LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-27149 publications

Actualiser

Optiver Head of AI Engineering Noyan Tokgozoglu Discusses Rapid Model Deployment and Cost Control in High-Frequency Trading

Sujet · OptiverPrise rapide
2026-10-27 15:00 GMT+8

Noyan Tokgozoglu, Head of AI Engineering at Dutch market maker Optiver, recently participated in a podcast interview discussing AI engineering practices. The episode focuses on how Optiver balances the extreme speed requirements of high-frequency trading with the need to manage rapid iteration and cost constraints within its internal AI infrastructure.

According to the episode description, Tokgozoglu detailed the core components of Optiver's on-premises AI platform. This includes an inference gateway for routing and budgeting, a process for benchmarking new frontier models against internal golden sets to enable same-day adoption, and a "meta harness" that is agnostic to specific models and testing frameworks. He also shared insights on how the company regained control over AI spending following a series of major large-model releases during the summer.

For readers interested in AI infrastructure investment, this resource offers a specific first-hand perspective on how enterprises in highly sensitive financial environments balance model update velocity with operational costs through architectural design. While a full transcript is unavailable to verify specific technical details, the direct statement from a technology leader at a leading market maker provides clear industry reference value regarding the mentioned architectural patterns.

Sources :youtube.com

Recherche

South Korean AI Security Firm Xenon Launches Enterprise Generative AI Guardrail Solution 'GenGuard'

Prise rapide
2026-10-27 17:04 GMT+8

South Korean AI security company Xenon announced on October 26 the launch of "GenGuard," an enterprise-grade generative AI security guardrail solution designed to address security threats facing generative AI services. The core capability involves evaluating and blocking AI service inputs and outputs according to corporate environments and security policies.

Originally embedded within Xenon's generative AI platform "GenOS," the guardrail functionality has now been developed into a standalone solution. GenGuard allows users to configure security policies using natural language without requiring separate model training. It controls AI requests and responses while defending against harmful content, jailbreak attempts that bypass AI safety mechanisms, and prompt attacks. The solution also includes privacy protection features, detecting both structured data like ID numbers and account details, as well as unstructured personal information such as names and addresses, masking sensitive data sent to external AI services to prevent leakage.

Regarding technical validation, Xenon stated that GenGuard supports an observation mode, allowing enterprises to verify blocking results and their business impact before full activation, as well as analyze false positive cases. The solution also supports on-premise deployment within internal servers or closed networks.

Xenon simultaneously released its proprietary lightweight guardrail model "GenGuard 1.0 Light" and a Korean-language policy evaluation dataset named "KSPG-Bench." According to the company, in its own performance evaluation, GenGuard 1.0 Light ranked first with an average score of 84.3 among five public guardrail models. This figure represents vendor self-testing and has not yet undergone independent third-party verification.

Sources :미래를 보는 창 - 전자신문

Recherche

South Korean Cybersecurity Firm Igloo Reports 83% Auto-Response Rate with New Agentic SOC

Sujet · Igloo代理式SOCPrise rapide
2026-10-27 16:00 GMT+8

South Korean cybersecurity firm Igloo Corporation announced on October 27, 2026, that it has deployed an "Agentic Security Operations Center (SOC)" within its own AI-integrated security center. According to the company's self-reported data, the average time for security tasks dropped from 30 minutes to 12 seconds, and the proportion of automated responses among all incidents rose from 0% to 83%.

The system integrates SIEM (Security Information and Event Management), SOAR (Security Orchestration, Automation and Response), and specialized AI agents. Igloo stated that when attacks occur that are not covered by existing policies, AI agents query device logs, use large language models to filter actual threats, and automatically generate new detection policies. This mechanism reduced the time required to generate detection policies from 120 minutes of manual work to 10 minutes, increasing daily policy generation from 20 to 50.

In terms of event handling, SOAR-based automation increased daily event throughput from approximately 200 to 3,000, a 15-fold increase. While AI handles most analysis and initial responses, Igloo emphasized that critical actions such as server isolation still require final human approval. The company currently operates a "human-in-the-loop SOC" and plans to evolve toward a "human-on-the-loop SOC" in the future.

Sources :미래를 보는 창 - 전자신문

Recherche

South Korean Tech Giant Naver Provides Up to 18.74 Trillion KRW Payment Guarantee for AI Factory, Shifts Operations to Naver Cloud

Sujet · NaverMatériel
2026-10-27 19:54 GMT+8

South Korean internet giant Naver announced on October 27, 2026, that it has signed a $9 billion (approx. 12 trillion KRW) computing infrastructure acquisition contract with Canadian asset manager Brookfield. As part of this agreement, Naver decided to provide a payment guarantee of up to 18.74 trillion KRW for GPU computing service fees payable by its subsidiary Naver Cloud to Brookfield. This amount represents 64.73% of Naver's own capital.

This signing satisfies the prerequisite conditions of the previous $1 billion strategic investment agreement between Naver and Nvidia. Under the July deal, Nvidia acquired approximately 4.5% of Naver's shares through a third-party allotment capital increase, but required Naver to secure at least $9 billion in independent computing infrastructure. Naver stated that with the condition met, it will promptly complete the capital increase procedure for Nvidia.

The transaction structure shows that Brookfield will handle infrastructure investment through its domestic investment vehicle and an Asset Company (Asset Co.), while Naver Cloud will procure GPU-as-a-Service (GPUaaS) from Asset Co. and provide it to customers. Notably, the operating entity changed from the originally planned 100% new subsidiary to the existing Naver Cloud. Although Asset Co. is wholly owned by Brookfield, it will be consolidated into Naver's financial statements under K-IFRS accounting standards because its business depends entirely on Naver Cloud.

Regarding the 18.74 trillion KRW guarantee scale, Naver emphasized that this is the nominal total service fee expected to accrue until the end of 2034, assuming Brookfield fully executes the $9 billion investment. Since fees are paid based on actual usage and occur in stages, this guarantee does not equate to immediate cash outflow or cost burden, but rather serves as a credit mechanism to ensure contract performance.

In terms of deployment plans, Naver will launch 55 megawatts (MW) of computing power in the first half of 2027, expecting to reach a cumulative 100MW by the end of 2027, and expand to 200MW by 2028. These facilities will use 'Gak Sejong' as a core hub and gradually expand into overseas markets.

Sources :바이라인네트워크

Recherche

South Korea's Shinhan Financial Group Reports Record Cumulative Net Profit, Announces Dividend and Buyback

Sujet · 新韩金融集团Q3财报Matériel
2026-10-27 13:50 GMT+8

Shinhan Financial Group, a major South Korean financial holding company, reported on October 27, 2026, that its cumulative net profit for the first three quarters of 2026 reached KRW 5.027 trillion, a 12.7% year-over-year increase and a historical high. This growth was primarily driven by increased net interest income due to rising interest rates, along with continued asset expansion and improved net interest margins (NIM).

Despite the strong cumulative performance, the standalone net profit for the third quarter was KRW 1.5846 trillion, representing a 12.9% sequential decline from the second quarter. The company attributed this drop mainly to a decrease in non-interest income. However, the overall profit growth trend remains robust, supported by balanced increases in fee income and stable management of credit loss costs.

Regarding shareholder returns, the board of directors resolved to set the third-quarter dividend at KRW 740 per share and approved a total share buyback program worth KRW 450 billion. Of this amount, KRW 200 billion will be repurchased within the current year, while the remaining KRW 250 billion is scheduled for January 2027. These measures aim to boost investor confidence through active capital return policies.

Shinhan Financial Group CFO Jang Jung-hoon stated that as results toward the 10% Return on Equity (ROE) target become visible, the company will maintain profitability-centric portfolio management and efficient capital allocation to ensure sustainable growth. Additionally, the group is pursuing the acquisition of a property insurance company to strengthen its non-banking business base and further improve medium-to-long-term profitability.

Sources :zdnet.co.kr

Recherche

Reducto Founding Engineer Omar Alhait: Data Under PDF Redaction Boxes Can Still Be Extracted

Sujet · PDF遮蔽提取漏洞Matériel
2026-10-11 19:30 GMT+8

In November 2025, the U.S. House Committee on Oversight and Accountability released documents related to the Epstein case. The JMail team—built by Riley Walls and Luke Igel within five hours on the night of the release—created an interface allowing users to browse these files like an email inbox, later expanding into applications such as JFlights, JDrive, and JPhotos. Reducto was responsible for parsing over three million pages of documents involved, including redacted PDFs, scans, complex tables, and handwritten records from numerous individuals.

Omar Alhait, a founding engineer at Reducto, described an unexpected issue encountered while processing flight logs at the 2026 AI Engineer World Expo: some PDFs contained black boxes drawn over text, but because Reducto parses PDF metadata rather than just rendered images, the underlying text data remained fully readable despite being visually obscured. Alhait stated that you cannot simply place a box on a PDF and expect it to be truly redacted. He further explained that while some files were correctly redacted, others were not, depending on whether the original document creators understood the layer structure of PDFs.

For the JMail team, this meant they had to implement an additional layer of redaction to ensure information intended to be hidden remained invisible in the final product. Alhait summarized this dilemma with the phrase: "Our tool was working a little too well."

The stronger the extraction capability, the easier it is to violate the document's intended disclosure limits.

This failure mode is easily masked in demonstrations because small samples often do not reveal problems. Alhait recalled that on the night the first batch of files was released, he pushed throughput hard to complete parsing quickly, repeatedly triggering infrastructure alerts. On-call engineer Alvin messaged him late at night saying, "Hey bro, can you slow down?" Small-scale Studio tests can verify if extraction results for a single document are reasonable, but they cannot expose the true performance of a three-million-page corpus under concurrency, diverse formats, and consistency of redaction.

Alhait outlined a two-step evaluation path: first, perform intuitive checks (which he calls "vibe evals") on a small number of documents in Studio to confirm outputs look reasonable; then expand to one hundred or one thousand relevant documents for systematic evaluation using appropriate configurations. He emphasized that configuration itself affects evaluation results—customers might test a vendor using unsuitable API settings, yielding numbers that do not represent the vendor's performance under correct configurations. Therefore, meaningful comparisons must include the actual processing settings used, not just the vendor name.

Regarding the relationship between parsing and extraction, Alhait explained that Reducto’s pipeline first parses (restoring document structure) and then extracts (requesting specific fields according to a schema). The parse output becomes the input for extraction, enabling intermediate operations: for example, with a math textbook, one can first parse the content, split it by chapter boundaries, and then execute extraction on each chapter separately to generate chapter summaries. The extraction schema is like a paintbrush on the canvas—it determines what the system returns, but correctness still depends on the ability to actually extract the requested information from the document.

Sources :youtube.com

Recherche

Warp Founder Zach Lloyd: Software Engineering Is Becoming Factory Engineering

Sujet · Warp工厂工程论Matériel
2026-09-28 01:30 GMT+8

Zach Lloyd, founder of Warp and a former principal engineer who led engineering on Google Docs, made this argument at the 2026 AI Engineer World’s Fair: software engineering is becoming factory engineering. His own experience serves as the starting point for this thesis—despite being the founder of Warp and having previously led engineering at Google Docs, he still ships products frequently but has not written a single line of code in the past six months. The factory he describes operates as a closed loop: ideas enter the system, agents perform triage, complex tasks generate product specifications and technical specifications, humans review the specifications, agents implement the code, both agents and humans conduct reviews, agents verify the output, humans review the final product, it is shipped, monitored, and observed results are fed back into the input stream. The engineer’s role shifts from writing code to building and managing the system that executes these steps.

Lloyd connects his factory thesis to the business logic behind Warp’s decision to open source its project. His reasoning is that when building software becomes cheap, cloning it also becomes cheap, making it difficult to establish a moat based solely on the product itself. Advantages such as distribution, ecosystem, community, brand, data, and capital lie outside the product, yet startups typically lack these resources. While open source can cultivate an ecosystem and community, the traditional maintenance burden—noisy issues, sloppy pull requests, and lengthy code reviews—has historically made this path unrealistic for resource-constrained teams. Lloyd stated that automation changed Warp’s calculus: they built an entire suite of automations around their open-source project, making the public factory both a development system and a mechanism for handling contributions.

The factory’s self-improvement mechanism is a key point emphasized by Lloyd. He cited the example of a code-review agent: the agent leaves comments, which a senior engineer then corrects. Without an additional feedback loop, these corrections only help the current review; with an observer agent, the corrected comments become input for improving the review skill, allowing the next run to use the revised skill. Lloyd explicitly noted that this is a proposed feedback mechanism; the talk did not provide measured improvement metrics or specify how skill revisions are evaluated.

Lloyd also acknowledged the risks inherent in the factory metaphor: it can sound like it is mechanizing engineering work. He concluded by emphasizing that while a factory can organize how changes are built, humans must still guide it toward things worth building. He described his own work as figuring out what customers want and what holds value for them. For recent graduates entering the industry, he advised focusing on adaptability, critical thinking, and speed, while retaining an understanding of underlying systems and architecture—because someone must be able to reason about the code and specifications generated by agents. He mentioned that Warp’s hiring experience has involved hiring more people than ever before, but he clarified that this is merely an observation specific to Warp, not a conclusion about the broader labor market.

Sources :youtube.com

Recherche

Hengrui Pharma Q3 Net Profit Jumps 89% as Innovative Drugs and Licensing Drive Beat

Matériel
2026-10-27 19:30 GMT+8

Jiangsu Hengrui Pharmaceuticals, one of China's largest drug makers, reported a net profit of 2.46 billion yuan for the third quarter, a significant 88.92% year-on-year increase that exceeded the Bloomberg analyst consensus estimate of 2.38 billion yuan. Revenue rose 20.55% to 8.95 billion yuan, also beating forecasts of 8.55 billion yuan.

The company stated in a Tuesday filing that improving innovative drug sales and out-licensing revenue contributed to the growth. This stronger quarter helped Hengrui recover from a disappointing first half, when revenue declined 1.9% year on year to 15.5 billion yuan, and net profit edged up just 0.3% to 4.47 billion yuan.

Hengrui attributed its first-half revenue decline to China's national and regional centralized procurement policies, which aim to lower medicine prices through bulk purchasing. Generic medicine sales plunged 16.1% year on year for the six months ended June 30. The company said it had "proactively" scaled back resources allocated to its generic medicine business to limit related risks.

In contrast, innovative medicines maintained steady growth. For the nine months ended September 30, innovative medicines generated 13.87 billion yuan in revenue, up 17.05% year on year, accounting for 64.59% of total drug sales. Out-licensing revenue jumped 31.12% from a year ago to 2.78 billion yuan. Earlier this month, Hengrui entered into a major licensing deal with Danish pharmaceutical giant Novo Nordisk involving its oral GLP-1 weight-loss drug HRS-1596, worth up to US$2.6 billion.

Despite the beat, Hengrui noted that growth in innovative drug revenue so far this year had fallen short of the company's expectations at the beginning of the year, citing "slower-than-expected progress in hospital access" and intensifying competition.

Sources :South China Morning Post

Recherche

Anthropic S-1 Excludes Co-founder Ben Mann from Founder LLC; Company Says He Is Still Fairly Called a Co-founder

Sujet · Anthropic上市Matériel
2026-10-27 08:31 GMT+8

In its Form S-1 registration statement filed with the U.S. Securities and Exchange Commission (SEC) on October 26, 2026, American AI company Anthropic explicitly excluded Benjamin Mann from its "Continuing Founders" and "Founder LLC" in a footnote on page 248. This means Mann will not participate in determining how Class F common stock is voted and is not part of the cohort of seven founders who jointly control more than 50.1% of shareholder voting power.

This disclosure contrasts with the company's recent public statements. Just two weeks prior, on October 14, Anthropic announced the opening of its first Singapore office and referred to Mann as "cofounder Ben Mann." An Anthropic spokesperson confirmed on Monday that there have been no changes to Mann’s role and stated, "it’s still fair to refer to him as a cofounder."

Ann Lipton, a professor at University of Colorado Law School, noted that while "Founder" is not a strict legal term, the S-1 footnote clarifies the boundaries of Mann's governance rights: he does not have continuing governance rights in the firm, though his economic rights, such as share percentage, remain unknown. Lipton suggested this may be an intentional legal distinction to separate factual founders from those defined as having governance control, aiming to avoid confusion for investors.

Sources :The San Francisco Standard

Recherche

Anthropic Pledges No Ads in Claude in IPO Filing Cover Letter

Sujet · AnthropicMatériel
2026-10-27 19:20 GMT+8

US artificial intelligence company Anthropic has explicitly stated in the cover letter accompanying its Initial Public Offering (IPO) filing that it will maintain a ban on advertisements in its flagship chatbot, Claude. This declaration aims to convey the company's core values to investors and users: prioritizing user experience and trust over maximizing short-term ad revenue.

Anthropic explained in the document that users often bring their most important work and most personal questions to Claude, and therefore the company wants Claude's only incentive to be providing helpful answers. The firm noted, "Once advertising becomes part of a company's revenue targets, the pressure to optimize for engagement tends to grow. We have chosen not to introduce that pressure into Claude."

This policy stance contrasts with competitor OpenAI. Previously, Anthropic criticized ChatGPT for running chat ads, warning they could be more manipulative than traditional ads. In the IPO filing, Anthropic categorizes the "no ads" pledge as one example of placing company values above profit. As another illustration, the document mentions that Anthropic delayed the release of its most powerful model, Claude Mythos Preview (later released as Claude Fable 5), due to safety concerns. Although this move cost the company revenue and market share, it provided approximately $100 million in free Claude credits to cyber defenders.

As future models become stronger, Anthropic expects to face more decisions where it foregoes maximum profit to maintain safety and values. This statement reflects not just product strategy but also the company's strategic choice regarding its long-term business model and brand positioning ahead of going public.

Sources :The Decoder

Recherche

Arize AI's Laurie Voss Discusses Code Review Bottlenecks at AI Engineer Conference: Autonomous Agents Generate More Code but Human Review Fails

Sujet · AI代码审查瓶颈Matériel
2026-09-30 21:30 GMT+8

In a video published on September 30, 2026, from the AI Engineer conference, Laurie Voss, Head of Developer Relations at Arize AI and co-founder of npm, analyzed the new bottlenecks in the era of AI-assisted programming. Voss pointed out that although developers using autonomous agents wrote significantly more code (described as a 741% increase), the actual increment in delivered software was much lower (only about 30%). This gap indicates that code generation is no longer the limiting factor; instead, the human review stage has become the new obstacle.

Voss cited multiple pieces of industry evidence in her talk to support this view. She mentioned OpenAI's zero-human-code product, METR's finding that approximately half of PRs passing SWE-bench tests would not be merged, and Cognition's FrontierCode benchmark results (scoring 88% on SWE-bench Pro but only 29% on real-world mergeability). These data points aim to illustrate the significant gap between existing automated testing/evaluation standards and the ability of code to actually enter production environments.

Additionally, the video covers specific practices currently adopted by the industry to address this challenge. Voss discussed how Cursor and GitHub run review processes in production, and how multi-pass reviews and default suspicion strategies reduce false positives. She also referenced Carlini's agent-built C compiler case and Bun's port of a million lines of Zig code to Rust, which resulted in 13,044 unsafe blocks, to explore the risks of removing humans entirely from the loop. Finally, she noted that automated review tools can be fooled by prompt injections that humans might catch, leaving production as the last 'reviewer' standing.

Sources :youtube.com

Recherche

Microsoft, NVIDIA and Infineon Launch 'Physical AI' Accelerator in Europe to Support Robotics Startups with Cloud Resources and Hardware Access

Sujet · 物理AI增长实验室Matériel
2026-10-22 17:16 GMT+8

On October 22, 2026, Microsoft, NVIDIA, and German semiconductor manufacturer Infineon Technologies announced a partnership with UnternehmerTUM Accelerator, an incubation platform affiliated with the Technical University of Munich, to launch the "Physical AI Growth Lab." The project aims to help European robotics and industrial AI startups overcome barriers from prototype to commercial deployment without requiring any equity in return.

The accelerator is scheduled to run from January to April 2027, selecting 10 to 15 eligible startups. Participants will receive $350,000 in Azure cloud credits, access to NVIDIA's software development kits (SDKs) and open models, and foundational hardware support from Infineon for building physical AI systems, including power semiconductors, microcontrollers, and sensors. Additionally, founders will gain mentorship from over 170 deep-tech experts and access to a network of more than 200 investors.

Startups participating in the program must be based in or targeting the Europe, Middle East, and Africa (EMEA) markets, have a working prototype or minimum viable product (MVP), maintain teams of no more than 20 people, and have been founded within the last three years. This initiative reflects major tech companies' efforts to preemptively position themselves in next-generation applications of embodied intelligence and industrial automation by providing infrastructure and technological ecosystems.

Sources :Source EMEA

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

On schema generation, Alhait distinguished two methods: regular generation builds a JSON schema based solely on user requests via a quick LLM call; enhanced generation first parses the document and then tailors field descriptions based on the document's actual content. The latter makes the schema closer to the structure of real materials, but enhanced generation alone does not establish extraction accuracy across corpora—this still requires subsequent systematic evaluation to verify.

Boundaries to retain: All of the above describes the Reducto team's experience with the JMail project, not an independent benchmark or proof of performance. The team declined to disclose the specific models used for parsing and extraction during the conversation, nor did they provide comparative results or evaluation protocols to establish their claimed performance advantages. The redaction issue reflects the handling of specific government documents and cannot be generalized as a universal defect in all PDF redactions. The evaluation advice (inspect a small sample first, then systematically scale up) is close to engineering common sense; its incremental value lies in its binding to specific failure cases—throughput pressure, evaluation distortion dependent on configuration, and redaction penetration—which only appear simultaneously at scale.