LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-0159 publications

Actualiser

Researcher Finds Lenovo Fingerprint USB Unencrypted; Uses OpenAI Codex to Break Other Vendors' Protections

Sujet · 联想指纹U盘未加密争议Prise rapide
2026-11-01 21:00 GMT+8

Security researcher Xusheng Li disclosed in a March conference talk and subsequent blog posts that Lenovo's ThinkPlus FU100 fingerprint USB drive, marketed as having AES-256 hardware encryption, actually contained only a plain, unencrypted SD card internally. The fingerprint module merely enabled the USB connection, leaving data trivially accessible to anyone with physical possession of the device. Li noted that "hardware security" can be pure marketing.

In a recent follow-up study, Li used OpenAI's coding assistant tool, Codex, combined with reverse-engineering techniques, to automatically analyze other USB storage products claiming security protections. The results showed that Netac's tool was broken in about seven minutes because it contained a fixed key built into the program and wrote user passwords to diagnostic logs. KakaSoft's USB Copy Protection software allowed Codex to rebuild its internal password and decrypt a test file in around 26 minutes. Nexcopy's Copy Secure solution was compromised in about 10 minutes, with its key unlocked via a certificate embedded in the viewer app. Additionally, SecZure's YutianDun product was found to have two hidden "backdoor" accounts with simple passwords.

Regarding the Lenovo vulnerability, the company issued a high-severity advisory requiring users to contact their China service line for a firmware upgrade or replacement, strongly advising data backup and wiping before sending devices in. For other vendors, Li stated he did not warn KakaSoft, Nexcopy, or SecZure, arguing that drives already in an attacker's hands cannot easily be recalled or updated. He advised users to verify security claims themselves, stating that "client-side validation is never secure." As of his March talk, Netac had no patch available.

Sources :Tom's Hardware

Recherche

Open-source UEFI tool unlocks Nvidia's GPU AI limiter on Windows, boosting FP16 performance by up to 65%

Sujet · 英伟达GPU AI限速器Prise rapide
2026-11-01 22:00 GMT+8

A new open-source UEFI tool called nv-uefi-unlocker has been released, which can unlock the half-speed AI limiter on Nvidia RTX 20, 30, and 40-series desktop and laptop GPUs by booting from a USB drive. The tool runs before Windows loads the Nvidia driver, allowing the tensor units to run at full speed and boosting FP16 math performance by up to 65%.

Unlike previous methods, this tool does not require flashing the card's VBIOS or patching the driver. It uses the community V67 payload with read-back verification to trick Nvidia's signed Booter firmware into opening the write lock on two speed-selector registers. Because it must be run again on every boot and requires the system to be booted in UEFI mode with Secure Boot either off or set to trust a self-signed EFI, the tool cannot bypass Secure Boot requirements for games that use anti-cheat software.

Sources :Tom's Hardware

Recherche

OpenAI Tests Merging Chat and Work Tabs; ChatGPT and Codex to Share Usage Allowance

Sujet · ChatGPT购物Matériel
2026-11-01 12:14 GMT+8

OpenAI has sent emails to a small group of users announcing that it will begin testing a combined experience for the Chat and Work tabs in ChatGPT in the coming weeks. This change aims to allow users to move from exploring ideas to executing tasks in one place, while Codex remains a separate experience in the ChatGPT desktop app.

A more critical adjustment lies in the billing model: as part of this unified experience, users' usage of ChatGPT and Codex will draw from the same allowance. OpenAI stated it is preparing carefully for this transition and expects most users won't need to wait for their usage to reset more often than they do today. Specific usage will depend on the model and task, with more advanced models and complex tasks consuming more of the allowance.

Users will be able to see their combined remaining usage and reset times in Settings → Usage. If the full allowance is used, users can wait for it to reset, upgrade their plan, or buy credits to continue using eligible features. Unlimited Instant is included in the plan and does not draw from this usage.

Sources :V2EX

Recherche

TheSequence Editor: AI Models Improve Monthly, But Creditors Still Expect Payment on Tuesday

Matériel
2026-11-01 20:03 GMT+8

TheSequence Issue 963 editorial proposes a seemingly simple but easily overlooked judgment: building artificial intelligence requires two kinds of models—one that predicts the next token, and another that persuades a credit committee the GPUs will pay for themselves. Both were busy this week, and the author argues that the tension between them will not disappear due to progress by either side.

The author illustrates this point with three disclosures released simultaneously. Microsoft’s September quarter generated $58 billion in operating cash flow against $39.4 billion in capital expenditure, with Azure revenue growing 48% while repaying $4.9 billion in debt—a massive construction project attached to a massive cash machine. Alphabet’s situation was more dramatic: Google Cloud revenue grew 103%, but $51.4 billion in capital expenditure exceeded $42.3 billion in operating cash flow, leaving $9.1 billion in negative free cash flow, supplemented by approximately $29 billion in debt issuance. Meanwhile, SpaceX secured a binding commitment for approximately $40 billion in Apollo-managed GPU equipment finance leases covering NVIDIA Vera Rubin hardware, expected to close before year-end with equipment delivery beginning in the first half of 2027. SpaceX management stated that its superintelligence infrastructure investments are achieving payback periods of less than one year.

The author’s interpretation of this data is not about who is superior, but rather points out a structural issue: debt does help when revenues arrive later than equipment bills. However, the risk lies in lenders demanding predictable repayments while AI companies face uncertain utilization rates, hardware replacement cycles, and aggressive pricing. Microsoft’s existing business gives it room to finance the next project; Alphabet’s customers are indeed growing, but so are the invoices; SpaceX has obtained compute capacity alongside payment obligations. The author specifically notes that the claim of less-than-one-year payback is management’s assertion, which is a handy phrase to have when discussing a $40 billion transaction.

The author’s conclusion remains symmetrical: the bull case is that demand grows, useful products multiply, and financing brings productive infrastructure online faster; the bear case is that competition lowers compute prices, hardware depreciates, and fixed payments remain fixed. The editor’s own judgment is that the industry is making real technical progress while assuming real financing risk, and these two facts do not cancel each other out. Models can improve every month, but its lenders still expect the payment on Tuesday.

Sources :thesequence.substack.com

Recherche

The Information Reports Meta Collaborated with China's Fourier Intelligence on Humanoid Robot Prototype, Unconfirmed by Parties

Prise rapide
2026-11-01 16:16 GMT+8

Technology media outlet The Information reported on Saturday (October 31), citing people familiar with the matter, that US tech giant Meta has collaborated with Chinese startup Fourier Intelligence to develop a humanoid robot prototype and is currently working on a second-generation version. However, as of publication, neither Meta nor Fourier Intelligence has officially confirmed this news, and Reuters stated it could not immediately verify the report.

Earlier rumors circulating on social platform X claimed that Meta and Fourier Intelligence were co-developing a humanoid robot named "Max," weighing approximately 48 kilograms, with the hardware taking only three months from concept to completion. However, investigations found no formal information indicating that either company had released a robot named Max or jointly developed such products. Fourier Intelligence's publicly available product line does not include a model named Max; its existing GR-series robots weigh between 63 and 71 kilograms.

Founded in Shanghai in 2015, Fourier Intelligence initially focused on rehabilitation exoskeletons before launching the GR-1 humanoid robot in July 2023. The company's current public focus leans towards emotional companionship and elderly care applications. In contrast, Meta's moves in the robotics field are more clearly defined: it established a robotics research team within its Reality Labs division in 2025 and acquired Assured Robot Intelligence in May 2026 to strengthen its AI capabilities. While these actions demonstrate Meta's long-term interest in robotics, they do not yet prove a formal partnership with Fourier Intelligence.

Sources :INSIDE

Recherche

French National Assembly passes data traffic tax amendment, targeting tech giants with ~€0.055 per GB levy

Sujet · 法国按流量计费数字税Matériel
2026-11-01 00:11 GMT+8

According to a report by Tom's Hardware, the French National Assembly voted 11-10 on October 30, 2026, to approve an amendment introducing a new tax based on internet data traffic, adding it to the country's next budget. Proposed by Socialist lawmaker Laurent Lhardit, the measure would require content providers to pay approximately €0.055 (roughly $0.06) to the government for every gigabyte of content distributed to French users.

Unlike traditional digital service taxes that target advertising revenue or profits, this new levy directly targets entities generating substantial network traffic. The proposal explicitly names Google, Meta, Netflix, Amazon, and content delivery network Akamai as major potential contributors. Proponents estimate that with approximately 154 million users in France and around 200 billion GB of data passing through French internet providers annually—nearly half of which is attributed to five major companies—the tax could generate about €10 billion ($11.2 billion) in annual revenue.

To enforce the measure, telecommunications watchdog Arcep will oversee internet service providers like Orange Free and Liad, requiring them to measure data volumes and identify the companies responsible for sending it to French users. This data will be submitted quarterly to the French tax administration. The proposal also includes a cap: a company's total tax liability cannot exceed 15% of its French revenue or 2% of its worldwide revenue.

While the amendment has cleared the National Assembly, it still requires final approval from the French Senate to become law. Opponents have warned that such traffic monitoring could endanger user privacy and increase compliance burdens for businesses. However, Lhardit dismissed concerns that costs would be passed on to consumers, arguing that since platforms like Facebook, Instagram, and TikTok offer free access, tech companies cannot easily shift these infrastructure usage fees onto end-users.

Sources :Tom's Hardware

Recherche

Vals AI Study: Chinese LLMs Not Uniformly Pro-Government; Prompt Language Significantly Shifts Political Stance

Sujet · Vals AI智能体基准Matériel
2026-11-01 20:33 GMT+8

A new study by AI benchmark firm Vals AI indicates that Chinese artificial intelligence models behave in a more complex manner regarding politically sensitive questions than previously assumed, challenging the notion that they act merely as "government mouthpieces." Researchers posed hundreds of questions about China to various language models, including those from the US and China, while varying the prompt language, usage context, and response format.

The findings show that the language of the prompt has a strong effect on the political slant of responses. For most tested models, answers generated in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. This effect was present in both Chinese and American models but was far more pronounced in the Chinese ones. Vals AI also identified a smaller but consistent gap between Simplified Chinese (used in mainland China) and Traditional Chinese (common in Taiwan and Hong Kong).

Furthermore, system prompt context altered model behavior. When the system prompt indicated the model was being used internally by a company in mainland China, some models responded in a more pro-government way compared to when the prompt suggested US-based usage. Notably, Western models occasionally adopted official Chinese government positions, while certain Chinese models, such as GLM-5.3 Flash, explicitly acknowledged documented human rights violations when asked.

The study notes that the common assumption—that Chinese models primarily block sensitive questions—only held true for certain DeepSeek models. Most other Chinese models rarely refused to answer, instead sometimes adopting official government stances. For instance, Kimi K3 denied the existence of concentration camps in Xinjiang, referring to them as "vocational training centers" designed to combat extremism.

Sources :The Decoder

Recherche

Google Rolls Out Gemini AI Features in Google Maps for Apple CarPlay Users

Prise rapide
2026-11-01 14:15 GMT+8

According to German tech media outlet t3n.de, Google has started rolling out its Gemini artificial intelligence assistant within the Google Maps application for iPhone users utilizing Apple CarPlay. While Android Auto users have long had access to similar features, this update marks Google's entry of AI capabilities into the closed Apple automotive ecosystem.

The functionality is embedded directly within the Google Maps navigation interface rather than operating as a standalone app. Drivers can activate Gemini by tapping a star icon on the right side of the screen, allowing for multi-step natural language interactions. For instance, users can ask for scenic viewpoints nearby, check if a specific location is currently crowded to find quieter times, or verbally request toll-free routes when entering a destination, bypassing manual settings adjustments. Additionally, drivers can report traffic jams or accidents directly to Gemini while driving, providing real-time warnings to other map users.

Per 9to5Google, this rollout begins in the United States and is expected to expand globally over the coming weeks. Users who do not yet see the Gemini icon should await further updates.

Sources :t3n.de

Recherche

Google Team Introduces StreamVibe Video Agent; Tests Show Response Quality Drops Up to 61.5% in Continuous Long-Video Operation

Sujet · StreamVibeMatériel
2026-11-01 19:25 GMT+8

A research team comprising researchers from Google, Stanford University, Oxford University, and other institutions has introduced StreamVibe, a video agent designed to understand vague user intentions and decide when to offer help. According to The Decoder, the team also released ProactiveBench, a benchmark containing approximately 4,100 examples covering clear instructions, vague goals, and requests that change over time.

The tests show that while StreamVibe-VL (reportedly trained based on Alibaba's Qwen3.5-9B) improved response timing and quality by 10.7 percent over baseline models, all tested systems performed significantly worse when processing continuous videos longer than five minutes. As video length increased, models struggled to correctly time their responses, and key content metrics dropped by up to 61.5 percent due to compounding errors.

Researchers noted that existing video assistants are typically built for clear-cut commands, making it difficult to pick up on unspoken intentions or track completed steps, which leads to repeated advice or delayed responses. StreamVibe attempts to address this by distinguishing between explicit, implicit, and changing user intentions, but its core finding reveals a fundamental limitation in current AI for continuous interaction: performance plummets as errors compound when models must build upon their own earlier responses.

Sources :The Decoder

Recherche

SenseTime Open-Sources SenseNova Loop-9B-Preview: Exploring Small Model Performance via Looped Transformer Architecture

Prise rapide
2026-10-31 10:51 GMT+8

SenseTime officially open-sourced the SenseNova Loop-9B-Preview model from its Riri Xin series on October 31, 2026. This large language model employs an innovative Looped Transformer architecture designed to extend the effective computational depth of the model through cyclic computation, aiming to achieve or surpass the performance of traditional 27B parameter models with only 9B parameters.

The core design retains the first and last four layers of the original decoder modules while transforming the intermediate layers into parameter-sharing cyclic computation modules that increase information interaction depth through multiple iterations. To maintain signal propagation stability and enhance feature interaction, SenseTime introduced the mHC (Manifold-Constrained Hyper-Connections) mechanism, utilizing multi-path residual connections for constrained information mixing. This architecture attempts to address the issue where current LLM capability improvements rely excessively on scaling parameter counts, exploring a new path to release model potential while controlling parameter size.

According to internal evaluation results published by SenseTime, SenseNova Loop-9B-Preview demonstrates strong performance across code generation, office scenarios, and agent tasks. Specifically, the model claims to outperform Qwen3.5-27B in benchmarks such as WorkBuddy Bench, RaccoonPPT Bench, and AutomationBench. Ablation studies indicate that introducing the mHC mechanism improved scores on GPQA Diamond and TeamBench by 4.42 and 13.34 points respectively. It is important to note that these are vendor-reported results and have not yet been independently reproduced by third parties.

Currently, research community explorations of Looped Transformers remain largely academic, lacking direct open-source practices for daily use. The release of SenseNova Loop-9B-Preview fills this gap, with weights uploaded to GitHub and Hugging Face. SenseTime stated that future work will continue to explore scaling potential at larger sizes and adaptive 'fast/slow thinking' mechanisms to further improve inference efficiency.

Sources :leiphone.com

Recherche

Anthropic’s Claude Models Reportedly Block Game Cloning and Reverse-Engineering Tasks

Sujet · AnthropicMatériel
2026-10-31 20:20 GMT+8

According to multiple user reports compiled by Tom's Hardware, Anthropic's Claude Code and Opus 5.5 models have begun blocking attempts at video game recreation and reverse-engineering. The restrictions appear linked to an updated internal rule allegedly instructing the model to avoid "[recreating] a specific existing game," even if renamed or called original. This move seems designed to limit the flood of AI-generated clones of popular titles and application suites.

While the intent may be to prevent copyright infringement, legitimate developers report that the safeguards are also triggering refusals for non-malicious tasks. Complaints on GitHub indicate that Claude has refused to assist with standard game development halfway through projects, replicating specific mechanics, or patching tabletop-style games. Additionally, users claim the model blocks generic reverse-engineering tasks such as analyzing Windows crash dumps, clearing junkware on smart TVs, or using the popular Reverse Engineer Anything (REA) tool.

The timing coincides with Anthropic's introduction of its Cyber Verification Program on October 6, which limits access to advanced cybersecurity capabilities. Some purported game developers state that Claude blocked their own games' anti-cheat development and security reviews under the justification that safeguards were triggered. There is no official statement from Anthropic confirming these specific rules or the extent of the blocking.

The reaction among users has been swift, with many suggesting a shift to Chinese, open-weight, or local LLMs for reverse-engineering work. This highlights a growing tension: frontier AI companies trained on vast amounts of public data are now implementing guardrails to protect against copyright claims, while users seek control over devices and software they purchase. Scientific American notes that tools like REA allow AI agents to orchestrate decompilation programs, making reverse engineering accessible to ordinary users, a capability that legal frameworks struggle to contain.

Sources :Tom's Hardware

Recherche

Samsung Chairman Lee Jae-yong Visits Yokohama Advanced Packaging Lab, Prioritizing AI Chip Competitiveness

Sujet · 三星横滨先进封装实验室Matériel
2026-11-01 15:10 GMT+8

Samsung Electronics Chairman Lee Jae-yong visited the "Advanced Packaging Lab" (APL) in Yokohama, Japan, on November 1. This marked the first high-level on-site inspection since the lab officially opened on September 8, according to a report by South Korea's Electronic Times. Lee reviewed the progress of next-generation packaging element technology research with scientists and discussed cooperation plans with Japan's materials, parts, and equipment (Sobuja) industry.

Established in 2023, the APL is dedicated to developing and verifying the materials and processes required for advanced semiconductor packaging. Samsung plans to invest 40 billion yen (approximately 350 billion won) over five years starting from 2024 to solidify its local research base. The lab's core role is to connect Samsung's design and manufacturing capabilities with Japan's robust semiconductor supply chain ecosystem.

As competition in AI chips intensifies, advanced packaging has become a critical bottleneck determining the performance of GPUs, High Bandwidth Memory (HBM), and accelerators. Packaging technology directly impacts AI system bandwidth, power efficiency, and thermal performance. During the visit, Lee emphasized that securing advanced packaging technology is central to semiconductor competitiveness in the AI era.

Current R&D priorities at the APL include next-generation processes such as wafer bonding and glass substrates. Glass substrates are considered vital for future packaging due to their higher flatness and thermal stability compared to traditional organic substrates. Additionally, the lab aims to develop new packaging architectures suitable for mass-producing low-power, high-performance AI chips. Samsung plans to cultivate the Yokohama APL as a global hub for advanced packaging R&D, coordinating with domestic semiconductor research institutes.

Sources :미래를 보는 창 - 전자신문

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement