LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-30115 publications

Actualiser

ElevenLabs Releases SFX v3 Sound Effects Model with Improved Multi-Sound Prompt Handling

Prise rapide
2026-10-29 20:00 GMT+8

AI audio company ElevenLabs released its latest sound effects model, SFX v3, on October 29, 2026. The model is now available in the ElevenCreative platform and via ElevenAPI.

According to the official ElevenLabs blog, the primary improvement in SFX v3 is its ability to more reliably handle single prompts that describe multiple sounds. For example, users can request crowd cheering, a ball dropping, and subsequent booing in one instruction, and the model better integrates these elements. Additionally, the new version generates more convincing music and vocals, and reduces unwanted artifacts in sounds like gongs, chimes, and whistles, resulting in more natural overall audio quality.

ElevenLabs states that these improvements aim to help video and game creators explore scene sounds more efficiently without trial and error. Developers can integrate this feature into their own products through ElevenAPI.

Sources :ElevenLabs

Recherche

Microsoft Q1 Revenue Rises 19% to $92.3B as Cloud Offsets Device Decline

Matériel
2026-10-29 12:32 GMT+8

US technology giant Microsoft has reported its financial results for the first quarter of fiscal year 2027, ending in September 2026. The company generated $92.3 billion in revenue, a 19% increase year-over-year. Operating income rose 21% to $45.9 billion, while net income surged 32% to $36.6 billion. According to a report by German tech outlet heise online, these figures exceeded market expectations, prompting Microsoft shares to rise nearly 4% in after-hours trading.

This marks the first full quarter under Microsoft's new segment reporting structure. The 'Agents and Infrastructure' segment performed strongly, reaching $76.8 billion in revenue, up 25% year-over-year. Within this, Azure cloud services grew by 48%, and total Microsoft Cloud revenue hit $63.9 billion, up 30%. In contrast, the 'Devices and Consumer' segment lagged, with revenue falling 3% to $15.5 billion. Specifically, Windows licensing and devices revenue dropped 22%, and Xbox gaming declined 2%, although search and advertising grew 12%.

Artificial intelligence product adoption significantly contributed to the results. CEO Satya Nadella told analysts that usage of Microsoft 365 Copilot increased by 50% compared to the previous quarter. While Microsoft did not disclose specific user counts for Copilot, CFO Amy Hood stated that the Microsoft 365 E7 package, launched earlier this year and integrated with Anthropic's Claude technology, now serves millions of users.

For the current quarter, Microsoft guided revenue between $96 billion and $97.1 billion, implying approximately 19% year-over-year growth, which is above the consensus estimate of $95.6 billion. This outlook reinforces investor confidence in Microsoft's ability to monetize AI, demonstrating that cloud and AI demand can offset cyclical weaknesses in the traditional PC market.

Sources :heise online

Recherche

AWS: Combining Knowledge Distillation with Medusa Brings a 1.1-Billion-Parameter Model Close to 70-Billion-Parameter Quality, Accelerating Inference by 10 to 20 Times

Sujet · TinyLlamaMatériel
2026-10-30 00:51 GMT+8

Large language models face a core contradiction in travel platforms: larger models yield more accurate answers, but inference latency and costs also rise. Booking.com's AI trip planning assistant must extract user preferences from dialogue in under one second; smart filters need to convert natural-language searches into structured filtering conditions in real time; and accommodation description generation must control costs across millions of listings. These scenarios' demands on latency and cost make deploying 70-billion-parameter-level models directly unfeasible.

In a technical blog post published in October 2026, Daniel Zagyva, Senior Machine Learning Engineer at Amazon Web Services (AWS), and Laurens van der Maas, Senior Applied Scientist at AWS, demonstrated how to integrate two orthogonally optimized techniques into a single pipeline. The first is knowledge distillation: using a large 'teacher' model to generate predictions on labeled and unlabeled domain data, then training a small 'student' model on these predictions, allowing the student to inherit the teacher's knowledge with significantly reduced parameters. The second is Medusa: attaching multiple prediction heads to a base model, where each head predicts a subsequent token in parallel, transforming sequential token-by-token decoding into multi-token parallel decoding, thereby leveraging GPU parallel computing capabilities to reduce inference steps.

These two techniques operate on different dimensions. Knowledge distillation reduces the computational cost per token (because the model is smaller), while Medusa reduces the number of sequential steps required to generate a text segment (because multiple tokens can be predicted simultaneously). Their acceleration effects are multiplicative rather than substitutive: the distilled model itself is already about 6 times faster, and adding Medusa heads on top provides an additional 2 to 3.6 times speedup, resulting in a combined 10 to 20 times acceleration.

The key methodological contribution lies in integrating both into a unified training pipeline rather than processing them in stages. Specifically: first, fine-tune the teacher model (Llama-3.1, 70 billion parameters) to generate predictions on unlabeled, domain-relevant text; then, during student model training (TinyLlama, 1.1 billion parameters), attach Medusa heads directly to the student model and perform single-stage training using a mix of teacher predictions and human-labeled data. This produces a compact model that inherits both the teacher's knowledge and multi-token parallel prediction capabilities.

The team also compared the engineering trade-offs of two Medusa variants. Medusa-1 requires fine-tuning the base model first, freezing its weights, and then separately fine-tuning the Medusa heads; Medusa-2 trains all components simultaneously in a single stage. Both achieve comparable speedups with no significant quality loss, but Medusa-2's simpler pipeline is more practical for production teams managing multiple models.

Progressive experimental design clearly demonstrates the contribution of each step. TinyLlama with only domain-specific fine-tuning is already significantly better than few-shot GPT-4o, indicating that domain-specific fine-tuning can compensate for insufficient parameter count, though it still lags behind the teacher model. After adding knowledge distillation, the quality gap between student and teacher shrinks by over 90%, and inference speed becomes approximately 6 times the original. Adding Medusa-2 heads further brings the final model close to the 70-billion-parameter teacher in quality, with inference speeds reaching 10 to 20 times faster.

Sources :Amazon Science

Recherche

Unstructured.io Video: GPT-5.6 Sol Misses Numeric IDs on Scanned Invoices Due to Lack of Context

Prise rapide
2026-10-30 04:00 GMT+8

Document AI company Unstructured.io published a YouTube video short on October 29, 2026, sharing its team's observations on the performance of the frontier large language model GPT-5.6 Sol in invoice processing tasks. The video indicates that when handling images of scanned paper invoices, GPT-5.6 Sol's main recognition failures occurred on pure numeric strings such as Tax IDs, phone numbers, and bank account numbers.

In the video description, Unstructured.io explained that these numeric fields typically lack sufficient surrounding natural language text to provide context, meaning the model cannot use linguistic logic to reason or backtrack to correct errors. By contrast, the company stated that on clean digital invoices, GPT-5.6 Sol's reading performance was roughly equivalent to another model named Transform.

The video references an Unstructured.io blog post titled 'RAG Needed: Mediocre Data Agents Need Extraordinary Data' as the source for the full test results. This observation offers a specific failure mode reference for engineers working on intelligent document extraction: relying on general-purpose LLM reasoning may be limited in scenarios involving isolated numeric extraction without semantic context.

Sources :youtube.com

Recherche

Korean Controller Maker FADU Unveils 'File-Semantic' SSD to Offload File System Logic from Host CPU

Sujet · FADU AI存储控制器Prise rapide
2026-10-29 12:00 GMT+8

South Korean storage controller designer FADU has unveiled its File-Semantic SSD (FS-SSD) at the OCP Global Summit. The drive aims to address a bottleneck in AI data centers: in traditional architectures, the host CPU spends significant time handling file lookups, permission checks, and address mapping. FADU's solution embeds these operations directly into the firmware of its sixth-generation controller, allowing GPUs, DPUs, and network cards to access files without passing through the host file system or block layer.

Technically, FS-SSD borrows the mature Network File System protocol NFSv3, encapsulating its requests within standard NVMe commands. FADU states this design leverages the stateless nature of NFSv3 to simplify drive memory management. Additionally, the drive merges two lookup processes—normally handled separately by the file system and the flash translation layer—using a layout based on Linux's F2FS file system, aiming to reduce DRAM usage and optimize garbage collection efficiency.

FADU published preliminary performance figures, but they require careful interpretation. The company claims that in GPUDirect Storage random-read workloads, the host file system and block layer consume about 30% of CPU cycles. In remote file throughput tests, FS-SSD maintained approximately 460,000 ops/s as server cores were reduced from 16 to one, while a standard Linux NFS server dropped from roughly 150,000 ops/s to 40,000. However, these tests were conducted on a PCIe 5.0 host, and key metrics like a 63% reduction in write amplification come from simulations using NVMeVirt software developed by Seoul National University, not real hardware.

Currently, FS-SSD is not an industry standard; FADU has not submitted a file-semantic command set to NVM Express, so clients still require proprietary software. Partner samples are not expected until Q3 2027, when real-world testing in PCIe 6.0 environments will be crucial to verify if this architecture can truly alleviate AI storage bottlenecks.

Sources :TechRadar

Recherche

GE Vernova Q3 Earnings: Gas Turbine Backlog Hits 64 GW on Order Conversions, But Slowing New Signings Weigh on Stock

Sujet · GE VernovaMatériel
2026-10-30 01:15 GMT+8

US energy equipment manufacturer GE Vernova reported third-quarter results on October 28, 2026, with revenue climbing 27% year-over-year to $12.6 billion and adjusted earnings per share reaching $4.81, both exceeding analyst expectations. Despite strong performance, the company's stock declined approximately 2% on the day, reflecting market concerns about the sustainability of its growth.

A key shift lies in the quality of its order book. Firm gas turbine backlog—consisting of signed equipment orders—rose to 64 gigawatts by the end of the quarter, up from 53 GW as of June 30. This increase was driven by the conversion of 13 GW of production slot reservations into firm orders, a faster pace than the 10 GW converted in the previous quarter. However, new contract signings slowed significantly, totaling only 11 GW this quarter compared to 20 GW in the second quarter. New slot reservations also dropped to 8 GW from 18 GW.

In terms of profitability, the Power segment stood out, with organic orders surging 178% to $21.8 billion, pushing the segment's EBITDA margin up 800 basis points to 21.4%. Management raised its full-year Power margin outlook to 19%-21%, up from 17%-19%. Conversely, while the Wind segment posted a small profit in Q3, its full-year loss expectation widened to $400 million-$500 million. Electrification orders fell 9% organically, partly due to large synchronous condenser orders in the prior year period.

Regarding capital allocation, the company doubled its share repurchase authorization to $20 billion and raised its full-year 2026 free cash flow guidance to $16 billion-$17 billion. Management expects total backlog to exceed $200 billion in the fourth quarter.

Market reaction indicates investors are closely watching whether AI infrastructure demand has moved beyond the 'reservation' phase into 'signed contracts.' Barclays analysts previously warned that peak orders, peak free cash flow, and peak supply tightness could all arrive together in 2026. The current slowdown in new signings appears to validate this cautious view, preventing the stock from rallying despite robust backlog data.

Sources :MarketBeat

Recherche

Reddit Announces Shift to Logged-In-Only Experience to Curb AI Scraping and Boost Ads

Sujet · Reddit登录墙策略Matériel
Vérifié 2026-10-30 04:53 GMT+8

Reddit announced on Thursday that it is transitioning to a primarily logged-in experience, marking one of the most significant shifts in its 21-year history. The decision was revealed alongside its third-quarter earnings report, aiming to address growing data scraping issues in the AI era and further unlock the potential of its advertising business.

In an interview with Axios, Reddit co-founder and CEO Steve Huffman stated that the company is making this change to improve user experience, bolster its ads business, and prevent bots from scraping its proprietary data. Chief Operating Officer Jen Wong noted that this does not represent a change in strategy but rather a push to build a more direct relationship with users. Over the past decade, Reddit received significant search traffic because its data corpus was open, but this did not alter the company's core mission.

Huffman emphasized that when users are logged in, they spend more time on Reddit, allowing the company to better understand their interests and target ads more effectively. He revealed that over 80% of monetization comes from the app and logged-in users, limiting short-term risk. According to Wong, engagement among logged-in users is about five times that of logged-out users, resulting in more impressions and better business outcomes for advertisers. Performance advertising currently accounts for more than 60% of Reddit's ad revenue and is its fastest-growing segment.

Regarding implementation, Reddit will begin testing the fully logged-in experience in select markets outside the US this quarter, collecting data before rolling out broader changes gradually. The homepage will remain accessible to logged-out visitors, and users accessing via search engines or shared links will still see specific content. However, participation features like commenting and content ranking will be restricted to logged-in users. Additionally, Reddit is relaunching an incognito browsing mode within its app, allowing users to browse and search without affecting their accounts, though this mode will be read-only.

This shift occurs against a backdrop of declining traditional search traffic. In its shareholder letter, Reddit acknowledged that search referral traffic was "down meaningfully in Q3" and warned that daily users could decline by a "low-double-digit percentage" in Q4 if trends persist. Meanwhile, Reddit is attempting to protect the value of its data as an AI training source by restricting unauthorized scraping. The company previously reached content licensing deals with Google and OpenAI and sued Anthropic and Perplexity in 2025 for unauthorized scraping.

From an investment perspective, Reddit is converting its unique "collection of human experiences" into a more exclusive asset. As demand for high-quality, authentic human conversation data increases for AI models, controlling data access may enhance its bargaining power in licensing negotiations. However, this brings risks of slower user growth, and investors should closely monitor user retention and data licensing revenue in subsequent quarters.

Sources :Axios

Recherche

Intel Reports Q3 2026 Earnings: Data Center Revenue Surges 78%, Marking Strongest Growth in 15 Years

Sujet · 海光信息业绩Matériel
Vérifié 2026-10-30 04:53 GMT+8

U.S. semiconductor company Intel Corporation reported its third-quarter 2026 financial results on October 29, 2026. The company recorded total revenue of $16.5 billion, representing a 21% increase year-over-year. Earnings per share (EPS) attributable to Intel was $0.60 on a GAAP basis, while non-GAAP EPS stood at $0.52.

The Data Center and AI (DCAI) segment emerged as the primary driver of performance, generating $7.3 billion in revenue, a substantial 78% year-over-year increase. David Zinsner, Intel's CFO, characterized this as the company's "strongest data center revenue growth in 15 years." In contrast, the Client Computing and Physical AI Group (CCPG) saw revenue decline by 5% to $8.1 billion.

CEO Lip-Bu Tan noted that despite near-term headwinds in the PC industry, demand for Intel CPUs remains robust, particularly for the Xeon family, which he described as facing "extraordinary demand." The company's immediate priority is unlocking additional supply to meet these needs.

Looking ahead, Intel forecasted fourth-quarter 2026 revenue between $15.0 billion and $16.0 billion, with GAAP EPS guided at $0.10 and non-GAAP EPS at $0.25. Management explicitly warned that the slowdown in the PC market is expected to continue at least through the first half of 2027.

Additionally, Intel announced the completion of a $23 billion public equity offering to support future growth opportunities while maintaining an investment-grade rating. The release also highlighted a planned multiyear collaboration with Oracle to expand Intel Xeon-based computing on Oracle Cloud Infrastructure.

Sources :sec.gov

Recherche

Skyworks Solutions Global Head of AI/ML Mumtaz Vauhkonen discusses AI agents in semiconductor test automation

Prise rapide
2026-10-30 01:44 GMT+8

Skyworks Solutions' Global Head of AI/ML and Data, Mumtaz Vauhkonen, along with AI engineer Kordag Mehmet Kilic, recently discussed the company's deployment of AI agents in semiconductor workflows. In this LangChain podcast episode, they explain how LangGraph and LangSmith are being used for end-to-end test automation, root cause analysis, and design assistance.

A key metric highlighted is the improvement in root cause analysis efficiency. According to the episode description, a manual process that previously took about 20 days has been reduced to 10 days on a first pass using these agentic systems. Each use case within their workflow is tracked with its own specific Key Performance Indicators (KPIs).

The conversation also covers the transition from homegrown tooling to LangSmith. Kordag Mehmet Kilic explains that this shift allowed deployment, streaming, and thread management to come out of the box, while evaluations could run on their own datasets and metrics. Mumtaz Vauhkonen details why Skyworks chose LangGraph and LangSmith to handle complex environments involving on-premises and cloud data, third-party tools, and multiple databases, emphasizing quick iterations to get agents into production.

Sources :youtube.com

Recherche

Amazon's Operating Margin Hits Record High, But AI Capex Drives Massive Free Cash Flow Outflow

Sujet · 亚马逊AWS电力瓶颈Structurel
Vérifié 2026-10-30 04:39 GMT+8

U.S. cloud and e-commerce giant Amazon.com, Inc. announced its third-quarter financial results for the period ended September 30, 2026, on October 29. The company reported operating income of $28.1 billion and an operating margin of 13.9%, marking the highest operating margin in its history. However, despite this improvement in profitability, the company's cash position deteriorated significantly due to massive investments in artificial intelligence infrastructure.

According to the filing, Amazon's free cash flow for the trailing twelve months (TTM) shifted from an inflow of $14.8 billion in the prior year to an outflow of $42.4 billion. This reversal was primarily driven by a year-over-year increase of $94.8 billion in net purchases of property and equipment, which the company explicitly attributed to investments in artificial intelligence. Although operating cash flow increased by 29% to $168.3 billion, the heavy capital expenditures completely offset the cash generated from operations.

CEO Andy Jassy stated that AWS's investment in functionality and infrastructure is working, with AWS net sales growing 39.3% year-over-year in Q3 to reach a $184 billion annualized revenue run rate. He projected revenue growth rates of over 40% for both 2027 and 2028. These figures illustrate the typical financial profile of tech giants in the current AI race: demonstrating business viability through high growth and margins while absorbing substantial upfront cash burn to build computational capacity.

Sources :sec.gov

Recherche

Illumio Study Finds 70% of Organizations Delegate Decisions to AI Agents, But Only 49% Can Audit Them

Sujet · AI智能体控制缺口Matériel
2026-10-30 04:20 GMT+8

A study by cybersecurity firm Illumio has identified a growing disconnect between the authority granted to AI agents in enterprises and the ability to monitor their actions. According to the report published on October 29, 2026, 70% of organizations are now delegating decisions that were previously made by humans to AI systems. However, only 49% of these organizations possess the capability to see and audit the specific actions taken by these AI agents.

This 'control gap' poses significant security risks. The study found that 76% of organizations experienced at least one security issue related to AI in the past year. Furthermore, 35% admitted that following AI recommendations would result in a security risk, while 26% reported that AI agents had connected to other systems without necessary approval.

Raghu Nandakumara, VP of Industry Strategy at Illumio, noted that the biggest risk may not be a system breaking rules, but one that was never given meaningful limits. He advised businesses to set strict boundaries on what AI agents can access and do, enforcing these limits within applications and infrastructure to contain unexpected behavior.

Sources :TechRadar

Recherche

Zendesk Executive Shares Internal AI Practices: 600 Employee-Built Apps and Sales Agent Case

Sujet · ZendeskPrise rapide
2026-10-30 00:23 GMT+8

OpenAI's official channel published a video on October 29, 2026, titled "ChatGPT Enterprise at Zendesk: Putting Company Data to Work." In the video, Daniel Humble, VP of Enterprise Data and Analytics at Zendesk, shares specific progress on the company's internal AI adoption.

According to the publisher's description, Zendesk currently has about 600 applications built by different people within the company. These apps span sales, marketing, and operations, designed to address specific work-related challenges. ChatGPT Enterprise is available to all employees, allowing business users to query company data directly.

In the sales department, a particular AI agent connects product usage and ticket data per agent to surface expansion opportunities and helps salespeople explain customer benefits. This application sees about 1,000 daily active users. For readers interested in how large enterprises integrate AI into daily workflows, this video offers an observation window based on actual deployment data.

Sources :youtube.com

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

On Booking.com's three features, production deployment results vary in focus. Smart Filters: The KD-plus-Medusa combination improved response speed by 15 times for 99% of queries. AI Trip Planner: User engagement increased by 2.9% compared to GPT-3.5. Accommodation Description Generation: Users gave 'helpful' ratings 1.4% higher, at less than one-third the cost of GPT-4. This work previously won the Best Paper Award in the Industry Track at the Association for Computational Linguistics (ACL).

It should be noted that all performance figures above are self-reported by the team, and the blog post does not detail the specific datasets, metric definitions, or statistical significance tests used for quality evaluation. The 2.9% increase in engagement may be influenced by other product changes occurring simultaneously, making it impossible to attribute it solely to the model replacement based on this source alone. Furthermore, this pipeline targets specific domain tasks (travel dialogue, structured filtering, description generation); whether its conclusions can transfer to open-domain generation or other task types requires independent verification.

For engineering teams deploying LLM applications, the core transferable insight from this blog post is: model compression and inference parallelization are not mutually exclusive choices but orthogonal dimensions that can be stacked; integrating Medusa heads into the single-stage distillation training process is more practical than staged processing; however, the magnitude of specific benefits depends on task characteristics, hardware conditions, and baseline selections, requiring re-measurement in one's own scenario.