LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-29198 publications

Actualiser

AI Assistant Startup Wabi Raises $55M at $1.25B Valuation

Sujet · Wabi融资动态Matériel
2026-10-29 22:35 GMT+8

US-based AI startup Wabi announced it has raised $55 million at a $1.25 billion valuation. The round was led by venture firm IVP, with participation from Andreessen Horowitz (a16z), Forerunner, and Amazon Alexa Fund.

Founded by Eugenia Kuyda, who previously created the AI companion Replika, Wabi initially focused on allowing users to build apps via prompts. In September 2026, the company pivoted to a "personal agent" model with the launch of Wabi 2.0. This version integrates into group chats and builds small application interfaces on demand—such as fitness trackers, schedules, or restaurant recommendations—and can take actions like booking appointments.

The funding comes amid intensifying competition in the AI agent sector. Recently, Instinct raised $1 billion at a $10 billion valuation in September, while Tolan secured $30 million in October. Investors see potential in agents that combine conversation with dynamic software generation.

Sources :TechCrunch

Recherche

91APP CISO Lin Dawei: Past Success in Tracking Data Leaks Slowed Down This Investigation

Sujet · 91APP数据泄露事件Matériel
2026-10-29 21:00 GMT+8

Taiwanese e-commerce SaaS provider 91APP held a press conference on October 29, 2026, to release the "91APP Hacking Incident: 25-Day Investigation Report." Chairman Ho Ying-chi led company executives to apologize to affected consumers and brand clients. 91APP stated that due to the inability to confirm exactly which orders were involved, based on the maximum possible scope of impact, the incident affects 1,329 brands and 4,798,225 consumers. The information potentially accessed includes names, phone numbers, delivery addresses, email addresses, and order records, but does not include passwords, credit card details, or other payment data.

Beyond the scale of the impact, the investigation process of this incident is also noteworthy. 91APP was breached on October 3, and the following day, brands reported that consumers had received suspected phishing messages. However, the attack trajectory and vulnerability were not confirmed until October 14, with the breach blocked that same night. Why did it take 11 days from the initial breach to confirming the source of the intrusion? 91APP CISO Lin Dawei reconstructed this investigation process at the press conference.

Lin Dawei admitted that the prolonged investigation was related to the initial direction taken. He explained that an e-commerce order passes through multiple nodes, including the brand's operational computers, the order management system, and the 91APP platform itself. In the past, when investigating suspected order data leaks, the team habitually prioritized checking external nodes, a method that had proven effective over the years; thus, the same approach was used this time. However, the actual source of the intrusion was located within 91APP's internal systems. The team only grasped the attack trajectory after subsequently expanding the investigation to the internal environment. Lin Dawei acknowledged that the initial investigation direction was incorrect and that both internal and external environments should have been checked simultaneously from the start, rather than focusing primarily on external nodes.

According to the joint investigation by 91APP and Aoyi Smart Technology, attackers entered 91APP's internal network through an unpatched internal efficiency tool, Appsmith, and then obtained database keys from the code management platform GitLab, thereby illegally accessing an order database within Commerce Cloud. 91APP noted that the investigation found traces related to the ARTEX AI automated attack tool and attributed this incident to automated attacks using such tools. This incident not only involved third-party tools that were not updated in a timely manner but also included the issue of attackers gaining further access to databases after obtaining internal keys. 91APP stated in its investigation report that the failure of protection and patch management to keep pace with the speed of attacks is an area the company must reflect upon.

In response to this incident, 91APP stated that it has removed the exploited tools, changed system keys and passwords, readjusted GitLab permissions, isolated test and production environments, and restricted network connections between hosts and databases. Regarding future protection, 91APP proposed three major improvement directions: First, to promote a Zero Trust architecture by implementing micro-segmentation of networks and applications, enforcing least privilege, and conducting continuous verification to reduce the scope of access between different systems and lower the risk of lateral movement by attackers within the internal network. Second, to utilize AI to accelerate security detection and response speeds by continuously monitoring for anomalies, analyzing them, and automatically responding via AI, while also using AI to simulate attacks and proactively identify system vulnerabilities. Third, to introduce external security experts for continuous assessment, leveraging third-party independent testing, evaluation, and threat intelligence sharing to supplement potential blind spots in the internal perspective.

Sources :TechOrange 科技報橘

Recherche

NYT Tests AI Recipe Modification in Cooking App, Faces User Backlash Over Mandatory Interface

Sujet · 纽约时报烹饪AI功能Prise rapide
2026-10-29 22:39 GMT+8

According to The Verge, U.S. media outlet The New York Times is testing a new AI feature in its NYT Cooking app that allows users to modify recipes. Announced by NYT Cooking senior editor Genevieve Ko last Friday, the tool recommends substitutions for ingredients and cooking equipment, scales recipes to any number of servings, and aims to "make our recipes work for what a cook needs at that moment." For instance, if a user asks to use baby spinach instead of kale, the AI advises adding it later because baby spinach wilts faster.

The feature is currently rolling out gradually to all iOS Cooking subscribers and will be more broadly available soon. Customized recipes are private and cannot be published or shared via the app. However, the update has sparked drama among users. Some readers expressed concerns in the comment section regarding the environmental impact of widespread AI adoption and the general effect on user experience. When asked if there would be a way to disable the AI feature, Ko responded, "No, not at this time," noting that the feature is "completely optional to use."

This isn't the first time AI has been used by the NYT Cooking app. Ko stated it was used two years ago to convert imperial volume measurements to metric weight, and a serving-size conversion feature was added last year. Regarding the closure of comments on the announcement post, NYT spokesperson Jordan Cohen told The Verge that this is standard practice to allow moderators to focus on new discussions. He noted that many community members are already trying the tool and reporting good results.

Sources :The Verge

Recherche

NYT Reports US Intelligence Suspects Informant Close to Nvidia CEO Jensen Huang

Sujet · 黄乐涉间谍指控Structurel
2026-10-29 21:42 GMT+8

According to a report by The New York Times, relayed by technology media outlet The Decoder, US intelligence agencies believe that Alice Huang (Huang Le), an executive at Singapore-based data center operator Megaspeed, is an informant for China and is close to Nvidia CEO Jensen Huang. The report states that US authorities discovered she passes information to Beijing while monitoring her.

The report indicates that President Trump was briefed about a year ago but did nothing, likely to avoid disrupting Nvidia's chip sales or further straining trade relations with China. US officials believed at the time that Jensen Huang had no idea, so he was not warned.

The US government also suspects Megaspeed of helping China get access to Nvidia chips. Between June 2024 and February 2026, a Malaysian sister company bought at least $4.8 billion worth of AI hardware containing Nvidia chips, according to the report. An Nvidia spokesperson said meeting with customers is part of a CEO's job and that Megaspeed does not receive sensitive information. Megaspeed has announced an internal investigation. A spokesperson for the Chinese embassy dismissed the allegations as "groundless smears and slanders."

Sources :The Decoder

Recherche

French AI Firm Mistral Releases Open-Weight Model Mistral Large 4: Full Weights, Architecture Details and Benchmarks Available

Sujet · Mistral L4模型Matériel
2026-10-29 08:00 GMT+8

French artificial intelligence company Mistral officially released its latest multimodal large language model, Mistral Large 4 (codenamed "Le Chonk"), on the Hugging Face platform on October 29, 2026. The model features approximately 1 trillion total parameters and 49 billion active parameters, utilizing a Mixture-of-Experts (MoE) architecture that accepts text and image inputs. This release marks the transition from preview to fully open weights, allowing developers to access official checkpoints in FP8 and NVFP4 formats.

Technically, Mistral Large 4 combines sparse attention mechanisms with a Pixtral vision encoder. Its text decoder comprises 72 layers, employing Multi-head Latent Attention (MLA) and a DeepSeek Sparse Attention (DSA) indexer, which selects up to 2,048 key positions for each query. This design aims to balance computational efficiency and performance through two forms of sparsity: selecting feed-forward network experts and selecting context positions. Additionally, the model uses YaRN-scaled RoPE (LeRope) to optimize long-context processing.

Mistral highlights the model's defensive capabilities in cybersecurity. According to Artificial Analysis's Cyber Index test, Mistral Large 4 scored 82% in reproducing and patching vulnerabilities in open-source software, ranking among the top five globally. Mistral notes that many closed frontier models refuse such tasks due to safety filters, whereas open weights allow users to control deployment policies, which is critical for security teams needing to reproduce vulnerabilities for remediation. The model also shows competitive performance in coding agents and multimodal understanding benchmarks, achieving 61.7% on DeepSWE v1.1.

The model is distributed under the Mistral Ecosystem Open-Weight License (an MIT derivative with restrictions on model-as-a-service offerings). Day-0 support is available for Transformers, vLLM, and llama.cpp, with community-provided GGUF quantizations for local execution. This release offers industrial enterprises seeking to avoid vendor lock-in and deploy high-performance code and security analysis models in private environments a new verifiable option.

Sources :huggingface.co

Recherche

What Kubernetes Is Missing Is Not Another Platform, But Composition Primitives

Sujet · K8s组合原语之争Matériel
2026-10-29 21:00 GMT+8

Ask a platform team what they built last year, and the answer is likely a Kubernetes controller: it creates several objects in sequence, copies fields between them, contains no business logic, only wiring. A Deployment, preceded by a Service, plus a ConfigMap where the only interesting value is the hostname reported in the Service's status. Kubebuilder turns building the first controller into a pleasant afternoon; the tenth brings its own reconciliation bugs, upgrade path, and on-call rotation. It is the same controller, slightly different each time, existing in thousands of organizations.

In this CNCF-sponsored technical essay, Adam Crowder, an AWS engineer and contributor to the kro project, presents an architectural judgment: Kubernetes has all the ingredients—CRDs define new types, controllers give behavior, server-side apply tracks field ownership, CEL evaluates expressions, OpenAPI types every field—but provides no 'recipe.' Nothing says that several resources form one thing, or that a value in one comes from the status of another and must wait for it.

This missing piece is not hypothetical. Helm does composition at render time; Helm 4 added server-side apply and waiting for resource status. Argo CD has sync waves, hooks, and health checks because manifests themselves cannot say what order they need or when they are done. Crossplane has compositions and functions, with v2 composing native Kubernetes resources directly. Operators do it by hand in Go. Strip these tools down and the kernel is the same: a set of objects, dependencies between them, values flowing along edges, and something that keeps re-checking whether the cluster still matches. Each recent addition reaches for a part the tool originally lacked: order, readiness, per-field ownership, evaluation inside the cluster.

Crowder argues that since everyone is building the same piece, the question becomes whether it should be shared. This is exactly kro's position, which is why it moved into the Kubernetes SIG Cloud Provider in September 2025—as a primitive for other things to build on, rather than yet another end-to-end platform.

kro's approach is to write composition as data. A ResourceGraphDefinition holds two separable parts: the composition (the resource template behind each instance and the CEL expressions referencing between them) and the API (the schema, the generated CRD, and the instances created by application teams). The key mechanism lies in deriving the dependency graph from references within CEL expressions, meaning order is never explicitly declared. The ConfigMap template references the Service's hostname field; this reference itself constitutes the entire ordering rule—the ConfigMap reads from the Service, so the Service is created first, and nobody writes down the order. Explicitly declared order is a second description of a relationship already contained in the manifest, and the two can become inconsistent after renaming or rearranging. Derived order does not drift because it is derived rather than independently defined. Dependency is reference.

Sources :The New Stack

Recherche

AI Chipmaker Iluvatar CoreX Claims Tiangai 300 Matches Nvidia H200 Performance

Sujet · 天数智芯天垓300芯片Matériel
2026-10-29 22:00 GMT+8

Chinese artificial intelligence chipmaker Iluvatar CoreX claims that its new Tiangai 300 graphics processing unit (GPU) can match and even surpass Nvidia’s H200 in real-world AI workloads. This statement highlights efforts by domestic players to close the technology gap with the US chip giant.

According to a report by Chinese tech media outlet Xinzhixun, the Shanghai-based firm showed the Tiangai 300 outperforming the H200 in complex multimodal tasks during a live demonstration for mainland Chinese media outlets on Wednesday. These tasks included text-to-image and text-to-video generation using frontier AI models such as Z.ai’s GLM-5.3. The H200 is the most advanced Nvidia AI processor approved for export to China under current US trade rules.

In one text-to-image test, the Tiangai 300 achieved a 100 per cent completion rate with a throughput of 0.101 images per second, compared with the H200’s 87.5 per cent success rate and 0.086 images per second, although the Chinese chip took 0.64 seconds longer to complete the task. The demonstration corroborated findings from earlier third-party benchmark tests conducted by state-backed Shanghai INESA. Those tests showed the Tiangai 300, unveiled in July, reaching 1.05 times the throughput of the H200 “under typical deployment conditions”. That made it the first Chinese general-purpose GPU to match the US processor in empirical testing, the report said.

A separate report by Chinese financial outlet Wallstreetcn noted that the Tiangai 300 beat the H200 in mixture-of-experts (MoE) throughput by 15 per cent, while slashing first-token latency by up to 30 per cent in long-text processing. MoE is an architecture that routes tasks to specialised sub-networks to save computing power. On the AI training front, the Tiangai 300 delivered 1.33 times the effective throughput of Nvidia’s prior-generation H100 when fine-tuning Alibaba Group Holding’s open-source Qwen3-32B model, the report added.

The performance claims come as China’s top GPU contenders – including Iluvatar, Biren Technology, MetaX Integrated Circuits, Enflame Technology and Moore Threads – are racing to deliver viable home-grown substitutes under Beijing’s self-sufficiency drive and tightening US trade curbs. Nvidia said in August that it had commenced its first H200 data-centre processor shipments to China under the new US rules, ending an extended supply freeze.

Iluvatar, which went public in Hong Kong earlier this year, has laid out ambitious targets to counter US dominance. To prove commercial viability beyond lab tests, Iluvatar said earlier this month that its Tiangai series had reached 10,000-card-scale deployments, giving it crucial experience in system-level delivery and operations maintenance.

Sources :South China Morning Post

Recherche

XREAL Launches World's First Android XR Glasses 'AURA' at $1,279

Sujet · XREAL AURA眼镜Matériel
2026-10-29 17:06 GMT+8

Chinese AR company XREAL officially launched the world's first split-form-factor spatial computing XR glasses running Android XR, named XREAL AURA, during the Google Android Developer Summit in the early hours of October 29. The product starts at $1,279 (approx. 8,520 RMB), with a limited-time launch price of $1,099. Initial pre-orders are open in the US, Canada, Japan, and South Korea, with shipments expected before December.

On the hardware front, the AURA weighs approximately 95 grams and features transparent lenses that allow users to see their surroundings while viewing digital content. It utilizes XREAL's proprietary X Prism optical engine, offering a 70-degree field of view, 1200P resolution, and up to 120Hz refresh rate. System operation and computation are handled by a separate compute unit equipped with the Qualcomm Snapdragon Reality Elite platform, while the glasses themselves use XREAL's custom X1S chip for spatial processing and low-latency display.

Software and ecosystem integration are the primary focus of this release. Built on Android XR, the AURA allows users to access apps and games directly from Google Play. According to XREAL, there are currently over 150 applications specifically designed for XR, along with more than ten XR apps adapted specifically for the AURA. Additionally, XREAL released accessories such as a game controller co-designed with GameSir, which features a dedicated Gemini button to invoke AI assistants during gameplay.

In practical scenarios, the AURA can connect to laptops or handheld gaming devices via USB-C, supporting multi-window displays that project work, entertainment, or gaming interfaces onto a large private spatial screen. For instance, users can watch up to three NBA games simultaneously or expand board game tables into virtual spaces. Overall, the AURA is positioned as a lighter, more affordable, and more open alternative to Apple's Vision Pro.

Sources :zhidx.com

Recherche

Education Publisher McGraw Hill Announces MCP-Based Agentic AI Plugin for Microsoft Copilot

Matériel
2026-10-29 08:00 GMT+8

McGraw Hill, Inc. (NYSE: MH), a leading global provider of education solutions, announced on October 29, 2026, that it will launch an agentic AI education plugin for Microsoft 365 Copilot. The initiative aims to enable higher education and school district customers with Copilot licenses to securely access McGraw Hill's proprietary course materials through Copilot to support teaching and learning.

The integration utilizes the Model Context Protocol (MCP), a standard that allows AI tools like Copilot to connect securely with external content and systems. McGraw Hill becomes one of the first major education companies to make its K-12, higher education, and professional learning content available to Copilot users via this protocol. Through the plugin, educators and students can summarize, translate, and adapt licensed course materials within the Copilot environment to generate personalized study aids, lesson plans, and other resources.

Dave Cortese, Chief Digital Information Officer at McGraw Hill, stated that AI represents a new growth opportunity, enabling the company to bring trusted content to more people across a growing range of platforms while maintaining control over how its content is used and valued. Deirdre Quarnstrom, Vice President of Microsoft Education Experiences, noted that the collaboration helps institutions adopt AI responsibly in a secure, governed manner, leveraging McGraw Hill's credibility and scale.

This plugin is part of McGraw Hill's broader agentic AI strategy, which already includes embedded AI tools like AI Reader and Sharpen in its digital products. According to the official announcement, McGraw Hill will begin a limited beta of the Copilot plugin in early 2027. This move reflects how traditional content publishers are adjusting their distribution strategies as general-purpose large models become primary workflow entry points.

Sources :mheducation.com

Recherche

UK Asset Manager Baillie Gifford Interviews Vinted CFO Maurizio D’Arrigo on Second-hand E-commerce Growth and Future Strategy

Sujet · Vinted增长战略Prise rapide
2026-10-29 19:42 GMT+8

UK asset manager Baillie Gifford has published the latest episode of its "Conversations with private companies" series. The guest is Maurizio D’Arrigo, Chief Financial Officer of Vinted, the European online marketplace for second-hand fashion, interviewed by Robert Natzler of Baillie Gifford.

According to the publisher's description, Maurizio D’Arrigo discusses Vinted's remarkable growth across Europe and the company's future direction. Vinted has made buying second-hand commonplace, bringing millions of items gathering dust in cupboards into the online market.

For professional investors focused on private markets, particularly in consumer technology and the circular economy, this interview offers a first-hand perspective from the company's core management. Although no full transcript is currently available, the reputation of the asset manager and the specific role of the guest make this a resource with clear reference value and reading purpose.

Sources :youtube.com

Recherche

Qatar Investment Authority Reaffirms $500 Billion US Investment Target, Doubles New York Staff

Sujet · 卡塔尔投资局对美投资Matériel
2026-10-29 22:03 GMT+8

Mohammed Saif Al-Sowaidi, CEO of the Qatar Investment Authority (QIA), stated at the CNN Global Perspectives conference that the fund remains committed to investing $500 billion in the United States over the next decade. He noted this target should be achievable even if the fund needs to draw on reserves to support Qatar's domestic economy.

Al-Sowaidi cited the attractiveness of the US economy, which accounts for 25% of global GDP and 60% of tradable financial assets, as the primary driver for increased allocations. To prepare for this deployment, QIA is doubling its New York team to approximately 45 people.

The CEO's mandate upon appointment included preparing QIA to invest windfalls expected from Qatar's liquefied natural gas expansion. This latest statement clarifies the long-term strategic direction of one of the region's largest sovereign wealth funds.

Sources :semafor.com

Recherche

Titus Winters, Former Head of Google's Readability Program: AI's Value to Engineering Lies Not in Accelerating Old Work but in Launching New Projects

Matériel
2026-10-29 21:37 GMT+8

In a podcast, Titus Winters described a picture he repeatedly heard from engineer peers at various companies: compared to a year ago, code output has risen significantly, incident rates have also increased by about 30%, but the number of delivered features has remained basically unchanged. Interpreted through traditional developer productivity metrics, this looks like treading water—or even leaning negative because of the rise in incidents.

Winters believes the problem lies with the metrics themselves. He says people are caring about the wrong things. Traditional productivity measures—feature delivery counts, iteration speed, lines of code—measure throughput on known types of work. AI's unique contribution may not lie in this dimension, but rather in lowering the feasibility threshold for projects: making projects that were previously unworthy to start due to labor costs, complexity, or timespan realistic.

This judgment has the background of his own engineering experience. In 2010, shortly after joining Google, he participated in a large-scale refactoring covering approximately 250 million lines of C++ code. At the time, the entire codebase was maintained by about 10,000 engineers, while the team executing the refactoring consisted of only twelve people. Without large language models, they relied on MapReduce to parse the abstract syntax trees of the entire codebase, match call patterns, automatically generate edits, shard and distribute them, lock down old patterns, and finally delete old APIs. This specialized toolchain later gave birth to clang-format. Winters' analogical logic is: having twelve people complete a migration involving 250 million lines of code did not rely on making each person write faster, but rather on how the toolchain changed the boundaries of feasibility for this task.

Winters projects the same logic onto AI: its value does not lie in allowing the same engineer to write the same code in less time, but in enabling a team to launch projects they would never have considered before—perhaps an architecture migration across the entire codebase, or a system design previously vetoed due to excessive maintenance costs. From the perspective of incident rates, AI might currently be negative; but from the perspective of feasibility, it opens up new working space.

The boundaries of this argument are also worth noting. First, 'previously unfeasible' is a counterfactual judgment—it is difficult to strictly prove that a certain project would indeed not have been attempted without AI. Second, the 30% increase in incidents cited by Winters comes from informal conversations with peer engineers, not controlled measurements or public datasets. He himself admits that from the perspective of incidents alone, this technology might currently be negative. His optimism is built on another yet-to-be-quantified dimension: those projects that did not exist before but were launched because of AI.

Winters' advice is not to abandon measurement, but to change the way we ask questions. Instead of asking 'how much faster did AI make the same work,' we should ask 'what previously wouldn't-have-been-done things has AI made possible?' The former is easy to quantify but may obscure real changes; the latter is harder to measure but may point to AI's deeper impact on software engineering.

Sources :newsletter.pragmaticengineer.com

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Lin Dawei pointed out that automated tools allow attackers to find intrusion paths faster, making traditional periodic protection operations difficult to respond to in time. This comparison reveals the speed gap between automated attacks and manual responses: when attackers use AI tools to accelerate vulnerability discovery and path planning, defense systems relying on regular scans and manual investigations suffer from structural delays. 91APP's investigation experience also illustrates that past successful models can themselves become cognitive anchors, causing teams to stick to old paths when facing new attack vectors, only breaking the deadlock after external experts intervene.

Regarding subsequent consumer handling, 91APP has notified brands within the affected scope and expects to proactively notify potentially affected consumers starting November 2, once brand authorization is obtained, while uniformly accepting inquiries and compensation applications. According to 91APP, consumers can apply for compensation regardless of whether they have already suffered financial losses. However, the actual scale of fraud-related financial losses, the total compensation amount, and the specific compensation methods for those without financial losses are still under investigation and have not yet been determined.

Crowder further points out that kro 0.10's Graph primitive reveals a more fundamental separation: the API was never the input, only one kind of input. A composition reconciles against cluster state, and a new instance of a custom resource is just one of many state changes it can react to. The Graph exists as the composition part alone—with namespaces, but no schema, no CRD, no instances. Label a Service, and an Ingress rule appears; remove the label, and it disappears. Nobody creates an instance.

He uses a precedent from Kubernetes itself to draw an analogy: rolling updates used to be a kubectl subcommand until the Deployment controller took over, making everything above it simpler. Composition looks like it is due for the same migration moment.

It is worth noting that Crowder himself is a contributor to kro, so this argument carries a project bias. The strongest counterargument is that fragmentation may be an organizational problem rather than an architectural defect—teams do not share because there is no motivation or trust, not because the primitive does not exist. Helm, Argo CD, and Crossplane have each already solved the composition problem, and users take what they need. Whether extracting composition into a shared primitive truly reduces total complexity, or merely shifts complexity from writing code to defining data and owning APIs, Crowder offers a partial answer: ten codebases are gone, but ten APIs remain. The cost has moved from writing abstractions to owning them. Teams should count APIs, not controllers.