LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-18133 publications

Actualiser

Samsung Launches Beta Plugin Connecting SmartThings to ChatGPT for Natural Language Appliance Control

Sujet · Samsung SmartThings ChatGPT IntegrationMatériel
2026-11-18 15:00 GMT+8

South Korean tech outlet ET News reports that Samsung Electronics has released the "Samsung SmartThings (Labs)" plugin and initiated a test service on OpenAI's ChatGPT platform. Users can connect their SmartThings accounts to ChatGPT to control home appliances through natural language commands.

The plugin allows users to check the status of lights and appliances, view recent device operation history, execute preset automation scenes, or switch home modes (such as at-home, away, or night). Currently in an early testing phase, some devices only support status queries, while control functions for security-related devices like door locks are restricted.

This move signals a shift in the smart home industry from reliance on proprietary apps and voice assistants toward deep integration with external general-purpose AI services. Previously, LG Electronics connected its ThinQ platform to ChatGPT in August, and Google introduced the Model Context Protocol (MCP) to Google Home in September to enable access for various generative AIs including Claude and ChatGPT.

Sources :미래를 보는 창 - 전자신문

Recherche

US AI Firm Liquid AI Releases d1.2 Series Decision Models, Tops Third-Party Benchmark and Plans to Open Weights

Sujet · Liquid AI d1.2 SeriesMatériel
2026-11-17 11:23 GMT+8

US AI startup Liquid AI released two new full-modality decision models, d1.2 and d1.2-ultra, on November 17. Both models process text, image, and audio inputs, aiming to provide enterprises with faster, low-cost automated decision-making capabilities. Liquid AI stated plans to open-weight these models in the future.

In the Artificial Analysis (AA) Decision Index v2.1 leaderboard, d1.2-ultra ranked first with a score of 75, leading the second-place Drex Max v1.1 by three points. Competitor Jev, also focused on structured decisions, scored 66 and ranked 13th; d1.2 scored 69 and ranked 9th. This ranking shift marks a performance milestone for Liquid AI in specific decision tasks, surpassing previously dominant models.

Beyond composite scores, Liquid AI published internal probability calibration data. Lower calibration error indicates that predicted probabilities align more closely with actual accuracy, which is critical for developers deciding when to escalate tasks to human review. In tests involving five models, d1.2-ultra and d1.2 achieved calibration errors of 0.8 and 1.2 respectively, significantly lower than Jev 1.14's 2.7 and GPT-6 Luna Decisions API's 7.2.

Functionally, the d1.2 series adds support for five core tasks: information reading, text snippet extraction, image object localization, numerical estimation, and multiple-choice judgment. For example, the model can extract payable amounts directly from invoices or box all trees in a photo with coordinates. According to Liquid AI's self-tests, the accuracy for reading receipt values reaches up to 98%, and image object localization accuracy is 94%.

Regarding pricing, both models continue the strategy of charging only for input tokens, maintaining prices identical to the previous generation. d1.2 costs $0.02 per million input tokens, while d1.2-ultra costs $0.042. This low-price strategy combined with high calibration precision targets enterprise users processing large volumes of documents, image classification, and routing decisions, aiming to capture simple decision tasks currently handled by general-purpose LLMs.

Sources :zhidx.com

Recherche

Kling AI Launches Image Model Kling Image 4.0: Supports Native 8K Generation and Agentic Planning

Prise rapide
2026-11-17 18:17 GMT+8

Kuaishou's subsidiary Kling AI launched its latest image creation model, Kling Image 4.0, on November 17, 2026. According to a report by Zhidx, the model focuses on upgrades in visual realism, text expressiveness, complex instruction understanding, and cross-image consistency. Official introductions state that the new model supports native 8K image generation, can render up to 2,000 characters of text in a single image, and is capable of generating up to 10 coherent images in one go.

Technically, Kling Image 4.0 adopts a unified architecture for understanding and generation, incorporating planning and Agentic capabilities. This design aims to enhance the model's ability to understand physical laws, spatial relationships, and complex instructions, thereby better completing intricate visual tasks such as e-commerce detail page layouts and storyboard design. Hands-on tests showed high detail retention and style transfer capabilities in scenarios like realistic portraits, landscape photography, and game character modifications.

Sources :zhidx.com

Recherche

Gallup Report: Over One-Third of Gen Z Rely on AI for Companionship, Linked to Higher Loneliness Risks

Sujet · Gen Z AI CompanionshipMatériel
2026-11-18 13:01 GMT+8

A new report titled "Friendship in an Online-Offline World" by public polling firm Gallup and the Digital Wellness Lab at Boston Children’s Hospital reveals that more than one-third (33%) of Gen Z rely on artificial intelligence (AI) for companionship. This figure spikes to 50% among daily AI users and stands at 42% among those who have online-only friends.

The survey, conducted between May and August 2026, sampled 15,925 people aged 15 to 22 across 16 countries. It found that while in-person friendships were rated highest for "strength and support" and associated with the most positive emotional experiences, young people who rely on online-only friends or AI as a social companion face higher odds of loneliness than those who do not.

David Bickham, research director of the Digital Wellness Lab, noted that adolescence is a time when individuals thrive through socialization, leading them to seek various contexts for connection. He argued against viewing this as a major deficit, suggesting instead that young people growing up in the AI era are increasingly engaging with the social side of the technology as they use it for information.

The report indicates that while AI may help fill a genuine social gap, it has not yet replaced real-life bonds, with in-person and digitally mediated friendships coexisting. It also highlighted gender differences: Gen Z men were more likely to encounter new acquaintances through gaming, while women cited social media as their primary avenue for connection. For boys with higher social anxiety, AI chatbots offered a pathway to seek connections.

Sources :WIRED

Recherche

London-based AI Legal Tech Firm Perry AI Raises $2.2M Seed Round Led by Haatch

Sujet · Perry AI FundingPrise rapide
2026-11-18 13:00 GMT+8

Perry AI, a London-based startup, has announced the completion of a $2.2 million seed funding round. The round was oversubscribed and led by Haatch, with participation from Antler, Syncorp Ventures, Araya Ventures, and several US super angels.

The company has built an AI legal operating system specifically for private capital, including private equity and venture capital funds. Its platform uses specialized AI agents to review legal documents, flag issues, and prepare drafts while maintaining a persistent record of every right, obligation, and commitment across a fund’s investments. Perry AI argues this institutional memory addresses the lack of context in generic AI tools, which often fail to track how earlier transaction commitments affect later investments.

According to Perry AI, its platform is already used by funds managing more than $1.5 billion across over 1,000 portfolio companies. The company reports that customers experience 90% faster workflows, with deal signing times on live transactions falling from 84 days to 13 days. The new funding will support expansion into the US and Asia, alongside continued growth in the UK and Europe.

Perry AI was co-founded by Shashwat Patel, CEO, and Vaneesa Agrawal, Chief Product Officer. Patel spent 15 years as a private equity and funds lawyer at Debevoise & Plimpton and Simmons & Simmons, while Agrawal brings more than 12 years of experience as a funds lawyer and entrepreneur.

Sources :Tech Funding News

Recherche

IAB Releases AI-Era Attribution Blueprint: Only 8.8% of Post-ChatGPT Visits Are Direct Clicks, Majority Misattributed to Search

Sujet · IAB AI Attribution BlueprintMatériel
2026-11-18 09:30 GMT+8

The U.S. Interactive Advertising Bureau (IAB) released the "AI-Era Attribution Blueprint" on November 12, 2026, addressing the failure of traditional attribution models to accurately record AI contributions in purchase journeys. Citing data from Similarweb, the report notes that among U.S. desktop users who received brand recommendations from ChatGPT, only 8.8% visited the brand's website within seven days via direct clicks from the AI interface. More than half of this traffic was still categorized as coming from search. This indicates that many conversions influenced by AI are erroneously credited to search engines or direct traffic, distorting marketing budget allocations.

The blueprint categorizes AI's role in transactions into three pathways: Human-Owned, AI-Assisted, and AI-Executed. The "AI-Assisted" category is further divided into "AI-Influenced" (where AI sparks interest but humans make the decision) and "AI-Recommended" (where AI narrows options and substantially sways the final choice). IAB emphasizes that even without trackable digital signals, if AI influenced intent or decision-making, it should be classified accordingly to avoid systematically underestimating AI's impact.

Regarding credit allocation, the blueprint introduces the concept of "credit claims," distinguishing between "decision claims" and "contributory claims." When AI determines the specific product purchased, the AI platform can assert a decision claim. When AI merely facilitates the process while humans make the final call, only a contributory claim applies. For instance, if a consumer selects a specific sunscreen brand based on AI advice, the AI platform holds the decision claim; if AI provided comparison information but the consumer made the final judgment independently, it constitutes a contributory claim. This distinction aims to clarify eligibility for credit and the evidence required to support such claims.

In terms of evidence, IAB classifies currently available proof into deterministic evidence (such as UTM tags and referrer domains) and directional evidence (such as shopper surveys and lift tests). The report acknowledges a significant gap in measuring publisher content influence: if an AI response cites a review but the user purchases directly from a retailer without clicking the link, this contribution remains unproven. IAB calls for greater transparency from AI platforms regarding paid recommendations and source citations to bridge this measurement void.

Sources :TechOrange 科技報橘

Recherche

Former US Commerce Secretary Raimondo Criticizes Trump Admin Chip Policy for Prioritizing Revenue Over National Security

Matériel
2026-11-18 03:24 GMT+8

Former US Commerce Secretary and current CEO of RAISE US, Gina Raimondo, publicly criticized the Trump administration's chip export policy at Semafor's Silicon Valley & The World event, arguing that it failed to prioritize national security over corporate revenues in sales to China.

Raimondo recalled that during her time leading the Commerce Department under the Biden administration, she spent significant time with former Nvidia CEO Jensen Huang and AMD CEO Lisa Su to establish rules on what could and could not be sold. "They weren't always happy," she said, adding that she always put national security first and that one cannot "put a price" on America's national security.

She accused the current administration of lacking this consistency, describing the shifting message on chips as: "'No, you can't sell chips.' 'Yes you can, if we can get a 25% cut.' You're either for America's national security or you're not — it's not at a price." This comment directly references the so-called "illegal export tax" mechanism recently investigated by Democratic lawmakers, where the Trump administration allowed Nvidia to sell H200 chips to China provided they were first shipped to the US and subjected to tariffs, effectively taking a cut of the sales.

Regarding new export rules cracking down on China's ability to rent compute, Raimondo stated she started work on that two years ago but then the "buzzer rang and the game was over." While glad they are moving in that direction, she questioned why it took two years. She emphasized that China "wakes up every day" thinking "how do we beat America?"

When asked whether she would run for president on the Democratic ticket in 2028, Raimondo deflected, stating, "My next job is my current job."

Sources :semafor.com

Recherche

Empirical Study by the Federal Reserve Bank of New York: Borrowing from the Discount Window Triggers Market Penalties, While the Bank Term Funding Program Does Not Elicit a Similar Reaction

Matériel
2026-11-17 20:00 GMT+8

How does the market interpret it when banks borrow from the Federal Reserve during a crisis? If borrowing is viewed as a signal of financial weakness, banks may hesitate to seek help, thereby undermining the intended function of the lender-of-last-resort mechanism. In a study published in November 2026, Asani Sarkar, a financial research advisor at the Federal Reserve Bank of New York, alongside Natalia Fischl-Lanzoni, a research assistant at MIT FutureTech, and Martin Hiti, a Ph.D. student at the MIT Sloan School of Management, provided a concrete answer using actual stock data from the March 2023 bank run: borrowing from the traditional discount window indeed triggered market penalties, whereas borrowing from the newly established Bank Term Funding Program (BTFP) did not.

Their method did not directly observe stock price fluctuations but instead tracked changes in investor sensitivity to specific risk factors for each bank. The research team constructed four stock portfolios reflecting weaknesses in bank balance sheets—the ratio of uninsured deposits, unrealized losses on securities, cash holdings, and capital adequacy—and then estimated the beta for each bank relative to these portfolios. A higher beta means that investors begin to more closely link that bank’s stock price movements to a particular type of risk, indicating heightened alertness to that weakness. Before the bank run began on March 7, 2023, these betas were statistically close to zero, suggesting that the market was not particularly focused on these risks. After the run erupted, the betas rose sharply. The key question was: after a bank’s first borrowing from the Federal Reserve, did investors’ sensitivity to bank weaknesses increase further?

The results showed a clear asymmetry. Between March 8 and March 31, 34 banks in the sample first used the Bank Term Funding Program, while 33 first used the discount window. For banks that used only the discount window, investor sensitivity to uninsured deposits and capital adequacy rose significantly within 4 to 7 days after borrowing. For banks that used only the BTFP, there was no significant change in investor risk perception. The authors attribute part of this difference to the terms themselves: for the same batch of underwater bonds, the discount window offered only market-value collateral, whereas the BTFP valued them at face value; additionally, borrowing rates under the BTFP were typically lower than those at the discount window. A bank willing to accept worse terms may have already exhausted better options, or may have needed to pledge illiquid loans that only the discount window would accept as collateral—investors could reasonably interpret this choice as a signal of distress. Furthermore, historical reputation may also play a role: the stigma accumulated over decades of banks shunning the discount window contrasts with the BTFP, a new tool free from such historical burden.

The authors explicitly acknowledge the limitations of this study: it observes a small sample of banks over a handful of trading days in a single, rapidly evolving event, making it impossible to fully separate the signal sent by the act of borrowing from the actual reasons the banks needed to borrow; furthermore, the head-to-head comparison between the two tools is not statistically decisive. However, this asymmetric pattern itself carries policy implications: emergency tools designed for specific crises, with more generous terms and without historical baggage, are more likely to be actually used by banks, thereby functioning more effectively as a lender-of-last-resort facility.

Sources :Liberty Street Economics

Recherche

Mckenzie Lock, Former General Manager of Product, Operations & Creative at Netflix: Just Because It’s Technically Possible Doesn’t Mean It’s Commercially Worth Doing

Sujet · Netflix AI StrategyMatériel
2026-11-18 11:30 GMT+8

Netflix offers over 17,000 titles in nearly 200 countries and regions, supports more than 30 languages, and serves over 300 million users. As AI opens up new possibilities for promoting and distributing such a massive library, Mckenzie Lock, who previously oversaw product, operations, and creative at Netflix, faced a core question that went beyond what models could do: should this be done, and if so, how is it best done? She held firm to an uncompromising goal: identify which parts of the workflow are cumbersome, slow, or flawed, use AI to manage that chaos, while ensuring content quality remains unaffected and the labor of creators is not undermined. Her summary was: most product decisions are essentially choices about how to handle complexity, and AI is no different; good design doesn't eliminate complexity, it places it in the right spot.

Lock's team did not start with a checklist of capabilities from the latest models. Instead, she posed two open questions: What final experience do we want to create for our members? In what ways do current production workflows genuinely fail to meet member needs? The team conducted an imagination exercise—what would we want users to be able to do if we temporarily ignored technical constraints? For example, if a viewer paused a show for three months and then returned, could they first see a personalized recap of the plot? These ideas might not become final products, but their value lies in allowing the team to clearly depict the goal before working backward to determine the necessary technology and workflow. This is the opposite of seeing a new model first and then looking for application scenarios for it.

The challenge Netflix faces is not just about viewer experience. Behind every title launch, there are many professionals involved, including writers, editors, animators, translators, and voice actors. They are familiar with difficulties in the process and may be the most concerned that new technology will change their jobs. Lock's team went to observe frontline work, opened surveys to different roles, and organized discussions among experts and managers around the results. They invited people with differing viewpoints to present their reasons, making potential disagreements visible before the project advanced. Unlike technical teams, large creative teams do not adopt AI just for the sake of adoption; they need to see their own future direction. The team created two-column comparison tables for each workflow: the left column listed budget constraints, quality variance, linear dependencies, and time costs of today's processes; the right column described how one person's work would change with AI, what tasks would decrease, and which judgments would become more important. The team also created imaginary press releases and social media posts about upcoming process changes to report on this transformation in advance, stimulating anticipation among creative workers.

Once the vision was clear, the team still needed to decide where to direct resources. Netflix writes important strategic choices in a binary format, where both sides must be attractive options. In the AI transition, the key question the team faced was: learn in low-risk scenarios first, or cover the works with the largest impact first? Blockbuster hits like Stranger Things have many viewers and wide reach; if AI applications succeed, the returns might be more significant. However, the quality requirements are higher, and the cost of errors is greater. The team chose to experiment first in content and processes with relatively lower risks, learning from real production, and then gradually expanding usage based on technological maturity and confidence. Another question was whether to invest in measurement first or scale up as quickly as possible; Netflix chose the former. Lock emphasized that AI investments must align with the company's core metrics, not just AI-specific new metrics, nor should the primary goal of most AI investment be built on cost reduction.

Sources :红杉

Recherche

Microsoft Redefines Copilot as 'Operating System for Work' at Ignite 2026, Launches Code 2 Models and Microsoft IQ

Sujet · Microsoft MAI Code 2Structurel
2026-11-18 11:00 GMT+8

Microsoft CEO Satya Nadella officially redefined the company's generative AI assistant, Copilot, as the "new operating system for work" during his keynote address at the Ignite 2026 developer conference held in San Francisco on November 17. This strategic positioning marks Copilot's evolution from an early office productivity aid to a core platform layer connecting human users with autonomous AI agents, designed to unify knowledge work, programming, and enterprise business processes.

Nadella stated that Copilot now integrates various forms of human-computer interaction, including chat, collaboration, coding, and autopilot, serving as a single entry point. Similar to how traditional operating systems manage hardware resources, Copilot is described as the "kernel" responsible for scheduling AI models, managing organizational data context, and coordinating diverse business applications. This shift implies that boundaries between previously siloed software like Office, ERP, and CRM will dissolve, with all workloads managed and executed through Copilot.

To support this architecture, Microsoft released several key technologies and product updates. These include new AI models Code 2 and its lightweight variant Code 2 Flash, specifically designed for enterprise coding tasks with a focus on cost-efficiency in code deployment, debugging, and multi-file implementation; as well as the Cyber 2 Flash security model, offering high-performance vulnerability detection and response at lower costs. Additionally, Microsoft launched the Microsoft IQ platform, which uses technologies like WorkIQ to build semantic relationship networks among people, documents, emails, and projects within an organization, enabling AI agents to make decisions based on rich business context.

On governance and security, Microsoft announced the "AI Transparency Initiative," planning to collaborate with the U.S. National Institute of Standards and Technology (NIST) to develop open industry standards ensuring AI behaviors can be independently verified and controlled. Meanwhile, Azure infrastructure received upgrades, with the newly launched Cobalt 200 virtual machines offering up to 50% performance improvement over the previous generation, set to expand globally to support growing AI workload demands.

Sources :바이라인네트워크

Recherche

OpenJDK Discloses 'Java4Shell' Critical Vulnerability with CVSS 9.2 Leading to Remote Code Execution

Sujet · Java4ShellMatériel
2026-11-18 05:35 GMT+8

OpenJDK publicly disclosed the 'Java4Shell' high-severity security vulnerability on November 17, 2026. Located in OpenJDK's Base64 decoder, this memory corruption defect allows for pre-authentication remote code execution (RCE). The vulnerability carries a Common Vulnerability Scoring System (CVSS) 4.0 score of 9.2, rated as critical.

Affected software includes widely used enterprise applications such as Keycloak, Jira, Confluence, JFrog, Jenkins, and PingFederate. Systems like Apache Kafka, Solr, and ElasticSearch also face potential risks due to their reliance on related components. The root cause lies in a logic error within the AVX-512 optimized assembly code, leading to buffer overflows under specific inputs. Oracle has released a critical security patch update to address the issue, and administrators are advised to upgrade affected Java versions immediately.

Sources :phoronix.com

Recherche

Microsoft Unveils Seven In-House AI Models and Weight Licensing Service at Ignite 2026, Challenging OpenAI and Anthropic on Cost

Sujet · Microsoft MAI Model FamilyMatériel
2026-11-18 10:34 GMT+8

US tech giant Microsoft announced on November 17, 2026, at its Ignite conference that it has developed and released seven in-house AI models through its Microsoft AI (MAI) division. These models cover programming, cybersecurity, image generation, real-time voice, medical transcription, and model routing. Mustafa Suleyman, CEO of Microsoft AI, stated that as enterprise token spending rises rapidly, Microsoft aims to provide solutions with performance approaching top-tier models from OpenAI and Anthropic but at significantly lower costs.

In core programming tasks, Microsoft introduced MAI Code 2 and MAI Code 2 Flash. According to Microsoft, MAI Code 2 achieves 98% of the performance of OpenAI's GPT-6.1 Sol across four benchmarks while costing only 40% as much per task; compared to Anthropic's Claude Opus 5.5, it reaches 96% performance at just 5% of the cost. MAI Code 2 Flash targets high-volume repetitive work, with costs approximately 65% of GPT-6 Luna and 20% of Claude Haiku 5.5.

Beyond coding, the new portfolio includes MAI Cyber 2 Flash for continuous monitoring and vulnerability scanning, which Microsoft claims performs comparably to GPT-6 Astra in CyberGymE2E tests at less than 2% of the cost. For images, MAI Image 2.7 ranks first on the Artificial Analysis leaderboard, priced at one-fifth of competitor GPT Image 2.6 Sunburst. Microsoft also released MAI Realtime 1 for low-latency dialogue and Dragon Copilot Transcribe, a medical transcription service using the MAI-Transcribe-2 model, which is already used by about one-third of physicians in the US.

Notably, Microsoft also launched Decision 1.1, a model routing system that automatically selects the most appropriate AI model based on task requirements to avoid sending all workloads to expensive large models. Microsoft states that this system has been adopted by 4,000 customers, runs 50 times faster than GPT-Sol 6.1, and costs approximately 1/79th as much.

In addition to API services, Microsoft announced the "MAI Licensed Weights" service. This initiative allows enterprises to acquire trained model weights and deploy them within their own Azure environments, customizing them with internal data. This means enterprises do not need to rely entirely on Microsoft-managed APIs, giving them control over access permissions, proprietary data, and customization outcomes.

Sources :iThome

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

Experiments in the dubbing workflow best illustrate the difference between whether something should be done and whether it can be done. Netflix's product team once conceived an elegant solution: re-animating actors' lips in post-production so that lip movements matched the dubbed audio perfectly. This solution was technically difficult, produced stunning visual effects, and was the kind of thing that could silence a room during internal demos. They found directors and actors willing to try it out and conducted A/B tests on some titles. The result was that viewer stickiness indeed improved. However, AI also caused some visual glitches that required manual cleanup, and the improvement effect was not evenly distributed—it was only obvious in dialogue-heavy dramatic and romantic scenes. Continuing to invest heavily in scaling this capability was not cost-effective. As AI models continue to improve, this calculation might be revisited in the future, but now Netflix had to answer another question: if this thing should be done, how is it best done?

After finding that lip re-animation was not cost-effective, the team changed direction: they built an AI model to help voice actors identify moments where mouth shapes and translations require special attention, and improved matching by adjusting translations, among other methods. Compared to directly modifying visuals, this solution was less dazzling but closer to the problem the team hoped to solve. Just because it is technically possible does not mean it is commercially worth doing broadly—that is the judgment Lock repeatedly emphasized. The most dazzling solution is not necessarily the one worth expanding.

The same logic appeared in two other workflows: asset recommendation and content tagging. The asset recommendation system initially optimized only for viewing metrics, resulting in a homepage that might turn into a screen full of actor close-ups, unable to distinguish between suspense dramas and nature documentaries. The team worked with creative experts to establish more detailed guidelines and adjusted evaluation methods, focusing not only on viewing results but also on whether assets matched the content's vibe. Content tagging work demonstrated another approach: rather than gradually handing over old processes to machines, they first narrowed the scope of the problem—sorting and streamlining uniformly managed tag concepts, then using multiple models and human review mechanisms to handle remaining tasks. Under the new process, tags for a title go live within minutes of ingestion, whereas the old process took weeks.

Netflix's case demonstrates a method of continuous judgment: starting from user experience and listening to frontline voices; clearly describing future work; openly discussing resource trade-offs; testing in real-world scenarios; and being willing to adjust direction when results deviate from expectations. Lock herself holds a technologically neutral attitude toward AI: AI itself is neither good nor bad; goodness or badness depends on our decisions. Technology will continuously improve its ability to achieve goals, but the harder question is whether we have chosen the correct goals. AI can accelerate the speed of progress, but deciding where to go remains a human responsibility.