LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-0268 posts

Atualizar

Saudi Aramco Acquires 49% Stake in AI Firm Humain for $2.6 Billion

Tópico · Humain AI基建Material
2026-11-02 19:47 GMT+8

Saudi Aramco has agreed to acquire a 49% stake in artificial intelligence infrastructure company Humain for approximately $2.6 billion. According to The Decoder, the deal supports Saudi Arabia's strategic goal of becoming a global AI hub. Founded in May 2025 by Saudi Arabia's Public Investment Fund (PIF), Humain builds data centers, develops cloud services, and trains AI models such as the Arabic language model ALLaM.

Per Aramco's statement, the contribution will occur in 2026 and includes the transfer of Aramco's own AI technology. Humain plans to build three gigawatts of data center capacity by 2030, aiming to serve both domestic demand and international markets as a leading AI exporter. Business momentum is already visible: AMD reported in late August that infrastructure built jointly with Cisco and Humain is now serving customers inside and outside Saudi Arabia.

For Saudi Aramco, this investment extends beyond finance into digital transformation. The oil giant plans to use these AI technologies in its own operations and generate additional revenue by selling its own AI solutions. Previously, through its investment arm Wa'ed Ventures, Aramco acquired a stake in Korean AI chip developer Rebellions. Despite regional geopolitical tensions affecting some construction projects, Humain CEO Tareq Amin stated that the company is leveraging Aramco's experience in securing oil infrastructure to enhance data center protection, noting that demand has not dropped due to the war.

Fontes:The Decoder

Pesquisa

Linux AMD Graphics Driver RADV Adopts LLM Contribution Policy, Allowing Discretionary Closure of AI-Generated MRs

Visão rápida
2026-11-02 19:16 GMT+8

The Mesa RADV driver, which provides Vulkan API support for AMD Radeon GPUs on Linux, has formally adopted a new policy regarding Artificial Intelligence/Large Language Model (LLM) code contributions. The initiative was led by Samuel Pitoiset of Valve's Linux graphics team to address a recent increase in LLM-assisted merge requests (MRs) from external contributors.

The new policy does not outright forbid AI-based contributions but explicitly discourages their use. According to the rules merged into Mesa Git, if a contributor finds a bug or shortcoming in RADV that they cannot address themselves, they should file an issue on the tracker. For LLM-assisted MRs, the policy states that they require a good understanding of the codebase and that reviewing them carries a real cost; reviewers and maintainers may close such MRs at their discretion. Additionally, LLM-written issues are discouraged due to being verbose and hard to read, with users asked to describe issues in their own words. However, LLM-generated reproducers are permitted provided they have been reviewed and demonstrate the bug.

Pitoiset explained in the merge request that the move was intended to get ahead of the growing number of AI-generated patches before developers become overwhelmed. This specific policy implementation reflects how open-source infrastructure projects are establishing clear governance boundaries to protect maintainer resources against the administrative burden caused by the proliferation of automated code generation tools.

Fontes:phoronix.com

Pesquisa

Philips Hue Launches Natural Language AI Automation, But Early Tests Show No Manual Editing and Logic Errors

Visão rápida
2026-11-02 19:26 GMT+8

Smart lighting brand Philips Hue (owned by Dutch company Signify) has launched "Custom AI Behaviors" in its app, enabling users to describe needs in natural language which the AI translates into internal programming languages for automation scenarios. For example, a user might request specific lights to turn on at sunset with changing color themes.

The feature is currently exclusive to owners of the Hue Bridge Pro. While Hue Head of Technology George Yianni described it as a "game changer," third-party tech outlet HueBlog's testing revealed significant limitations in this early version: when the AI misinterprets the scope of an instruction (such as turning on all lights instead of just the specified ones), users cannot manually modify the generated logic as they could with Apple Shortcuts, leaving them stuck with error messages.

This restriction highlights a key bottleneck in current consumer AI automation: without deterministic control fallbacks, total reliance on large model reasoning can lead to frustrating user experiences.

Fontes:TechRadar

Pesquisa

German Software Testing Practitioner Paul Behrend: Central Orchestration Agent Loads Skills on Demand, Replacing Simple Prompts and Multi-Agent Piling

Tópico · 智能体编排与提示词之争Material
2026-11-02 19:00 GMT+8

The application of generative AI in software testing suffers from two common failure modes: simple prompts often produce incorrect or incomplete results, while piling up multiple independent agents increases cost and complexity without necessarily improving quality. Paul Behrend, who has over ten years of experience in software quality assurance, proposes an intermediate path in his practice report—deploying a central orchestration agent that guides the testing process according to preset guardrails and dynamically loads corresponding skills based on the current task, covering test analysis, test design, and test case generation separately. This integrates originally scattered auxiliary tools into a clear and verifiable workflow, producing reliable, traceable results that directly serve test implementation.

However, Behrend discovered a practical boundary when implementing this method: the output quality of skills is tightly bound to the language model they are adapted for. Once the underlying model is switched, the answer quality will show significant deviation. The orchestration architecture itself cannot shield against the impact of model changes; teams still face the risk of quality fluctuations when selecting and migrating models.

Behrend also points out that agents are not a universal replacement—in specific scenarios, classic automation remains the better choice. Ensuring test quality ultimately depends on experienced testers, clear processes, and targeted cooperation with agents; all three are indispensable.

Fontes:Developer

Pesquisa

Alibaba Cloud Launches Qovest, an AI Investment Research Workbench for Professional Investors

Tópico · QovestMaterial
2026-11-02 15:09 GMT+8

On November 2, Alibaba Cloud officially launched Qovest, an AI investment research workbench designed for professional institutional investors. The product aims to delegate time-consuming and repetitive tasks in the investment research process to AI. Users can simply state their research needs in one sentence, and Qovest will automatically retrieve data, collect materials, conduct in-depth analysis, generate reports, and continuously track research subjects at set frequencies.

Unlike one-off Q&A interactions, Qovest's core differentiator is its "Research Space." This feature automatically archives research results, personal trading notes, and questions pending verification, allowing users to build upon previous studies rather than starting from scratch each time. For example, fund managers can define variables such as international gold prices and the US dollar index, enabling Qovest to update data weekly, decompose price fluctuation factors, and generate dynamic tracking dashboards. In future enterprise versions, the "Research Space" will serve as a central hub for collaboration and information sharing within organizations.

To meet the high standards of professional investors regarding real-time data accuracy and clear sourcing, the Qovest team procured authoritative financial databases, deployed them on Alibaba Cloud, and developed a proprietary Agent engineering framework (Harness) tailored for financial scenarios. This framework connects the entire workflow from data invocation to analytical output. According to Alibaba Cloud, in internal horizontal evaluations using the open-source financial benchmark FinFirst and IRAB, Qovest led mainstream products in answer quality, response speed, and cost. This claim remains attributed to vendor self-testing and has not been independently verified.

Currently, Qovest is open for invited testing by individual users via its official website. New users receive free trial credits upon registration, with subscription plans available at 99 RMB or 199 RMB per month. Alibaba Cloud plans to release an enterprise version soon to serve investment research teams at financial institutions and corporations.

Fontes:leiphone.com

Pesquisa

Apple Acquires Agentic Data Engineering Startup Ascend.io, EU Filings Confirm Deal

Tópico · Ascend.ioVisão rápida
2026-10-30 21:11 GMT+8

Apple has acquired Ascend.io, an artificial intelligence data engineering startup, according to reports by Mac & i and The Verge. The transaction details were disclosed in regular filings submitted by Apple to the European Commission under the Digital Markets Act (DMA), which requires large tech companies designated as "gatekeepers" to report such acquisitions.

Ascend.io specialized in "Agentic Data Engineering," developing software to automate and manage data pipelines. Its goal was to use AI agents to help data engineers better manage the data lifecycle. The company had previously offered tools like DataAware to simplify data engineering tasks.

The acquisition follows the pattern of an "acqui-hire," where Apple primarily acquires intellectual property rights and gains the ability to offer jobs to the startup's employees. Ascend.io announced on its website that it is "winding down operations as we prepare for our next great adventure."

This marks another recent small-scale AI acquisition for Apple. Earlier, the company acquired Samplab, a Swiss audio AI startup, and SPI Operations Limited, the operator of Swift Package Index. Similar to previous deals, the financial terms of the Ascend.io acquisition remain undisclosed, as the EU registry does not require detailed monetary disclosures for smaller companies.

Fontes:The Verge

Pesquisa

ECB Researchers: Distinguishing Frequency and Magnitude of Price Hikes, No Sustained Broad Repricing in Food Inflation by 2026

Tópico · 欧央行食品通胀研究Material
2026-11-02

After an energy shock, does inflation become widespread and persistent? For investors, CPI figures alone cannot answer this question—food inflation of 3% could stem from sharp price increases in a few products or from small price hikes across almost all products. The former is often temporary, while the latter indicates that cost-push inflation is taking root.

In a blog post published on November 2, 2026, researchers Chiara Osbat, Christian Höynck, Umberto Maglione, and Elisabeth Wieland of the European Central Bank provided micro-level evidence for this distinction using the "Daily Price Dataset" (DPD). Since 2022, the DPD has collected online supermarket prices in the euro area daily, mapping products to Harmonised Index of Consumer Prices (HICP) categories using an AI classifier. It calculates monthly repricing frequencies to filter out short-term fluctuations and temporary promotions, while allowing continuous updates to estimates within the current month.

The core analytical framework of the article decomposes pricing behavior into two margins: the extensive margin, which refers to how often firms adjust prices; and the intensive margin, which refers to the magnitude of each price adjustment. When shocks are small, firms adjust prices according to their established plans; when shocks are large enough, firms adjust prices earlier than planned—this is known as "state-dependent pricing." Once early price adjustments become widespread, the frequency of price increases rises, potentially leading to non-linear acceleration in inflation.

Data following the 2022 Russia-Ukraine war exemplified this scenario: the frequency of food price hikes continued to climb, and the share of product categories with a price hike frequency exceeding 10% expanded continuously over several months within the food basket. This indicated a widespread and lasting shift in firm behavior, consistent with state-dependent pricing.

The situation in 2026 was different. Following the energy shock caused by the Middle East conflict, the frequency of food price hikes rose in March but to a much lesser extent than in 2022. The gap between the frequency of price hikes and price cuts remained small, and the average magnitude of price hikes changed little. Crucially, after a brief rise, the frequency of price hikes stabilized rather than expanding month by month as it did in 2022. The distribution range of price hike frequencies across various food categories was also significantly compressed in 2026 compared to 2022.

Based on this, the authors concluded that there is currently no evidence suggesting that the 2026 shock is triggering widespread and persistent changes in repricing behavior. This conclusion aligns with broader research indicating that the inflationary effects of an energy shock depend on the composition of the shock itself and the economic conditions prevailing at the time. For investors, this implies that current food inflation is more likely a temporary fluctuation in cost transmission rather than the self-reinforcing price spiral seen in 2022. However, the data only cover up to September 2026, representing a limited time window, and the authors explicitly stated that these views do not necessarily represent the official position of the European Central Bank or the Eurosystem.

Fontes:European Central Bank

Pesquisa

Anthropic Announces CarPlay Support for Claude Voice Mode, Expands Languages to 33

Tópico · Claude企业采用率Visão rápida
2026-11-02 18:07 GMT+8

US artificial intelligence company Anthropic announced on October 29, 2026, via its official account that the voice mode of the Claude iPhone app now officially supports Apple CarPlay. This allows drivers to interact with Claude via voice while driving. This is Anthropic's first public explanation of the feature, although reports from Notebookcheck indicated that the functionality had appeared on some user devices as early as September without an official announcement at the time.

The update also brings other improvements: faster voice responses on iOS and Android platforms, better audio capture in noisy environments or when speaking softly, and more natural-sounding non-English voices. Both platforms have added multi-account switching capabilities, allowing users to switch between accounts without logging out.

Notably, the number of supported languages for Claude's voice mode increased by 24, reaching a total of 33 languages, including Bengali, Danish, Swahili, Thai, Turkish, Ukrainian, and Vietnamese. Currently, five AI chatbots are available on CarPlay: ChatGPT, Perplexity, Grok, Meta AI, and Claude. Google's Gemini is not yet on this list (although Gemini integration within Google Maps is gradually rolling out).

Fontes:INSIDE

Pesquisa

Capgemini CEO Aiman Ezzat Counters AI Disillusionment Narrative with Specific Client Outcomes

Tópico · 凯捷AI投资前景Material
2026-11-02 18:15 GMT+8

In an interview published on November 2, 2026, Aiman Ezzat, CEO of French IT services giant Capgemini, stated that despite numerous stories about organizational disenchantment with perceived ROI from AI investments, he sees enthusiasm remaining high at the C-suite level. "Our clients are committed to continue to invest. They believe that AI will drive value for them, and this remains quite strong," Ezzat said. He noted that while IT budgets are often scrutinized, they cover only part of the business, and his focus is on the actual customer appetite for AI transformation.

To support this claim, Ezzat provided specific examples of delivered outcomes. For an asset management firm, Capgemini built an AI-enabled unified desktop replacing 180 fragmented applications, resulting in approximately EUR 3 million in software development life cycle (SDLC) savings and 30% to 50% faster UX delivery. In another case involving a global shipping and logistics player, an agentic system handling over 27 million transactions annually delivered 40% to 50% productivity improvement and shortened collection times by five to eight days.

Ezzat also highlighted that joint work with hyperscalers and major software providers accounted for over two-thirds (65%) of Capgemini’s bookings for the first half of the year. He argued that clients are moving beyond tool deployment to asking how AI changes their business at scale. However, these figures and conclusions are self-reported by Capgemini's leadership and have not been independently verified, representing a vendor perspective on market confidence rather than an objective audit of industry-wide ROI performance.

Fontes:diginomica

Pesquisa

WIRED Investigation: Over 400 AI Apps Disguised as Daily Tools Hide Abusive Deepfake Features

Tópico · 深度伪造滥用应用网络Material
2026-11-02 18:00 GMT+8

A WIRED investigation published on November 2, 2026, revealed a network of more than 400 connected domains and applications that appear to be ordinary AI tools, such as the "pet portrait" generator Viefur, language tutors, or AI workout coaches. However, under specific conditions, these apps stealthily load web interfaces allowing users to generate sexually violent and child sexual abuse material (CSAM) deepfakes by uploading photos.

The network was discovered by researchers tracking links from thousands of video ads containing explicit deepfake content published on Meta’s Facebook and Instagram. These ads reached tens of thousands of people and prompted investigations by regulators worldwide. According to WIRED's review, the apps and domains appear to be operated by Chinese developers and linked to a Hong Kong registered entity called KeLing Technology Limited.

"This is an organized, commercially run operation: One operator behind hundreds of disposable apps and domains, cloaked to pass App Store review and pushed through tens of thousands of Meta ads under fake payer names," said Paul Bouchaud, lead researcher at the nonprofit AI Forensics. He added that the operators did not expect enforcement to catch up, and so far, it hasn't.

Technical analysis shows that each time an app is downloaded, it sends user location, language, and unique identifiers to a server. The server then determines whether to display the benign version or load the explicit deepfake web interface. One trigger appears to be downloading the app after clicking an associated ad on Meta’s platforms.

Maël Le Touz, a staff threat researcher at security firm Infoblox, identified nearly 500 linked domains, noting they run the same code and reference one another. Hidden within the websites' code is a JavaScript bundle containing configuration details for the entire network, including other domains, encryption keys, and Facebook advertising trackers.

Data indicates the apps began appearing in late 2025, with a spike in spring 2026. By September, around 27 such apps were live on Google’s Play Store, collectively downloaded more than 350,000 times, with many rated suitable for all age groups. One popular app, DiffAI, had over 100,000 downloads and was removed shortly before WIRED contacted Google.

After WIRED's inquiry, Apple removed 19 apps from its App Store, including Viefur. Apple spokesperson Adam Dema stated, "We designed the App Store to be a safe and trusted place for our users, and nudification apps and other pornographic content are strictly prohibited." He explained that while the apps complied with rules during review, developers subsequently changed the website the app pulled content from, replacing approved material with pornographic content.

This incident highlights significant vulnerabilities in major tech platforms' review processes when facing dynamic content loading and "bait-and-switch" tactics. Despite claims by Google, Apple, and Meta that they enforce policies against harmful deepfake content, the network continued publishing new apps weekly for approximately a year.

Fontes:WIRED

Pesquisa

Druid AI Co-founder Pleșea: The Root of Enterprise Agent Misunderstanding Is Ontological

Tópico · Druid AIMaterial
2026-11-02 17:50 GMT+8

A seller describes bricks in one specification, while another buyer searches using different terminology; both sides use different words but refer to the same batch of bricks. For two programs to reach a transaction, they must first agree on what constitutes a brick. Andreea Pleșea, co-founder of Druid AI, addressed this exact problem in her 2011 doctoral thesis: enabling software agents buying and selling goods to find a common language. She discovered at the time that the difficulty lay not in whether agents could execute actions, but in their fundamentally different understanding of the same word—one side described quotes using synonyms, while the other expressed needs through specific requests.

Fifteen years later, she believes enterprises are rebuilding the same maze. Today, every ERP and CRM tool comes with built-in agents. To check information in Salesforce, employees must consult Salesforce’s agent; to check another CRM, they must consult that CRM’s agent. Pleșea notes this returns organizations to the state of 2011 or earlier—every system exists, yet no one knows where information originates, nor is there a unified source of truth. She argues that what is truly needed is not more agents, but a secure, governable framework that orchestrates all agents in one place and audits what each agent does.

Joe Kim, CEO of Druid AI, offers another perspective. He describes traditional integration as working by design—you pre-draw flowcharts and instruct the system on step one, then step two. Agents, however, work by intent—a salesperson’s true goal is to close the deal, meet targets, and earn bonuses; they may follow processes or bypass them, but the optimization direction is the result itself. Kim believes agents will optimize similarly, so they must be designed from the start to be trustworthy and monitorable, rather than mapped incrementally.

However, the intervention of LLMs has not made this problem disappear; it has merely made it easier to handle. Kim states that early conversational platforms required matching user input to predefined intents and filling slots; LLMs have made intent understanding significantly easier. Yet he draws a clear line: when agents need to write data into business systems, this action must proceed through the enterprise’s existing deterministic systems. His reasoning is that while you can use LLMs for reasoning, the step of submitting to an ERP or CRM requires logs, control, and piece-by-piece fine-tuning—it cannot be left to the model to figure out independently.

Kim further points out that the most popular pattern in the current industry uses deterministic systems for routing rather than orchestration—the front end features a router deciding whether a query follows a deterministic workflow or a non-deterministic path. He identifies two problems with this approach: first, you cannot create something atop both paths that leverages both; business value becomes an either/or choice rather than a both/and gain. Second, even if tasks are routed to new models, if control is not exerted over the agents on that side, they will become uncontrollable due to ineffective constraints. Pleșea adds that agents running on desktops multiply this problem—on an enterprise scale, it is difficult to control what each agent does, measure its activity, or audit it.

Fontes:diginomica

Pesquisa

Japan's NIDS: AI Scales Cyber Attacks, Overwhelming Government Reporting Mechanisms

Tópico · AI加速网攻规模争议Material
2026-11-02 17:26 GMT+8

The National Institute for Defense Studies (NIDS) of Japan's Ministry of Defense released a new research analysis on October 30, addressing the series of unauthorized access and data breach incidents in Japan since September. The report, authored by Hayato Sasaki, a special researcher at NIDS' Network Security Laboratory and threat intelligence analyst at JPCERT/CC, offers a new perspective on AI-driven cyberattacks. He notes that while public discussion often focuses on AI's ability to discover unknown vulnerabilities or launch zero-day attacks, recent events show that the primary change brought by AI is not a sudden leap in attack technology, but rather its capacity to scale attacks rapidly, accelerate their pace, and combine multiple techniques to target different systems simultaneously.

According to previous disclosures by JPCERT/CC, these incidents involved various attack methods including exploitation of known vulnerabilities, API abuse, Metabase flaws, and Web Shells. Subsequent investigations confirmed that attackers used AI to scan for known vulnerabilities and attack web services via APIs, with related attack infrastructure also being used against services outside Japan. Sasaki argues that when numerous companies are attacked in quick succession and report to authorities, governments often struggle to quickly identify common attack sources and methods across different incidents. Furthermore, victim companies, busy with their own crisis management, provide limited technical information about the attacks to other security agencies.

The report examines Japan's current cybersecurity incident reporting system, noting that reporting channels are divided by industry regulator, type of compromised information, and attack method. Facing cross-industry, multi-technique automated attacks driven by AI, information tends to be scattered across different agencies, increasing the difficulty of inter-agency integration and analysis, and making it hard for existing response mechanisms to keep up with the speed of attacks. This challenge is not unique to Japan; traditional security alerts often recommend protections for specific vulnerabilities, systems, or industries, which may have limited effect against AI-automated attacks as adversaries can quickly pivot to other vectors.

Regarding defense strategies, Sasaki suggests focusing more on attack infrastructure, particularly core servers deploying AI agents, attack tools, and related data. Sharing this technical intelligence allows security vendors to continuously track attack activities, potentially helping to disrupt them. He emphasizes that for enterprises, recent attacks still primarily exploit known vulnerabilities and management oversights, so basic protective measures like patching vulnerabilities and securing API settings remain effective. At the institutional level, he argues for re-evaluating how incident damages are assessed and prioritized, and establishing pre-defined principles and roles for government-private sector cooperation, to prevent traditional layered reporting processes from lagging behind the speed of AI-automated attacks.

Fontes:iThome

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

The practical implication of this argument is that the bottleneck in multi-agent integration is not connectivity capability, but semantic alignment. LLMs lower the threshold for understanding intent, but business writes still require deterministic channels. For readers evaluating various agent platforms, the key question is not whether an agent can perform a task, but whether two agents can use the same word for the same thing.

It should be noted that Druid AI itself is a company specializing in agent orchestration; its arguments naturally lean toward positioning ontology orchestration as the core bottleneck. Other possible explanations—such as insufficient API standardization, organizational process issues, or limitations inherent to model capabilities—are not sufficiently discussed in this article. Pleșea and Kim’s argument represents a practitioner’s specific judgment, not a conclusion verified through controlled experiments.