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

2026-11-0917 投稿

更新

CoreWeave Launches Physical AI Launchables: One-Command Simulation Templates

クイックテイク
2026-10-21 19:32 GMT+8

U.S.-based AI cloud infrastructure provider CoreWeave launched Physical AI Launchables on October 21, 2026, a catalog of pre-configured templates designed to streamline physical AI development environments. The service addresses the common bottleneck where robotics and autonomous vehicle (AV) engineers spend significant time setting up simulations, wiring storage, and waiting for schedulers before running workloads.

The initial catalog covers major simulation tools including NVIDIA Isaac Sim, CARLA, MuJoCo, NVIDIA AlpaSim, and NVIDIA Cosmos 3. Each template has been built and tested by CoreWeave engineers. Users can select a template closest to their task, adjust key settings such as scenarios or parallelism, and execute the workload with a single command. Specific examples include Franka arm pick-and-place tasks in Isaac Sim and GPU-parallel reinforcement learning sweeps in MuJoCo.

The platform also supports integration with AI agents. CoreWeave states that coding agents like Claude Code and Codex can utilize these skills to run simulations directly, bypassing the need to manage cluster infrastructure. Results, including reports, metrics, and media, are returned through a unified interface.

It should be noted that this is a vendor announcement. The claimed efficiency gains and template stability are based on CoreWeave's internal testing and have not been independently verified.

ソース:wf.coreweave.com

リサーチ

Trucking AI Company Hemut: The Key to Agentic Action Is Treating Emails, Calls, and Sensors as System Records

トピック · Hemut AI卡车代理重大
2026-11-09 22:03 GMT+8

During an interview with diginomica at Confluent’s Current event, Pranav Guda, Head of Product at Hemut, offered a useful perspective for understanding where agentic AI fits into operations: the trucking industry’s system records have never been limited to ERP alone. He noted that while people typically treat ERP as the system of record, emails and call logs are also systems of record, carrying just as much—if not more—information than the core ERP. Previously, these were true silos because no one could query in real time what was written in an email or what a driver said on a call. Guda believes agentic AI changes this dynamic: when AI can read and classify these unstructured streams in real time, they transform from unqueryable noise into actionable signals.

Hemut’s mechanism involves streaming three types of data—emails, truck sensor feeds, and driver messages—through Confluent Cloud, allowing the company’s AI agent, Hemi, to read and correlate them. Specifically: when a broker asks when a shipment will arrive, Hemi does not ask the dispatcher but instead checks the truck’s actual location; when sensors report a belt failure, the system notifies the fleet before the driver even discovers it. Hemut Founder and CEO Loki Cheema shared an example: a customer told him the system indicated his truck’s belts had torn; he called the driver to confirm, who replied that indeed the belts were torn and the truck would break down within an hour. Cheema said this is what becomes possible after streaming fault data—the signal comes from the sensors, and the action trigger comes from the AI agent reading that stream and notifying relevant parties, rather than relying on someone’s memory or experience.

At the operational level, Hemut distinguishes between the core pain points of two types of customers. For small fleets, the bottleneck is load booking—which still relies largely on manual phone calls and emails, negotiating one rate at a time. Guda said that aggregating load sources into the system allows agents to mass-send emails, follow up, and even make phone calls to negotiate rates, handling 30 bids simultaneously instead of one. This difference could determine whether a load yields a 15% margin or a 5% loss. Larger carriers face the opposite problem: they have steady freight but are overwhelmed by data entry. Guda cited a carrier with 300 trucks whose order-entry workload became so burdensome that dedicated staff stopped doing it, forcing the accounting team to take over. Cheema estimated that automating just this process saved approximately $235,000 in labor costs. In Hemut’s largest deployment, the company identified roughly 37,000 hours of manual work per year across eight core workflows, equivalent to about $1.3 million in labor costs.

Cheema emphasized an often-overlooked premise: although the trucking industry is highly fragmented, its processes are actually highly standardized. He stated that if you can map out all standard operating procedures (SOPs) for a customer, you can completely compress their workload. This explains why an AI agent can be reused across multiple small fleets—because the underlying logic is identical. However, it must be clearly noted that all figures above are self-reported estimates by the company, with no independent audit or controlled experiments. Hemut participated in the Confluent startup accelerator, the interview took place during Confluent’s own event, and the article was reported by technology media outlet diginomica. The mechanism itself is verifiable—streaming, categorizing, triggering actions—but the upper bound of evidence for performance claims is attributed company statements. The practical value for readers lies in that framework judgment: in a fragmented, low-margin industry, the entry point for agentic AI is often not replacing core systems, but making previously ignored operational shadow data queryable and actionable.

ソース:diginomica

リサーチ

Data Collection Software Firm FormAssembly Announces Platform Overhaul with AI Automation and MCP Support

トピック · FormAssemblyクイックテイク
2026-11-09 22:00 GMT+8

Online data collection and workflow automation software provider FormAssembly Inc. announced a complete platform overhaul on November 9, 2026, expanding its capabilities beyond building forms to automating complex business processes using artificial intelligence. The company stated that the new platform, which took seven months to develop, combines form creation, workflow automation, integrations, and compliance management in a single no-code environment.

The update introduces two major features: conversational AI agents capable of collecting information through natural language interactions, and support for the Model Context Protocol (MCP), enabling external AI agents to interact directly with FormAssembly’s software. MCP is an open standard that allows different AI systems to share context and tool calls.

Founder and Chief Executive Cedric Savarese said users can “explain what you need in plain English to your agent, and the agent will build the entire process in a couple of minutes.” For example, a municipal government could describe an application process for obtaining a permit to hold an event in a public park, and the platform’s assistant, called Fai, would generate the necessary forms, approval workflows, and documentation requirements.

Savarese claimed the AI-assisted builder can reduce development time from days or weeks to minutes. He noted that these estimates were based on recreating existing customer workflows using the new platform internally.

Technically, FormAssembly does not develop its own models but primarily uses frontier AI models from Anthropic PBC. Savarese stated the company has zero-retention agreements with Anthropic and OpenAI Group PBC to protect customer information. The revamped platform also adds governance capabilities that evaluate workflows against predefined compliance policies before publication, flagging potential violations for human review.

ソース:SiliconANGLE

リサーチ

William Imoh, Founder Who Never Wrote Swift, Builds Native Voice App with Codex: 1,000 Days on Duolingo Didn't Teach Him Norwegian

トピック · Codex原生应用开发重大
2026-11-09 21:04 GMT+8

After moving to Norway, William Imoh spent 1,000 days studying Norwegian via Duolingo but still found himself unable to conduct a full conversation. This specific frustration directly led to the creation of Mural, a voice-learning application he later built using Codex.

His background does not point to that of a mobile developer. He holds a degree in chemical engineering, taught himself to code, worked as a product manager for four years, and had never written a single line of Swift. However, in an episode of Lenny's Newsletter's How I AI podcast, he demonstrated how he used Codex and GPT-Live to build native applications for both iOS and Android. Mural is a real-time voice application whose core function allows learners to practice conversations in Norwegian.

During the podcast, he discussed the importance of design language—a specific design prompt made Mural’s appearance distinct from other language-learning apps on the market. He also addressed the per-session cost of GPT-Live-1 and latency thresholds, factors that determine whether a voice-tutoring experience is viable. While these specific figures were detailed in the show, this summary confirms only that they were discussed.

On the operational side, William Imoh organized his team of Grok Bot agents by assigning roles based on characters from The Lord of the Rings. He mapped the various responsibilities of a solo founder onto different characters, using this system to assist in managing a team of about 25 people. The episode also noted that he granted his agents access to 15 open tabs, with one exception.

He mentioned that even after model improvements, there are still certain tasks he must complete personally. What keeps him awake at night is not the fear of being replaced, but rather the fear of dependency.

The practical boundaries of this case are significant: Mural is a voice-conversation application with a relatively clear task structure (inputting voice, outputting voice, and evaluating conversation quality), and Imoh himself is the target user. The conditions that allowed a product manager who had never written Swift to complete this project may not be directly transferable to applications requiring complex state management, backend architecture, or large-scale user testing. His experience illustrates the actual capabilities of current AI coding tools for specific types of tasks, rather than representing a universal replacement for software development.

ソース:lennysnewsletter.com

リサーチ

Korean Creator Solutions Firm Adoba Launches AI Content Platform with Tencent Cloud, Enabling Comparison of Up to 12 Models

トピック · Adoba腾讯云AI平台クイックテイク
2026-11-09 07:58 GMT+8

Global creator solutions company Adoba has announced the launch of 'adobaRo AIGC Studio', an integrated AI content creation service developed in technical cooperation with Tencent Cloud. The platform aims to streamline the fragmented workflow where creators often switch between multiple AI services to generate and verify video and image content.

The core feature, dubbed 'AI Battle', allows users to input a single prompt and view results from up to 12 different generative AI models side-by-side. This design addresses the challenge of selecting the most suitable model for specific creative tasks, as different models excel in different areas, thereby reducing the burden of choice for creators.

Adoba stated that the service leverages its experience in helping over 1,500 creator teams enter the Chinese market and operating more than 2,500 global channels, combined with Tencent Cloud's AI technology. An Jun-han, CEO of Adoba, noted that choosing among numerous AI models has become a new challenge for creators, and the platform was designed to reduce this complexity. Heo Jung-pil, Country Manager of Tencent Cloud Korea, emphasized the significance of connecting diverse generative AI technologies into a practical service for creators.

ソース:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

リサーチ

Investigation: Wearable Sleep Trackers Using Brainwaves to Train AI Models Without Explicit Consent

トピック · 可穿戴脑电AI训练争议重大
2026-11-09 20:00 GMT+8

According to a report by Tom's Hardware, a recent investigation by Science magazine and the Pulitzer Center's AI Accountability Network found that several consumer neurotechnology companies are using users' electroencephalography (EEG) data—electrical signals generated by brain activity—to train artificial intelligence models, without having previously disclosed this use clearly in their privacy policies. These companies have accumulated tens of millions of hours of EEG recordings.

Canadian firm Interaxon, maker of the Muse sleep and meditation headband, is one of the named entities. Its devices monitor electrical activity in the brain alongside other physiological measurements while users sleep. Nadia Kumentas, Interaxon's Chief Marketing Officer, stated that the company has already trained an AI model using about 50,000 hours of EEG recordings and claims it removes identifying information such as names, emails, and addresses before using the data. However, prior to speaking with Science, Interaxon's privacy policy did not explicitly disclose that users' brain recordings could be used to train foundation models. Kumentas acknowledged, "We agree users should be explicitly informed." For customers whose recordings had already been used, this classification came late, likely exposing the company to potential lawsuits.

Emotiv, another wearable EEG maker, is also training its Large Brain model partly on recordings from customers using its high-end devices. Following questions from Science, Emotiv updated its privacy policy to disclose the use of de-identified neural data for model training subject to opt-in consent.

The logic behind these systems differs from training large language models. Instead of feeding AI enormous amounts of text, developers supply it with recordings of electrical activity in the brain. The model then learns recurring patterns that can be adapted to recognize different brain states or assist with specific research tasks. For instance, AI startup Piramidal trained its model on more than 4 million hours of EEG data in partnership with Cleveland Clinic, developing a seizure-detection system that correctly identified nine out of ten seizures in a pilot study. However, this does not mean consumer headbands can read thoughts; most wearable EEG devices use only one to six electrodes, considerably fewer than the 19 to 25 typically found in clinical EEG systems, and their recordings are generally less comprehensive.

Privacy implications in wearable brain technology are not entirely new. In February, an engineer discovered he could access other people's brainwave recordings through a smart sleep mask he had purchased. Furthermore, the investigation revealed that more than two-thirds of the companies studied explicitly allowed corporate transfers of data without user consent. This could be significant if a small neurotechnology firm is acquired, potentially putting its accumulated recordings in the hands of a new owner. Currently, Colorado and California in the United States have amended their consumer privacy laws to cover neural data, although other states lag behind.

ソース:Tom's Hardware

リサーチ

Google Pixel's Moat and Achilles' Heel: When the Phone Becomes a Distribution Channel for AI Services

トピック · 谷歌Pixel AI战略重大
2026-11-09 18:45 GMT+8

After an extended trial of the Pixel 11 Pro XL, INSIDE editor Sisley reached a conclusion: the reasons he cannot part with this phone are almost none of them new to this generation. Recorder has been available since Pixel 4, and Magic Eraser is a software feature accumulated over several generations. What truly changed the editor's perspective was the randomly included Google AI Pro subscription—but this service can be subscribed to independently and is not exclusive to Pixel. In other words, what changed his mind was not the phone itself, but the six-month trial period included with the purchase, which gave him the opportunity to use the service in depth.

This points to a larger strategic shift: the value of Pixel flagships is moving from the phone itself to the software and services Google delivers through it. Hardware remains stable with minor upgrades, while true differentiation lies in software and AI services. The benefit is a consistent user experience continuity, combined with a seven-year system update commitment, meaning once purchased, the device can be used for a long time. The cost is that when hardware is no longer the selling point, software becomes the sole reason.

From the perspective of the subscription economy, the inclusion of an AI Pro trial with the Pixel Pro series is, in a sense, Google seeking long-term users for its own AI services. Notably, the previous-generation Pixel 10 Pro Fold came with a full year of Google AI Pro, whereas the Pixel 11 Pro series offers only six months. The shortening of the trial period suggests that Google may have accumulated sufficient confidence in user conversion or is controlling customer acquisition costs.

However, when a phone's commercial value is built on software, software reliability equals product reliability. During the trial, the editor encountered Recorder recording anomalies more than once: at the moment it appeared to be recording normally, the app had actually stopped working, and it was discovered afterward that the file had never been saved. Both 9to5Google and Android Authority reported similar issues in July 2026. For writers, interview content usually cannot be repeated; once audio is lost, it is difficult to remedy.

External reviews also confirm the judgment that hardware upgrades are limited. David Imel of The Verge gave it an 8/10 but stated plainly that the hardware is very similar to last year's, with even one downgrade, and software features are almost the entire selling point. Will Sattelberg of 9to5Google believes the AI vision Google wants to push comes too early, and most promises cannot yet be realized in actual use. MKBHD gave the Pixel 11 Pro camera a B+ rating, noting good performance but hardware specifications that do not qualify as top-tier.

There is another variable outside Google's control: memory prices. Shakil Barkat, Vice President of Devices and Services at Google, confirmed before the launch that due to severe supply-side memory shortages, the Pixel 11 series must adjust pricing and memory configurations. The price per 1GB of RAM has risen from $2.80 in 2025 to $12 this year, a sixfold increase. As a result, in the Taiwan market, the Pixel 11 Pro XL starts at NT$42,990, up from NT$39,990 for the previous generation, but the 256GB version's RAM shrinks from 16GB to 12GB.

ソース:INSIDE

リサーチ

Alibaba Cloud DevOps Launches 'Review Memory' to Cut AI Code Review False Positives

クイックテイク
検証済み 2026-11-09 18:59 GMT+8

Alibaba Cloud DevOps has launched a new feature called 'Review Memory' for its Intelligent Code Review tool. This capability allows development teams to save technical constraints, business rules, and team preferences clarified during code reviews as reusable rules. These memories are then automatically applied in future merge requests (MRs) to reduce false positives and repetitive explanations caused by missing context.

In practice, AI code reviews often suggest changes that seem reasonable but ignore repository-specific history, such as legacy compatibility requirements or standard exception handling patterns. 'Review Memory' makes this tacit knowledge explicit, enabling the AI to recognize valid differences and avoid erroneous flags in subsequent reviews. For example, if DTO classes in a specific package must retain no-argument constructors for reflection-based compatibility, teams can store this rule, preventing the AI from suggesting their removal.

According to the Alibaba Cloud DevOps team, in an internal test set they constructed, enabling 'Review Memory' reduced incorrect review comments caused by missing repository context by 96.9%. This figure applies specifically to that test set and does not represent overall accuracy across all repositories or scenarios. The feature is now available on the Alibaba Cloud DevOps platform and can be enabled in repository settings.

ソース:Alibaba Cloud Community

リサーチ

Box Solutions Engineer Hébrard: The Bottleneck for AI Agents Is Not the Model, It's Content

トピック · AI代理数据治理重大
2026-11-09 17:23 GMT+8

As enterprises shift from AI copilots to autonomous agents, a constraint that is often underestimated is the data the agents access themselves. Eléonore Hébrard, a solutions engineer at Box, offered a concise assessment during a discussion jointly hosted by Sifted and Box: agents can hallucinate and fail when fed disorganized data or encountering duplicate or outdated files—the problem lies not with the AI, but with the content. She further noted that an agent's speed and capacity make the blast radius of errors far greater than human operational mistakes; therefore, data governance is not optional but a prerequisite.

Leo Rosales, co-founder and CTO of Ralio, approached the issue from another angle: maintaining skepticism toward third-party AI models themselves. He described his company's approach as assuming the model has been compromised and building control layers around that assumption. They define what an agent can and cannot do, directly blocking it when it attempts to execute actions beyond its authorized scope. This is not a localized validation of model output, but a fundamental distrust of the entire model's reliability—here, 'zero trust' means a default hostile assumption toward the model itself.

Mick McNeil, founder and CEO of Deliverance AI, compared agent deployment to new employee onboarding. He would not grant a new salesperson full access to all company systems on day one; instead, he defines their role scope, gradually opens system access, and has managers continuously monitor performance. An agent's autonomy should be earned over time, not granted in one lump sum. He emphasized that while giving agents knowledge, you must also provide context—just as you define role boundaries for humans, you must instruct the agent on how to think and act when it encounters data.

Regarding specific permission management, Hébrard stressed that agents should not have broad system access; permissions must be task-specific, time-bound, and context-aware. Hanah-Marie Darley, co-founder and Chief AI Officer at Geordie AI, pointed out two overlooked risks. The first is silent failure: the agent is running, but the organization lacks visibility into its real-time behavior, meaning problems may occur without triggering any alerts. The second is the dilution of human oversight: many teams fail to explicitly designate who the human is and what information they need to make judgments, resulting in supervision that appears present but actually reduces effectiveness.

The common premise behind these recommendations is that agents are shifting from generating text to executing operations; once connected to databases, the consequences of errors escalate sharply. McNeil pointed out that humans must always remain accountable—humans grant agents the capability to execute, not the other way around. Hébrard requires organizations to enforce dynamic security controls at the content layer, logging every prompt and reasoning step to ensure that 100% of events can be explained during audits.

Boundaries to note: This discussion was sponsored by intelligent content management platform Box, with the four guests representing companies in AI agent development, content management, AI payments, and AI security. Emphasizing the importance of governance complexity and content quality naturally aligns with their business models. Furthermore, all recommendations are practitioner experience-based judgments; no quantified implementation costs, failure rate comparisons, or independently validated framework effectiveness data were provided. They serve as useful practical starting points, not verified standard answers.

ソース:Sifted

リサーチ

lastminute.com CEO Alessandro Petazzi: AI Is Not an Inevitable Disruptor of Online Travel, but a New Channel for Demand

重大
2026-11-09 17:00 GMT+8

Will online travel companies be replaced by AI agents? The answer from Alessandro Petazzi, CEO of lastminute.com, is no. In an interview with diginomica, he addressed a common narrative: whenever travel companies show signs of short-term weakness, the market tends to attribute it to AI. He believes this reasoning conflates a difficult market environment with AI-driven disruption. The real key is not whether AI can chat or make recommendations, but who bears responsibility after the transaction is completed.

Petazzi’s core judgment is that AI platforms will change how holidays are discovered and sold, but they will not change who is responsible for delivering them. Taking Meta’s newly launched Muse personal AI agent as an example, he emphasized that Meta did not announce its intention to replace all online travel companies, but rather needs partners like lastminute.com. The reason lies in their differing business models: AI platforms aim to create demand and surface the right offers to the right audience, whereas online travel companies focus on fulfilling that demand and managing the holiday itself.

This division of labor is underpinned by regulatory and operational responsibilities. Petazzi pointed out that package holidays in Europe come with specific protection requirements; consumers enjoy certain rights once they book, and assuming these responsibilities requires a tour operating license. In other words, when users actually place an order, the platform cannot merely stop at recommendations and price comparisons. If issues arise—such as flight cancellations, destination conflicts, or stranded travelers—the travel company is responsible for rebooking, accommodation, and follow-up care. He attributes such competitive moats to business models and regulatory obligations rather than technical capabilities.

Therefore, lastminute.com does not view AI as an inevitable substitute, but rather as a new customer acquisition channel. Petazzi believes that Google and social media have long played roles in generating travel inspiration and demand, and AI platforms are now joining this space, increasing the number of entry points. If more consumers begin planning trips through AI conversations in the future, the company wants to appear early in these channels rather than passively adapting after demand has already shifted.

The company has already taken concrete steps. Petazzi stated that lastminute.com’s flights, hotels, and packages are now bookable through ChatGPT and Claude. Internal development processes are also changing: in some teams, up to 70% to 80% of code is generated by Claude Code. Additionally, the company has built an internal platform allowing non-technical employees to create and publish AI tools, resulting in over 400 internal tools so far, used to accelerate information access, data analysis, and automation of daily tasks.

These changes are interpreted as improvements in operational efficiency and product experience. Petazzi cited the example of the company developing a dynamic packaging platform within months using AI, whereas it might have taken years previously. This platform can combine flight, hotel, and package prices in real time, helping users find the best travel times within limited budgets. This capability will soon enter conversational interfaces via integrations with ChatGPT and Claude, aiming to improve user experience and increase conversion rates.

ソース:diginomica

リサーチ

Jens Polomski, Founder of German AI Knowledge Platform Snipki: The Value of AI Agents Lies in the Orchestration Layer, Not the Model Itself

重大
2026-11-09 16:00 GMT+8

In February 2026, Jens Polomski, founder of the German AI knowledge platform Snipki, became one of the earliest adopters to deploy AI agents in practical work. He built a small agent team using OpenClaw, with a core coordination agent named Neo responsible for handling emails and other work tasks. Nine months later, Neo was shut down. However, Polomski believes that what truly holds value has survived: not a specific tool, but the underlying idea. OpenClaw itself was merely a combination of various components; today, many vendors have caught up with this approach, and numerous frameworks have emerged.

From this, Polomski offers a clear recommendation for enterprises: select one tool and avoid migrating back and forth between different platforms. He refers to this behavior as 'Tool-Hopping'—exporting context from here, importing it there, and bringing it back, only for everything to change again a month later. He considers this unworkable and crazy in a corporate environment. A better approach is to make choices based on practical factors such as regional considerations, consumer needs, or user experience, and then deepen your engagement with that choice. His conclusion is that it matters less and less which tool you use, and more how you use it.

Behind this judgment lies a more specific mechanism. Polomski points out that on platforms aggregating multiple models, the gap is not created by the models themselves, but by the harness—the set of orchestration tools, agent workflows, and integration methods surrounding the model. Aggregator platforms often only gradually add these capabilities. Meanwhile, if an enterprise faces twenty optional models simultaneously, it can lose its sense of direction. His current practice involves maintaining usage across different vendors like Claude and ChatGPT while running agents on different servers; his driver is actual need rather than loyalty to any specific framework.

However, there is tension within Polomski’s own stance. When asked whether an upper-level agent could be introduced to automatically select the optimal model, he admits that defining a perfect model is difficult, and he himself is curious about it. He believes that this direction will indeed exist in the long term, driven by efficiency, speed, and token costs.

ソース:t3n.de

リサーチ

Economist Noah Smith's Bull Case for Anthropic: The Change Is Not in Employment, But in Corporate Capital Structure

構造的
2026-11-09 15:59 GMT+8

Noah Smith has put forward a bullish thesis for Anthropic’s IPO that stands in stark contrast to the mainstream AI investment narrative. He first clarifies the boundaries of his argument: he is not envisioning a world where Anthropic dominates all businesses or captures 100% of every industry, nor one where AI replaces human jobs. Instead, he focuses on a more specific structural shift—how the capital structure of an efficiently producing economy will be reshaped.

His core mechanism operates as follows: there is a persistent capability gap of several months or more between frontier AI companies (such as Anthropic, OpenAI, and Google) and distilled models. For any enterprise to win in its respective industry, it must secure internal proprietary access to these frontier models; simply purchasing external services at market price is insufficient. Smith contends that truly gaining this advantage requires firms to tie their own capital structures to the frontier AI company, effectively becoming a subsidiary in some sense.

He uses an analogy to distinguish between two possible economic landscapes for AI. If AI were like oil—where everyone must buy oil or downstream products—it would be an arm’s-length transactional relationship. However, Smith envisions a different scenario: every company must become a subsidiary of Anthropic, OpenAI, or Google to survive. These AI giants would act as a few large conglomerates competing simultaneously across every industry through their networks of subsidiaries. In this oligopolistic ecosystem, Anthropic is merely the largest fish.

Smith points out that recent research by economists at Google DeepMind provides formal modeling support for this dynamic. A paper by Imas, Koh, and Legg (2026) demonstrates that as AI coordination capabilities improve, the Coasean constraints on firm size traditionally found in economics may be broken. Consequently, market structure could undergo a phase transition toward large, highly concentrated, AI-heavy firms. Smith believes his vision is slightly more complex than this model—as it retains elements such as human-AI complementarity and humans possessing private data—but the fundamental direction of the dynamics aligns.

However, Smith’s position carries inherent conflicts of interest. He discloses at the end of his post that while he currently holds no position in Anthropic, he will likely purchase stock after its IPO unless major changes occur; meanwhile, he holds stock in TypeSafe, a competitor of Anthropic. In this discussion, Dan Wang and Brad DeLong express greater skepticism regarding Anthropic’s prospects, whereas Smith himself is the proponent of the bull case.

ソース:noahpinion.blog

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

The internal tension of this strategy is that software is both the moat and the Achilles' heel. The editor even tried using Gemini Live's screen-sharing feature to let AI help find the correct system settings page, but Gemini repeatedly provided similar instructions yet failed to guide him to the right location—the AI could see the screen but was still lost within its own system. For a phone that wins on software, damage caused by issues like Recorder missing recordings may be greater than any lag in hardware specifications.

If you pursue top-tier hardware, the Pixel 11 Pro XL will not be the best choice; but if you value software experience and smart features used daily, it is still worth considering. Provided, however, that Google first makes these software features reliable enough.

However, Petazzi also acknowledged that the actual revenue contribution from current AI integrations remains small. His argument is more about strategic positioning: if demand indeed shifts to AI platforms, lastminute.com is already there; if demand still primarily comes from search and social media, the company will not treat AI as an inevitable threat. This judgment pulls the impact of AI on the online travel industry away from abstract valuation concerns and back to specific responsibility boundaries: who generates demand, who completes the transaction, and who bears the cost of errors.