読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-18133 投稿

更新

Independent test finds PolyAI's Dialog-RSN-1 fastest voice agent

重大
2026-11-16

UK-based AI voice company PolyAI announced that its dialog model, Dialog-RSN-1, was found to be the fastest responding voice agent in tests conducted by independent lab Veris AI using the VAmoS Pro Bench.

Veris AI simulated electric utility billing inquiries by sending 300 calls from various environments (quiet rooms, streets, cafes) and emotional states (calm or angry) to multiple AI agents. The results showed that half of Dialog-RSN-1's replies started within 1.24 seconds after the customer stopped talking, whereas most other agents took more than two seconds.

PolyAI stated that Dialog-RSN-1 uses an audio-native architecture where a single model handles turn-taking, hearing, and replying together, eliminating the need for a separate speech recognizer. This design removes steps that add delay and allows the model to maintain awareness of the full conversation context during reasoning.

These results were published by third-party organization Veris AI, with detailed data available on their VAmoS Pro leaderboard.

ソース:PolyAI

リサーチ

ScaleOps Founder Yodar Shafrir: The Idle-Spike Pattern of AI Agents Renders Static Resource Planning Obsolete

トピック · ScaleOps重大
2026-11-18 04:35 GMT+8

Yodar Shafrir, founder and chief executive officer of ScaleOps Inc., offered a specific assessment regarding AI agent infrastructure during an interview with theCUBE Research at KubeCon + CloudNativeCon NA (United States). He argued that the behavioral patterns of agent workloads differ fundamentally from those of traditional microservices, creating dual risks of waste and failure for static resource configurations.

Shafrir described this agent behavior pattern as follows: an agent waits for input from a human or another agent, calls upon a large language model, and then executes its task in a burst. Teams must provision resources for the highest peak, while the agent continuously holds memory during the entire waiting period. When resources are insufficient, crashed steps must be rerun, consuming tokens again. He stated that ScaleOps has observed agents crashing due to resource shortages 80% more frequently than traditional microservices.

This mechanism explains why traditional auto-scaling strategies fail in agent scenarios. The idle-to-spike switching of agents occurs without warning, meaning static resource configurations either result in waste or face the risk of failure. Shafrir believes that humans and all traditional scaling capabilities are largely irrelevant at this current stage because they are too static and require manual configuration; when changes are this dynamic, they no longer function in the same way.

ScaleOps's response is real-time resource decision-making: adjusting allocations, replica counts, and node placement for every application, agent, and Model Context Protocol server. The company claims this can reduce cloud costs by 70% to 80%. Additionally, ScaleOps allows other workloads to utilize resources while agents are idle and employs memory swapping so that agents receive memory only when needed. The company states this improves the cost efficiency of running agents by 90% without requiring a move away from Kubernetes.

Customers such as Adobe Inc. and Salesforce Inc. initially ran the platform in read-only mode to build trust. Shafrir stated that ScaleOps learns the historical patterns of each application, understanding their connections, communication partners, and business states.

It should be noted that this interview was sponsored by ScaleOps, and all performance figures are self-reported by the vendor, lacking independent benchmark verification. Shafrir's statement regarding the 80% higher agent crash rate stems from ScaleOps' own observations; the specific measurement methods and sample scope were not disclosed in the interview.

ソース:SiliconANGLE

リサーチ

Fashion E-commerce Startup Daydream Launches AI Shopping Agent That Can Complete Purchases

トピック · Daydreamクイックテイク
2026-11-18 00:32 GMT+8

Fashion e-commerce startup Daydream launched "Daydream Agent" on Tuesday, an AI tool designed to act as a personal stylist. The agent helps users plan outfits, virtually try on clothes, and complete purchases.

Unlike earlier chatbots that only offered suggestions, the new agent remembers user style preferences, fit requirements, and budgets. Shoppers can upload photos to virtually try on saved items and proceed through a one-click checkout process. The agent also monitors sales and brand drops, allows users to save items to a persistent "clothing rack" for comparison, and integrates with calendars to plan outfits around existing wardrobes.

Technically, Daydream states the agent combines fine-tuned versions of OpenAI, Gemini, and Claude models with its own proprietary models, pulling real-time pricing and availability from a catalog of over 10,000 brands. To address the subjective nature of fashion, the company hired international stylist Francesca Salih as its first stylist in residence this month. Salih provides content to shape the agent's outfit-pulling logic, maintaining a human-in-the-loop approach.

Founded by former Stitch Fix COO Julie Bornstein, Daydream has raised $50 million from investors including Forerunner Ventures. The company reports that more than 2 million shoppers have used its platform since its June 2025 launch.

ソース:TechCrunch

リサーチ

Janus Henderson Research Analyst Tim McCarty: The Three-Stage Value Path of AI Surgical Robots, from Physician Performance Reports to Capacity Bottlenecks

トピック · AI手术机器人价值路径重大
2026-11-17 17:55 GMT+8

Tim McCarty, a research analyst at Janus Henderson, characterizes the role of AI in healthcare as "under the radar and omnipresent." He argues that while most investors are only familiar with AI's application in drug discovery, AI is permeating the entire healthcare value chain, providing an additional growth driver for the healthcare sector, which also possesses defensive characteristics. In this interview published on November 17, 2026, he elaborates on the three-stage value path of AI in surgical robots.

The first stage, which has already been realized, is current: AI acts as a "report card" for surgeons. McCarty gives the example of a gallbladder surgeon who may take 20 minutes longer than best-in-class standards and use two extra instruments. AI can decompose the surgery into specific steps, pointing out, "You are deficient in step six," and matching training modules on practice robots. He uses the analogy of college basketball game film analysis—watching game footage, identifying deficiencies, and doing targeted drills.

The second stage is what McCarty calls the "not-too-distant future": AI eliminates medical errors and democratizes surgery. His core argument is information asymmetry—the healthcare system is rife with it, with bottom-quartile doctors responsible for a disproportionate amount of medical errors, which cost the system tens of billions of dollars annually. However, patients cannot judge a doctor's skill level before surgery, as "no one walks in saying I'm a bottom-quartile doctor." McCarty believes this asymmetry will be broken when AI can objectively evaluate every surgeon's performance.

The third stage is a long-term vision, which McCarty refers to as the "Waymo of healthcare." He points out that capacity is one of the major problems facing the current healthcare system: one surgeon per one patient. He envisions a world where, as robots autonomously complete more surgical steps, a single surgeon can "drive" multiple surgeries simultaneously, monitoring several parallel operations like a pilot, thereby alleviating supply-demand mismatches.

Sarah de Lagarde, Janus Henderson's Brighter Future Strategist, provides firsthand calibration for this narrative. She wears a bionic arm but admits its practical advantages in daily life are limited—"My biological left hand is much stronger than this arm." She notes that engineers tend to solve problems that are not truly needed in patients' daily lives, and distinguishing between "over-marketing" and "genuine benefits to patients" is precisely the job of portfolio managers. The only function she considers irreplaceable for the bionic arm is the ability to modulate grip strength at will via a smartphone app.

McCarty also mentioned AI's application in cardiovascular diagnostics. He focuses on severe aortic stenosis—a valvular disease with a 25% mortality rate if untreated—and notes that approximately 500 million people worldwide suffer from cardiovascular disease. Currently, minimally invasive replacement surgery is available at 850 centers in the United States. The opportunity for AI lies in improving diagnosis and treatment rates from primary care physicians to cardiologists and across various levels of the healthcare system.

ソース:US Advisor

リサーチ

Microsoft Releases Decision-1.1 Built on Gemma 4 and MAI Models; Foundry Adds Enterprise Knowledge Graph and Continuous Improvement Loops

トピック · Decision-1.1重大
2026-11-18 04:05 GMT+8

US technology company Microsoft announced on November 17, 2026, the rollout of new generative artificial intelligence models and major updates to its unified enterprise platform, Foundry. The centerpiece of this release is the generally available Decision-1.1 model, a generative AI model designed for fast, cost-effective decision-making in software at scale. Microsoft states that the model is built on Google LLC's open-source Gemma 4 26B A4 base and includes multimodal support for analyzing image instructions. Unlike large language models that rely on "thinking it through," Decision-1.1 is designed to quickly reach answers within constrained definitions and provide calibrated confidence scores, allowing applications to defer when uncertainty is high.

In addition to the decision model, Microsoft expanded its first-party MAI (Microsoft AI) model family. This includes MAI-Code-2 for demanding agentic coding work, MAI-Code-2-Flash for high-volume everyday coding, and the MAI-Image-2.7 series for sharper editing and consistent image generation. In cybersecurity, the company introduced MAI-Cyber-2-Flash, a purpose-built model available in preview for Codename MDASH, its agentic code security system. These models are now integrated into the Foundry platform for enterprise customers.

Updates to Foundry focus on enhancing context understanding and continuous optimization for enterprise AI applications. Foundry IQ now provides an enterprise knowledge graph out of the box, delivering a full estate of unstructured and application data to provide rich context for building applications. Furthermore, Microsoft introduced "learning loops" that allow autonomous agents to establish baselines for quality, safety, latency, and cost, enabling continuous behavioral improvements through internal evaluations. This process combines human annotations of agent traces (currently in preview) with optimization tests against real-world results, ensuring each improvement becomes the next baseline as teams measure results and test changes to enhance quality.

ソース:SiliconANGLE

リサーチ

Salesforce and Microsoft Announce 'Copilotforce' Partnership, Deeply Integrating CRM Workflows into Copilot and Teams

トピック · Copilotforce重大
2026-11-18 02:05 GMT+8

US software company Salesforce (NYSE: CRM) and US tech giant Microsoft (NASDAQ: MSFT) jointly announced a new partnership named 'Copilotforce' at the Microsoft Ignite conference on November 17, 2026. The collaboration aims to deeply embed Salesforce's customer management context and sales workflows into Microsoft's enterprise productivity ecosystem, including Microsoft Copilot, Microsoft Teams, and Microsoft Fabric.

According to the official announcement, key features include allowing sellers to access Salesforce customer data directly within Microsoft Copilot. Leveraging the Work IQ intelligence layer, which synthesizes insights from Teams meetings, Outlook emails, and SharePoint documents, users can prepare for meetings, analyze deal health, and review pipelines without leaving the interface. Additionally, a bidirectional zero-copy data connection has been established between Salesforce Data 360 and Microsoft Fabric, enabling organizations to query Fabric data natively within Data 360 or share Data 360 intelligence into Fabric without duplicating datasets.

Miguel Milano, President and COO of Salesforce, stated that this partnership combines the '#1 CRM' with the '#1 productivity surface,' transforming AI from a standalone tool into a core driver of productivity and decision-making. Judson Althoff, CEO of Microsoft Commercial Business, emphasized that combining Salesforce's data and sales skills with the work context provided by Microsoft Copilot allows customers to unlock actionable insights while maintaining enterprise-grade governance.

Accenture is serving as a design and implementation partner and 'customer zero,' deploying these integrations within its own global sales organization. While the Sales Cloud in Teams app is currently in pilot, specific general availability dates for the full Copilotforce suite have not yet been disclosed.

ソース:Salesforce

リサーチ

Google AI Studio Launches Full-Stack Android App Building with Automated Backend Provisioning

重大
2026-11-17

U.S. tech giant Google announced on November 17, 2026, that its development platform, Google AI Studio, now supports the creation of full-stack Android applications. This update enables developers to go from a single prompt to a production-ready Kotlin Android app complete with data persistence, user sign-in, and AI features, which can be published directly to their personal Google Play library.

The core of this upgrade lies in automated backend configuration. Previously, AI-generated apps were often limited to local session storage, making cross-device synchronization or secure cloud API calls difficult. Now, when a user's prompt requires these capabilities, the Google AI Studio Build agent automatically provisions a Cloud Firestore database, configures Firebase Authentication for Google Sign-In, and sets up Firebase AI Logic. The latter ensures that Gemini model API keys are not exposed in the client APK but are protected via server-side mechanisms.

In terms of developer experience, Google has introduced a faster, high-definition Android emulator with real-time audio output, allowing developers to test voice-driven applications directly in the browser. Additionally, a new 'one-click screenshot to agent' feature lets developers attach live UI screenshots from the emulator to their prompts, guiding Gemini to adjust interface layouts. All new features are rolling out in Google AI Studio today.

ソース:developers.googleblog.com

リサーチ

Fei-Fei Li says universities and public sector are 'priced out' of AI

トピック · 李飞飞AI公平获取论重大
2026-11-18 03:40 GMT+8

Fei-Fei Li, co-founder and CEO of World Labs and incoming Executive Vice President and Chief Scientist at Advanced Micro Devices (AMD), stated that universities and the public sector are being "priced out" of the AI revolution.

Speaking at Semafor's "Silicon Valley & The World" event on November 17, Li told host Andrew Edgecliffe-Johnson: "I don't think universities and public sectors are keeping up because we're priced out." She emphasized that the resources available to these institutions in terms of tokens and compute are "extremely limited."

Li, formerly a professor at Stanford University where she led the Stanford Institute for Human-Centered Artificial Intelligence (HAI), argued that lawmakers in Washington should allocate resources to the public sector to allow it to reap the benefits of AI given the high costs of usage. She also noted that universities are responsible for holding society accountable by researching and benchmarking AI.

"It would become one of the biggest failures of our time if we don't, as a society," Li said, "if we don't resource [the] public sector in a serious way."

ソース:semafor.com

リサーチ

California DMV Restricts Freeway Operations; Waymo Suspends Autonomous Freeway Rides Nationwide

重大
2026-11-18 03:49 GMT+8

The California Department of Motor Vehicles issued operational restrictions on Waymo on November 11, prohibiting its robotaxis from operating on California freeways. The decision followed several safety incidents, including three Waymo autonomous vehicles entering an active construction closure on U.S. Route 101 near Burlingame, and a separate incident involving California Highway Patrol vehicles on Highway 280 in San Francisco.

Although the restriction applies to California, Waymo confirmed it has complied and voluntarily extended the suspension of freeway operations to all 15 U.S. cities where it currently offers commercial services. Waymo spokesperson Dana Gallagher stated that this move aligns with the company's "safety-first principles" to evaluate performance around freeway construction zones. Waymo's ride-hailing service continues to operate safely on surface streets, serving over half a million trips each week.

Freeway access has been a central milestone for Waymo's expansion since late 2025, particularly in the San Francisco Bay Area, where it significantly reduces travel times to airports and Silicon Valley. This is not the first interruption; Waymo briefly halted highway operations in May 2026 due to similar navigation issues before returning in July after software updates. The California DMV described the current freeway operations as posing an "imminent hazard to public safety" and will maintain the restrictions until Waymo demonstrates effective corrective actions.

ソース:TechCrunch

リサーチ

US Software Stocks Slide on Tech Outage Fears; Salesforce Drops Over 8%

トピック · 美国软件股技术中断担忧重大
2026-11-18 02:54 GMT+8

Salesforce (NYSE: CRM) shares fell more than 8% on Tuesday, marking the stock's biggest one-day percentage decline since May 2024. The drop was not driven by company-specific earnings or product news, but by a broader sector-wide panic triggered by a series of recent technical disruptions.

As the US National Security Agency (NSA), the Federal Aviation Administration (FAA), and several major banks reported communication or system failures, investors began reassessing the reliability risk across the entire software industry. Although there is no publicly established link between these outages and Salesforce or other software companies' systems, market sentiment has translated these external infrastructure failures into a general concern for the SaaS sector.

Gil Luria, an analyst at D.A. Davidson, noted that the recent disruptions have increased investor concerns about software security, particularly the fear that bad actors could use AI tools to cause harm, leaving software companies vulnerable. This narrative of "AI-enabled security risks," while unproven in relation to these specific incidents, was sufficient to trigger valuation compression.

Other software names also came under pressure: ServiceNow fell 6.12%, Datadog declined 5%, Workday lost 4.15%, and Intuit dropped 5.19%. US Homeland Security Secretary Markwayne Mullin stated on Monday that the incidents were unrelated and did not indicate a broader threat to national security or the banking system, attempting to calm markets. However, for investors, the immediate focus remains on whether these disruptions will lead to greater regulatory scrutiny of software security and reliability, potentially altering the sector's risk premium.

ソース:invezz.com

リサーチ

Salesforce Launches Carter: An AI Shopper Agent That Replaces Keyword Search with Intent Understanding

重大
2026-11-17 22:00 GMT+8

US software company Salesforce officially launched Carter, an AI shopper agent integrated into its B2C Commerce platform, on November 17, 2026. The tool aims to replace traditional keyword search and category filtering with conversational interaction.

Carter's core capability lies in "intent understanding" rather than simple text matching. For example, when a user inputs "I need a few outfits for camping in cold weather," the system does not just search for products containing those words but recommends complementary items like hiking pants, puffy jackets, and sun shirts. This functionality is powered by Agentic Commerce Search, which handles ambiguous requests and provides visualized product comparisons and suggestions.

The agent consists of four pre-configured subagents responsible for product expertise, site FAQ answers, cart management, and order confirmation. This architecture allows users to complete the entire journey from discovery to payment within a single chat interface without navigating between pages. Merchants can integrate their own knowledge bases via Data 360 and customize follow-up questions and selling points in guided discovery.

Salesforce cited early adopters including personal care brand Conair, which deployed a customized version named Compass on Cuisinart.com to handle recommendations across more than 5,000 SKUs; and athletic retailer Hibbett, which went live with Hibbett AI in six weeks, reportedly handling 90% of core shopper journeys. SharkNinja has also unified its customer service and commerce interfaces under this AI agent.

Carter is now generally available as an add-on for B2C Commerce customers and supports all languages covered by Agentforce. While the vendor claims significant improvements in user experience, specific data on conversion rate uplift has not been publicly disclosed, and actual effectiveness will depend on each retailer's catalog structure and configuration strategy.

ソース:Salesforce

リサーチ

Datadog Engineer Khayyam Guliyev: When Restoring Historical State Becomes the Bottleneck, Handing Durable Correctness to the Storage Layer

重大
2026-11-17 08:00 GMT+8

Each new generation of Datadog’s metrics aggregation system had to support roughly 10× the metrics volume of the generation before it. In a blog post published recently, Engineering Manager Khayyam Guliyev detailed this evolutionary process. The central question remained constant: the aggregation service must remember which data has been incorporated into each time window, track message identifiers to avoid double-counting, and restore this state after a restart or failure to continue processing correctly.

The first-generation system used static partition ownership. Each partition group consisted of a leader and a follower pair; both performed aggregation work, but only the leader wrote results downstream. RocksDB retained several hours of aggregation buckets and deduplication identifiers locally. This design ran well at its initial scale, but the leader-follower model doubled compute load. When partitions grew too large, engineers had to split them manually—a documented split required approximately 1.5 hours of coordinated engineering effort, involved numerous steps, and introduced correctness risks.

The second generation separated partition ownership from specific instances. Kafka dynamically reassigned partitions, and each node continuously uploaded incremental SSTable changes to blob storage. When a new node took over, it restored the most recent time window first, loading older windows only when late-arriving data triggered them. This allowed the system to support roughly 10× higher data volumes and introduced autoscaling. However, Guliyev noted that as customer workloads continued to grow, state management itself became the primary source of system complexity: local databases, remote snapshots, recovery windows, and partition reassignments had to coordinate correctly. Upstream incidents and large backfills could force many nodes to recover or expand their state simultaneously, making state recovery the next scaling bottleneck.

The premise underlying the third-generation design changed fundamentally: the streaming aggregation service did not need to own the final durable aggregate. Aggregation and deduplication in the streaming layer became best-effort—merging on-time data as much as possible during normal operations and discarding retries identifiable within the active window. Late data, backfills, and work that could not be completed immediately continued flowing to storage without blocking the real-time pipeline.

This shift was feasible because Datadog’s timeseries storage engine already had to reconcile late and out-of-order writes before streaming pre-aggregation was introduced. Extending this existing reconciliation path allowed storage to understand a format representing partial aggregates rather than individual data points—each partial aggregate identifies its series and time bucket, specifies how to merge with other contributions, and includes enough information to detect duplicate deliveries. It is more natural to have the storage system handle these responsibilities than to replicate them in every worker’s RocksDB database. The storage system groups contributions for the same series and bucket, removes duplicates, and merges remaining results during reads and background compaction. This representation remains consistent as data moves from real-time storage to long-term storage.

ソース:Datadog

リサーチ
次へ →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

It should be noted that McCarty's investment thesis comes from Janus Henderson, an asset management company that sells healthcare-related investment products. In his three-stage framework, the first stage has specific descriptions but lacks quantitative adoption data, while the second and third stages are explicitly forward-looking visions; the original text does not provide implementation timelines, cost estimates, or verified efficacy comparisons. The "10-10 rule" (10 years to market, 10% success rate) cited in the drug discovery section is a widely referenced industry statistic, not a new finding in this article. While McCarty believes AI can "bend" this rule, he similarly provides no quantified expectations.

For readers interested in healthcare AI investment opportunities, the value of this interview lies in providing a trackable mechanism path—not just broadly stating "AI will change healthcare," but specifically explaining what AI does in surgery, how it changes physician evaluation methods, and the logic for solving capacity bottlenecks. Meanwhile, de Lagarde's firsthand experience reminds readers that there is a gap between technological hype and patients' actual experiences; distinguishing between the two is itself the core work of investment research.

The recovery model changed accordingly. The streaming layer can now reset directly to the live edge, while storage reconciles deferred data outside the processing path. During one production incident, an operator performed a reset directly through the Datadog Streaming Platform interface, while storage handled the delayed aggregation work. Guliyev wrote that this incident validated the architecture’s main benefit: restoring current traffic no longer required reconstructing all historical aggregation state first.

The broader implication, as Guliyev summarized, is that state does not have to disappear for a processing layer to become easier to scale. When recovery becomes a bottleneck, moving state between workers may not be sufficient—placing durable state into a layer designed to retain and reconcile it preserves correctness while keeping historical state out of the critical path for current work.

This design comes with costs. Guliyev explicitly lists the main limitation of the third generation in his architecture comparison table: reconciliation complexity in the storage and query paths. Correctness is no longer distributed across local databases in each worker but centralized in the storage layer, meaning read-merge and background compaction logic must simultaneously handle partial aggregates, duplicate deliveries, and out-of-order data. This complexity itself becomes a source requiring continuous maintenance.