (plus or minus $100 million). This projection follows a strong first quarter, during which the company's sales rose 55% year-on-year to $4.38 billion.

Management noted that demand for data storage continues to strengthen as cloud service providers and large-scale data center operators expand their artificial intelligence infrastructure. Consequently, Western Digital expects its adjusted gross margin for the second quarter to reach 60% to 61%, with adjusted earnings per share forecast at $5.80 (plus or minus 20 cents).

In the first quarter (ended October 2), Western Digital's adjusted earnings per share surged 161% year-on-year to $4.64, while operating income increased 158% to $2.05 billion. Gross margin expanded to 57.8% from 43.5% a year earlier. Cash generated from operations more than doubled to $1.60 billion, and free cash flow rose to $1.38 billion from $599 million.

Reflecting improved profitability, Western Digital announced a 17% increase in its quarterly cash dividend to $0.175 per share, scheduled for payment on December 17. Chief Executive Officer Irving Tan stated that robust customer demand and disciplined execution supported these results as the company progresses through technology transitions to meet growing storage requirements.

Fuentes:Proactiveinvestors UK

Investigación

Hugging Face Launches Tool to Check AWS Trainium Model Compatibility

Resumen rápido
2026-10-23 08:00 GMT+8

Hugging Face launched a Space application named "torch-neuronx model support" on October 23, 2026, designed to answer the developer question: "Does my model run on AWS Trainium?" The tool lists the actual running status of the most downloaded and trending Transformers models on the Hub when executed on Trainium chips.

For each tested model, the page displays whether it runs unmodified, the number of Trainium cores required, the total wall time including compilation, and peak device memory. Users can also view the specific Python script used for testing to reproduce the results in their own environments. Currently, tests cover text-generation and image-text-to-text tasks, with plans to expand to diffusers and other architectures later.

It is important to note that a "PASS" result only indicates that the model successfully loads and generates fluent English text on the specified Trainium instance (compared against CPU output for single-core runs, or judged by a small model for sharded runs). It does not guarantee throughput meets production requirements, nor does it imply compatibility across all tasks or longer context windows.

Fuentes:huggingface.co

Investigación

JFrog Expands AI Security: Local Shadow AI Scanning and BYOK Support

Tema · JFrog AI安全套件Resumen rápido
2026-10-23 14:32 GMT+8

Software supply chain security company JFrog has announced an expansion of its AI Security suite, extending shadow AI detection from traditional repositories and builds to developers' local computers. The new JFrog Agent Plugin, deployed via Mobile Device Management (MDM) or agent settings on workstations, is designed to discover unauthorized AI Skills, plugins, and MCP servers.

According to JFrog's '2026 Software Supply Chain Report', the company previously identified 969 malicious Agent skills and 495 malicious AI models on platforms like Hugging Face. This update aims to close the regulatory gap before these assets enter production environments. The scanning process executes deterministically locally, sending only metadata and risk assessments back to the JFrog platform, thereby balancing security with privacy.

Another significant update is the introduction of a 'Bring Your Own LLM Key' (BYOK) mechanism for semantic AI asset scanning. This feature allows customers to use their own subscribed model endpoints, such as Anthropic's Claude Opus 4.8, OpenAI's GPT-5.6 Sol, or Google's Gemini 3.1 Pro, for analysis without routing prompts and data through a JFrog gateway. This change makes security reviews of complex AI logic more flexible and compliant with data regulations.

Fuentes:Developer

Investigación

US FCC Bans First Two Devices Under New Chip-Level Rule Over Huawei Subsidiary AI Face Detection

Tema · FCC华为芯片禁令Material
2026-10-23 20:15 GMT+8

The US Federal Communications Commission (FCC) has recently enforced a new regulatory rule that allows it to ban devices based on their internal chips rather than final product brands or models. This move aims to close loopholes where products are rebranded to circumvent restrictions on components from Chinese manufacturers. According to Tom's Hardware, the FCC has used this new rule for the first time to prohibit the sale and marketing of two devices in the US: a dashboard camera and a wireless video transmitter.

These devices were banned because they contain the HiSilicon Smart Vision System-on-Chip (SoC), which features built-in face detection algorithms. HiSilicon, a subsidiary of Huawei, is on the FCC’s Covered List of entities deemed to pose a risk to US national security. The report identifies the specific chip as the Hi3516CV610 UHD Smart Vision, designed for vehicle dashcams and wireless CCTV systems.

The SoC includes a Neural Processing Unit (NPU) for AI capabilities. The FCC expressed concern over these logic-bearing “smart” components produced by entities on the Covered List. HiSilicon’s datasheet indicates that the Intelligent Video Engine (IVE) uses the NPU to run a built-in multimodal large model and transformer feature acceleration, enabling the understanding of object relationships across video frames to identify pets, faces, human figures, and parcels. While the vendor markets this as helping distinguish between a pet escaping and a parcel theft, the FCC believes the AI chip could be used to survey US citizens and log faces for foreign surveillance purposes.

Fuentes:Tom's Hardware

Investigación

US Biotech Firm Claims First Human Treatment with AI-Designed Phage

Material
2026-10-23 10:20 GMT+8

According to a report by TechNews, US biotechnology company Radical Numerics recently announced that a bacteriophage designed by its artificial intelligence model, Omnii-1, has been used in a human patient for the first time. The treatment was administered through a Compassionate Use mechanism to a patient suffering from a bacterial infection uncontrollable by antibiotics.

Radical Numerics stated that Omnii-1 is a multimodal genomic model capable of processing DNA, RNA, and protein information to generate phage-related sequences and predict their pairing relationships with bacteria. Before clinical use, the company tested the designs against antibiotic-resistant strains isolated from the patient, reporting that a cocktail including the AI-designed variants cleared the target bacteria in vitro.

Bacteriophages are viruses that specifically infect bacteria and are viewed as potential tools against antimicrobial resistance. Compared to broad-spectrum antibiotics, phages offer the possibility of targeting specific strains more precisely. However, the company did not disclose patient details or treatment outcomes, citing privacy concerns. Radical Numerics also cautioned that a single case is insufficient to prove the safety or efficacy of such therapies, noting that technical reports have not yet undergone peer review or been formally published in academic journals.

Fuentes:TechNews 科技新報

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión