LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-28154 publications

Actualiser

Databricks Releases Open-Source Vital Skills Library, Claims Improved Genie Code Performance on Biomedical Benchmark

Sujet · Vital SkillsPrise rapide
2026-10-28 03:00 GMT+8

Databricks released Vital Skills on October 27, 2026, an open-source Agent Skills library designed to enhance its AI coding agent, Genie Code, with specialized expertise in healthcare and life sciences (HLS). General-purpose AI coding agents often produce plausible code but lack domain discipline, such as silently selecting clinical thresholds or opting for generic statistical methods instead of industry-standard practices. Vital Skills addresses this by providing packaged playbooks that guide agents on which questions to ask, which steps are non-negotiable, and which tested tools to use.

According to Databricks' internal evaluation, the team tested Genie Code with and without Vital Skills across three runs on BiomniBench, a public benchmark for biomedical data-analysis tasks. The results showed that on five tasks within the skills' scope, Genie Code's average score rose from 71.9 to 76.9 out of 100. For eighteen tasks outside the skills' scope, scores remained stable (rising slightly from 69.1 to 71.2), indicating no negative interference. Additionally, in a specific cohort-building task, the free-text cohort F1 score improved from 0.62 to 1.00 when using the skill.

These findings are currently based solely on vendor self-testing and have not been independently reproduced. Vital Skills follows the open Agent Skills standard, allowing users to apply them directly or customize them to their needs, aiming to help analysts handle complex workflows in bioinformatics and clinical research more efficiently without immediate reliance on additional domain experts.

Sources :Databricks

Recherche

BofA, Blackstone, Citi, and Morgan Stanley Announce $63 Billion Financing for Anthropic's Broadcom Chips with Vendor Residual Value Support

Structurel
2026-10-28 13:14 GMT+8

Bank of America, Blackstone, Citi, and Morgan Stanley released a joint press release on October 26, 2026 (US Eastern Time), announcing the arrangement of a $63 billion financing package. The plan involves a newly formed Special Purpose Vehicle (SPV) purchasing Broadcom XPU AI accelerators and networking products to lease to artificial intelligence laboratory Anthropic. Chip deliveries are expected to begin in 2027.

In this transaction, Bank of America, Citi, and Morgan Stanley are arranging the first-lien financing, while Blackstone is the sole provider of the committed second-lien financing. Reuters, citing two people familiar with the matter, reported that the first-lien debt amounts to approximately $43.5 billion, representing 69% of the total financing, though the official press release did not disclose the specific breakdown.

A key risk-sharing mechanism lies in Broadcom's involvement. Beyond acting as the chip supplier, Broadcom provides "residual value support" to the SPV, equal to the amount of the first-lien financing. This means Broadcom will cover any shortfall if the future market value of the AI racks falls below expectations. In its subsequent 8-K filing, Broadcom disclosed that the potential liability cap for this guarantee is approximately $43 billion, describing it as a "low probability" contingent liability. The guaranteed amount increases as racks are deployed and decreases as Anthropic pays rent.

On the same day, Anthropic publicly filed its S-1 registration statement with the U.S. Securities and Exchange Commission (SEC). The S-1 document shows that Anthropic generated $16.3 billion in revenue in the first half of 2026 but suffered a net loss of $80.7 billion, largely due to accounting charges from fair value changes in financial instruments. Operating income turned positive in the second quarter at $974 million.

This transaction is part of Broadcom's previously announced strategic compute capacity agreement. In June, Broadcom partnered with Apollo and Blackstone to establish the AI XPV platform, aiming to support over 20GW of compute capacity by 2028. This $63 billion financing expands upon the platform's initial $35 billion deal. Hock Tan, President and CEO of Broadcom, stated that the deal underscores the power of their advanced technology and confidence in the AI ecosystem. Jon Gray, President of Blackstone, noted that rapid growth in AI demand creates massive opportunities for large-scale compute financing.

Sources :INSIDE

Recherche

Reid Hoffman Discusses AI Regulatory Coalitions and 'Dual Mandate' Strategy

Sujet · 霍夫曼AI监管策略Matériel
2026-10-28 15:01 GMT+8

In the latest episode of the 'Possible' podcast, LinkedIn co-founder Reid Hoffman and Aria Finger delve into the political coalitions forming around AI regulation ahead of the US midterm elections. The episode analyzes various proposals, ranging from Bernie Sanders and Greg Casar's bill to ban superintelligence to Anthropic's own draft frontier AI safety bill.

Hoffman explains why he believes pausing AI development for a congressional process would be a mistake. He argues that AI companies should work with any administration but remain transparent about their political spending, suggesting that government can borrow expertise through councils and working groups. He also makes the case for his 'dual mandate': keeping America in the lead on AI while ensuring it helps people now.

Beyond macro policy, the episode examines Satya Nadella's call to treat AI models as potential insider risks, the public feud between Factory and Cognition, and whether VCs can still back competing AI startups. For readers interested in the intersection of AI governance and business strategy, this podcast offers specific perspectives from an industry leader.

Sources :pdrl.fm

Recherche

Study: Non-English Languages Consume Up to 5x More AI Tokens, Creating a 'Tokenizer Tax'

Sujet · AI分词器税与语言不平等Matériel
2026-10-26 08:00 GMT+8

According to a report by Rest of World published on October 26, 2026, a study conducted by the Coalition for Language Equity in AI Research reveals that different languages consume vastly different amounts of tokens when interacting with large language models, creating what is termed a "tokenizer tax."

The research indicates that among the top 20 most popular languages on the internet, the most expensive ones use between 1.5 and 5 times more tokens than the cheapest. When the sample expanded to 200 languages, the gap reached up to 17 times in certain models. Languages using non-Latin scripts, such as Burmese, Tibetan, Tifinagh, Khmer, Lao, and Ol Chiki, consistently faced disadvantages across nearly all tested models.

This disparity affects more than just billing; it increases latency and reduces the effective context window available to users. For instance, the same sentence input in Malayalam may be split into more fragments than in English, consuming more computational resources. While previous academic work has noted similar phenomena, this study provides broader cross-language quantification, highlighting structural inequalities in AI infrastructure.

Sources :explainx.ai

Recherche

JetBrains Switches IDE Terminal Engine to libghostty to Boost AI Agent Compatibility

Sujet · JetBrainsMatériel
2026-10-28 16:11 GMT+8

Software developer JetBrains has announced that its Integrated Development Environments (IDEs) will switch their terminal emulation engine from JediTerm to libghostty, a library from the Ghostty project, starting with version 2026.3. This change aims to improve compatibility for command-line AI coding agents such as Claude Code and Codex, as well as tools like Neovim and tmux within the IDE.

As modern command-line tools adopt new terminal protocols, the legacy JediTerm lacked support, leading to broken keyboard shortcuts and character display anomalies. The new libghostty engine handles the parsing of terminal control sequences and state management, while JetBrains retains responsibility for rendering and IDE integration. A key improvement is the support for the Kitty keyboard protocol, allowing terminal programs to distinguish between Enter, Shift+Enter, and Ctrl+Enter, and to recognize key release and repeat events. This enables developers to use Shift+Enter for line breaks when inputting multi-line prompts to AI agents without prematurely sending incomplete content, and ensures complex shortcuts are correctly passed to the applications.

Additionally, the new engine correctly processes nearly 50 types of control sequences that JediTerm previously ignored or misinterpreted, significantly improving the display of Text User Interface (TUI) tools by fixing issues like missing line-drawing characters and incorrect color codes. The new terminal also supports theme synchronization, notifying compatible tools like Claude Code, OpenCode, Neovim, and tmux when the IDE switches between dark and light modes, allowing them to adjust their colors accordingly.

Although libghostty can already parse instructions for clipboard copying, desktop notifications, and image display, JetBrains has not yet completed the IDE-side integration for these features, so they remain unavailable in the new terminal for now. The new engine is enabled by default in the 2026.3 Early Access Program (EAP), with official statements confirming it will be included in subsequent beta and stable releases.

Sources :iThome

Recherche

Chainguard Engineer Patrick Smyth Discusses Open Source Security in the AI Era: Mitigating Attacks with SLSA and Sigstore

Sujet · ChainguardMatériel
2026-10-28 15:00 GMT+8

In the context of rapid AI development, the landscape of cybersecurity offense and defense is undergoing profound changes. In an interview about open source security, Patrick Smyth of Chainguard delves into how frontier AI makes attackers more dangerous and the challenges this poses to traditional security response mechanisms.

Smyth points out a significant current trend: "exploits now land before the patch." As AI-assisted tools lower the cost of discovering and developing vulnerabilities, the traditional 30-day remediation window no longer holds up. He reviews this year's waves of AI-assisted attacks and supply chain malware, emphasizing the systemic risks these new threats pose to open source infrastructure.

To address this situation, Smyth proposes covering four main attack surfaces: malware, known CVEs, zero days, and "side doors" like CI/CD, developer laptops, and agent skills. He recommends adopting more proactive defensive measures, including implementing cooldown periods to slow down automated attacks, accelerating CVE fix processes, and signing all build artifacts using SLSA (Supply-chain Levels for Software Artifacts) and Sigstore to ensure supply chain integrity.

Additionally, the episode discusses hardening agent skills, a new attack vector emerging with the proliferation of AI agents. For professionals focused on open source security and cloud-native infrastructure, this episode offers specific response strategies and technical perspectives from a frontline practitioner.

Sources :youtube.com

Recherche

SAP and NTT Announce Intention to Cooperate on Sovereign Cloud and Business AI

Sujet · SAP与NTT主权云合作Matériel
2026-10-28 16:00 GMT+8

SAP SE (NYSE: SAP) and NTT Inc. announced on October 28, 2026, their intention to cooperate in the areas of SAP Sovereign Cloud and SAP Business AI. The planned collaboration aims to bring together SAP’s enterprise application portfolio, SAP Sovereign Cloud, and SAP Business AI capabilities with NTT’s global infrastructure, full-stack integrated operations, deep industry AI expertise, and global customer reach. The goal is to help customers benefit from cloud and AI innovation while retaining control over sensitive data, operations, technology, and security.

As an initial area of cooperation, SAP and NTT plan to work together to provide a broad set of sovereign capabilities that combine the SAP Sovereign Cloud On-Site offering with NTT-managed infrastructure and operations for customers in key markets. SAP Sovereign Cloud On-Site is SAP’s deployment offering for customers with stringent sovereignty, regulatory, and compliance requirements. It enables SAP-managed cloud solutions to be operated in a customer-owned or customer-selected data center, providing organizations with a high degree of sovereignty across local physical control, data residency, and operational assurance.

By combining the capabilities of SAP Sovereign Cloud with NTT’s global reach and expertise in advising, designing, delivering, and operating secure infrastructure, the companies intend to support customers in geographies where strong local control is required. These capabilities are intended to help governments, defense organizations, defense industry companies, critical infrastructure operators, and other regulated enterprises modernize mission-critical processes while maintaining control over sensitive data and operations and addressing compliance requirements.

"For public sector organizations and regulated industries, sovereignty is not just a compliance requirement, it is the foundation for resilience, competitiveness and the responsible adoption of AI," said Dominik Asam, member of the Executive Board and Chief Financial Officer at SAP SE. Yutaka Sasaki, Representative Member of the Board, Senior Executive Vice President and Chief Financial Officer at NTT, stated that by combining SAP’s business applications and AI capabilities with NTT’s deep understanding of industry needs, they aim to help organizations transform how they work and accelerate business outcomes through AI. No specific financial terms or implementation timelines were disclosed in the announcement.

Sources :SAP News Center

Recherche

AI Engineer Interview: Kamil Potrec of Keycard Explains Agent Authentication and SDK/Gateway Tradeoffs

Sujet · 智能体身份验证标准Prise rapide
2026-10-28 16:00 GMT+8

When an agent sends an invoice using your token, who actually made the request? This is the core question posed in the newly published AI Engineer video "Who Is Making This Request? Agent Auth 101." Guest Kamil Potrec from Keycard breaks down how agents get credentials today, including personal tokens, static keys, and over-privileged service accounts, and highlights what's wrong with this picture.

Potrec further elaborates on what proper agent authentication needs: per-agent identity, delegation, short-lived scoped credentials, federation with legacy systems, and comprehensive policy and audit controls. These elements form the foundation of secure agent architecture.

Beyond the theoretical framework, the episode focuses on comparing two technical implementation paths: SDKs and Gateways. Potrec analyzes why SDKs alone may not be enough to keep credentials away from agents and discusses the advantages and tradeoffs of introducing external gateways for credential isolation. For engineering teams building or evaluating agent infrastructure, this perspective offers concrete reference points.

Sources :youtube.com

Recherche

ServiceNow and Google Cloud Expand Partnership to Integrate Action Fabric with Gemini Enterprise for Agentic AI

Sujet · 自动训练数据生成Matériel
Vérifié 2026-10-28 17:07 GMT+8

US software company ServiceNow (NYSE: NOW) and US cloud provider Google Cloud announced on October 27, 2026, an expanded partnership connecting ServiceNow's industry-specific operational context, 'Action Fabric', to Google's 'Gemini Enterprise'. This integration aims to make agentic AI more accessible for mutual customers across multiple industries within the assistant interface they use daily.

According to the joint press release, these new live and planned integrations mean Gemini Enterprise will be capable of reasoning and taking complex, multi-step actions using ServiceNow's widely-adopted industry workflows. With Gemini Enterprise as a unified, single interface, mutual customers can scale AI agents to support employees at headquarters, technicians in the field, or brick-and-mortar retail teams. The integrations maintain enterprise-level security, low-latency, and governance standards expected from both companies.

John Aisien, senior vice president of Central Product Management at ServiceNow, stated, 'Conversational AI is only as useful as the action it drives, and ServiceNow’s partnership with Google Cloud is turning that conversation into work that gets done.' He added that combining the workflow orchestration and governance power of the ServiceNow AI Platform with Gemini Enterprise allows teams to execute and resolve mission-critical tasks without leaving their chosen AI assistant.

Satish Thomas, vice president at Google Cloud, noted, 'Our goal with Gemini Enterprise is to empower organizations to turn AI potential into tangible business value.' He emphasized that this collaboration brings the agentic power of the Google Cloud platform straight into mission-critical ServiceNow systems, allowing teams to instantly trigger actions, resolve complex issues, and serve customers faster.

Initial integrations support common use cases in retail, telecommunications, sales, and IT. Specifically, integrations for Retail store operations, Customer relationship management (CRM), and IT service and operations management (ITSM/ITOM/CMDB) are generally available to mutual customers now. For example, in retail, the combination of ServiceNow Retail Operations with Gemini Enterprise allows brick-and-mortar teams to quickly report and diagnose equipment problems. In IT, technical teams can streamline lifecycle operations directly within chat.

Integrations for Telecommunications network inventory and service order management are expected to be generally available to mutual customers in early 2027. This integration combines ServiceNow Telecommunications Service Management (TSM) and Context Engine to enable telecom customer service teams to streamline the management of high volumes of account, technical, and billing inquiries.

Technically, ServiceNow Action Fabric integrates these workflows headlessly to Gemini Enterprise through Model Context Protocol (MCP) servers and Agent2Agent (A2A) connectivity. ServiceNow will also publish these connectors as plugins alongside joint demos and reference architectures in the Google Cloud Marketplace.

Sources :newsroom.servicenow.com

Recherche

ChipAgents Develops Build-A-Chip-Bench with Nvidia to Evaluate AI Agents on Full Chip Design Workflow

Matériel
2026-10-28 15:03 GMT+8

Semiconductor engineering firm ChipAgents is developing a new benchmark suite called Build-A-Chip-Bench, with contributions from Nvidia to the scoring rubric. The benchmark aims to evaluate AI agents' ability to complete the full chip-design workflow, from workload specification through to GDS layout, using open-source tools and an open-source PDK.

Unlike previous benchmarks that tested isolated steps such as writing RTL code from a specification, Build-A-Chip-Bench requires each design to pass a "hard gate" covering functional correctness against hidden verification IP, routability, hold timing, and DRC (Design Rule Check). Only designs passing all hard checks are scored on Power, Performance, and Area (PPA) constraints.

According to ChipAgents, the best-performing AI agent currently passes the hard gate on only 65% of tasks, achieving an overall score of 16.6%. Frontier models struggle to revisit earlier decisions, protect working behavior while optimizing, and plan around slow tool runs. These results highlight the limitations of current AI in complex engineering tasks, providing important context for investors and engineers tracking AI adoption in the EDA sector.

Sources :semiengineering.com

Recherche

South Korean fintech firm Webkeis joins LG consortium to build financial AI agents on K-EXAONE

Sujet · 韩国独立AI基础模型项目Prise rapide
2026-10-28 10:54 GMT+8

South Korean fintech company Webkeis announced it has joined a consortium led by LG AI Research to participate in the "Independent AI Foundation Model" project supported by the Ministry of Science and ICT (MSIT) and the National IT Industry Promotion Agency (NIPA). The project aims to secure technical competitiveness for domestic AI models and expand their application across industries.

Webkeis plans to integrate its data connection solution "OPERIA" with LG's self-developed AI model "K-EXAONE" to specifically develop AI agents for enterprise fund management and financial institution Management Information Systems (MIS). OPERIA converts natural language queries into SQL statements to extract database information and supports various Large Language Models (LLMs), enabling financial institutions to securely utilize internal data.

Through this collaboration, Webkeis seeks to establish use cases for AI applications in specific financial business scenarios and further expand its existing product lines targeting financial institutions, public agencies, and general enterprises, including fund management agents, agentic banking services, and AI MIS.

Sources :zdnet.co.kr

Recherche

Seoul City and ROKAF Launch 'Air Force AX Hub' to Connect Military AI Needs with Startups

Sujet · 首尔空军AX枢纽Prise rapide
2026-10-28 12:30 GMT+8

On October 28, 2026, the Seoul Metropolitan Government and the Republic of Korea Air Force (ROKAF) held an opening ceremony for the 'Air Force AX Hub' at the Industrial AX Innovation Center within the Seoul AI Hub. The initiative aims to connect the military's demand for AI technology with innovative solutions from private companies, supporting joint research and technology demonstrations to enhance the competitiveness of South Korea's defense industry.

The newly launched hub is an expansion and relocation of the former Air Force AI New Technology Convergence Center, which was previously operated within the main Seoul AI Hub center. The facility has grown from 314 square meters to 1,230 square meters and includes research spaces, server rooms, and a 'National Defense Data Safe Zone.' This secure area allows companies participating in Air Force AX research projects to access and utilize military data for R&D under strict security protocols, addressing the longstanding challenge where private firms could not easily access defense data or verify technologies in real military environments due to security constraints.

Seoul City and the Air Force have been collaborating since 2024 through the initial center and signed a memorandum of understanding in July 2026 to expand this partnership. Currently, they are advancing several open innovation and R&D projects based on Air Force technical needs, including weapon system supply chain risk prediction, unmanned access control systems, AI-based combat planning models, and automatic target identification models. Seoul City plans to support the discovery of five Air Force AX projects annually by 2030 and build a cooperation network involving large defense corporations and national defense-related institutions.

Sources :미래를 보는 창 - 전자신문

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement