LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-10-23129 posts

Atualizar

Anthropic Expands Cyber Verification Program to Three Tiers, Opening Restricted Claude Features to More Security Teams

Material
2026-10-07 20:05 GMT+8

On October 7, 2026, Anthropic announced an expansion of its Cyber Verification Program (CVP) to open powerful Claude model capabilities to a wider pool of security professionals with reduced safety restrictions. Previously, publicly available Claude models blocked most vulnerability research, malware analysis, incident response, and penetration testing because the same capabilities could also aid attackers. While basic tasks like code review or patching known flaws remained possible without special access, advanced cybersecurity tasks were strictly limited.

The new CVP splits access into three tiers. "Defense Access" is open to corporate security teams, government agencies, universities, critical infrastructure operators, open-source developers, and individual researchers, primarily for incident response and vulnerability validation. "Red Team Access" allows authorized attack simulations but is limited to organizations only. "Specialized Access" is reserved for testing systems like flight controls, power grids, or banking infrastructure; for this tier, Anthropic vets every applicant together with the US government.

Data from the predecessor program, Project Glasswing, provides initial evidence of effectiveness. According to Anthropic, partners found at least 129,000 confirmed vulnerabilities between April and July 2026, with more than 33,000 classified as high-severity or critical. However, these numbers come from surveys of a subset of partners, and Anthropic states the real-world impact is likely at least five times higher. Several partners reported that AI sped up their work by months or even years. While this structural change broadens defensive AI usage, actual adoption rates and independent verification of vulnerability discovery speed remain to be seen.

Fontes:The Decoder

Pesquisa

Hashed CEO Kim Seo-jun: Humans No Longer Execute; Agents Are the Workforce

Tópico · 人在回路上模式Material
2026-10-23 10:25 GMT+8

Kim Seo-jun is the CEO of Hashed, a South Korean crypto venture capital firm. At the 'Nitro Track Day' event held in Seoul, he put forward a clear organizational thesis: AI-native teams should not remain in the 'human-in-the-loop' (Human-in-the-loop) mode—where AI executes tasks and humans approve them step-by-step—but should instead transition to 'human-on-the-loop' (Human-on-the-loop). In this model, agents autonomously complete tasks and self-correct errors, while humans are responsible for designing work structures, setting intent, and taking responsibility for outcomes. He stated, 'Humans should no longer perform execution work; agents do the work, while humans design the structure, reflect intent, and take responsibility.'

Hashed operates internally according to this model. According to Kim Seo-jun, while employees were resting on the weekend, AI automatically submitted over 100 Pull Requests via Telegram. The development team for V, an Asian optimization agent service developed by Hashed using a venture-building approach, generated over 4,100 PRs in nearly one month. Within the team, individual and team-level agents communicate with each other to handle tasks such as schedule management and meeting invitation creation.

Lee Jae-hong, CEO of Across, an AI search optimization company, added a perspective on resource efficiency: not all tasks require high-performance models. 'You don't need to bring a cannon or nuclear weapon to solve something that can be handled with a pistol,' he said. The key is to identify the capability boundaries of each model and select them according to the scenario.

It should be noted that the above operational data are self-reported by Hashed and lack independent verification. The number of PRs does not equate to effective output quality, and the report does not explain the actual level and frequency of human review involvement. While 'human-on-the-loop' offers inspiration as an organizational design direction, its feasibility depends on the error tolerance boundaries of specific tasks and the actual reliability of agents, aspects which have not been elaborated upon in this report.

Fontes:바이라인네트워크

Pesquisa

Reuters: ByteDance, Alibaba and Tencent order $35 billion in Nvidia workstation chips, supply may meet only 5-10% of demand

Tópico · 英伟达AI生态垄断Material
2026-10-23 10:44 GMT+8

According to an exclusive report by Reuters on October 20, ByteDance, Alibaba, and Tencent have collectively ordered more than 3 million Nvidia RTX PRO 5500 workstation chips, with a total list price exceeding $35 billion. The figures come from two anonymous sources, and none of the three companies nor the U.S. Department of Commerce's Bureau of Industry and Security (BIS) responded to requests for comment at the time of reporting.

The report states that each RTX PRO 5500 is priced at approximately $11,400. Two other sources indicated that total orders from Chinese customers have exceeded 6 million units. Nvidia has informed clients that it cannot guarantee fulfilling these orders because Chinese demand far exceeds originally planned production volumes. Some partners may receive only 5% to 10% of their ordered quantities, while another source noted that Nvidia can produce a maximum of 3 million units globally in 2027.

The RTX PRO 5500 is a workstation chip with performance lower than data center-specific H200 and B300 models, though comparable to the discontinued RTX 6000D released in mid-2025. Sources say Chinese enterprises primarily intend to use it for AI inference rather than training. Under U.S. export control frameworks, B300s are banned from sale to Chinese buyers, and H200s require case-by-case review with licensing conditions, making the RTX PRO 5500 one of the few high-end Nvidia chips still accessible to Chinese buyers.

An Nvidia spokesperson stated that Chinese workstation chip businesses have seen "unprecedented growth" since controls were implemented in 2022, criticizing U.S. export controls as outdated and covering gaming products released nearly five years ago. Previously, Nvidia CEO Jensen Huang noted that restrictions had effectively reduced the company's share in China's AI data center chip market to zero, allowing local suppliers like Huawei to expand their market presence. All order numbers and delivery ratios remain unconfirmed claims attributed to anonymous sources cited by media.

Fontes:INSIDE

Pesquisa

Renesas Develops 800V-to-6V Direct Power System for AI Data Centers, Samples Provided to Major Manufacturers

Tópico · 瑞萨实体AI团队Material
2026-10-23 19:00 GMT+8

Japanese semiconductor company Renesas Electronics is developing a 20kW-class DC power system for AI data centers that directly converts 800V high-voltage DC to 6V low-voltage. According to EE Times Japan, Renesas has already provided samples of this platform to major AI data center manufacturers, including OEMs and ODMs. The company is currently using these samples to validate the power architecture and test foundational technologies.

As power density increases with GPUs and other computing devices, traditional 48V distribution schemes face challenges. Introducing 800V HVDC distribution reduces current and transmission losses, but requires stepping down voltage near the computing devices to levels suitable for chips. Renesas' new system aims to improve overall power supply efficiency and reduce module size by skipping intermediate voltage conversion stages like 48V or 12V. Its whitepaper, "Direct 800V-to-6V Conversion for AI Compute Infrastructure," indicates that the system uses Gallium Nitride (GaN) power devices and resonant LLC converters, targeting an efficiency of over 96% at 20kW output.

It is important to note that this efficiency figure is a prediction based on component loss calculations. Renesas stated that evaluation of actual hardware is still ongoing, and optimized measured data will be released once ready. The system was exhibited as part of the NVIDIA ecosystem partner area at the OCP Global Summit in October 2026.

Fontes:EE Times Japan

Pesquisa

Bitdefender Launches AI-Agent-Specific VPN: Isolates Connections with Disposable Containers, Free During Beta

Tópico · AI智能体安全Visão rápida
2026-10-23 09:39 GMT+8

Cybersecurity firm Bitdefender has recently launched a virtual private network (VPN) software specifically designed for AI agents. Unlike traditional always-on VPNs that cover entire devices, this new tool activates only when an AI agent initiates a network connection request and destroys the connection environment immediately after the task is completed.

According to Bitdefender, the service utilizes the mainstream Model Context Protocol (MCP), making it compatible with major AI development tools such as Claude Desktop and Cursor. The core mechanism relies on "disposable containers": whenever an AI agent processes a prompt and requires internet access, the software spins up an encrypted tunnel to route traffic through Bitdefender servers. Once execution finishes, the container—along with any cookies, cache, or session data—is discarded, ensuring no historical traces carry over to the next connection.

This design addresses growing privacy concerns surrounding AI agents. A Pew Research Center survey indicates that 71% of US adults worry that widespread AI use increases risks to their personal information security. The Cloud Security Alliance also categorizes AI agents as an "identity gray area" because they often borrow user credentials. Bitdefender notes that without proper isolation mechanisms, every web service accessed by an AI agent can be traced back to the user's real IP address.

However, the product currently faces significant limitations. It supports only macOS 13 or later on Apple Silicon chips, with no release date announced for Windows. Due to its reliance on container technology, system requirements for memory and disk space are higher than standard VPNs. Additionally, official marketing claims the tool helps AI agents "bypass geo-blocks," while terms of service state it "cannot bypass website IP blocks," presenting a contradictory description. The service is currently in beta, offered for free, and targets individual developers without team management features.

Fontes:TechOrange 科技報橘

Pesquisa

Fireworks AI DevRel Head Sophia Yang Demonstrates Glimmer Model Training and Performance on Kalshi Prediction Markets

Visão rápida
2026-10-23 06:03 GMT+8

Sophia Yang, Head of DevRel at Fireworks AI, released a video detailing how her team trained Glimmer, a 30B parameter open-source model, to perform better than Jev in Kalshi prediction markets. This content is part of Fireworks AI's "Free Weights Club," offering a recipe and model that can be fine-tuned for free.

According to the video description, the experiment used about 6,500 settled Kalshi events for training and tested on 8,765 later events. The technical approach involved K-way cross-entropy loss with a single LoRA adapter, followed by temperature tuning.

Fireworks AI claims that the tuned Glimmer model achieved an accuracy of 64.8%, compared to 56.4% for the base model and 45.9% for Jev. Additionally, calibration was reported at 86.8% when the model was at least 80% sure, with serving speeds 2–3x faster than standard chat completions.

It is important to note that these performance metrics are vendor-reported and based on paper trading rather than actual capital operations. The video serves primarily as a technical demonstration and resource recommendation for readers interested in AI applications in prediction markets.

Fontes:youtube.com

Pesquisa

Lumentum Senior Director Christian Urricariet Explains Proof-of-Concept for 1060 nm VCSEL Optical D2D Connectivity in AI Scale-Up Networks

Visão rápida
2026-10-23 02:32 GMT+8

Christian Urricariet, Senior Director of Product and Strategic Marketing at Lumentum, discusses the company's proof-of-concept configuration for high-density, 1060 nm VCSEL-based optical die-to-die (D2D) connectivity in a recent video. Developed in collaboration with ecosystem partners, the demonstration was showcased during ECOC 2026 in Málaga, Spain.

According to the video description, the configuration combines Lumentum's advanced-packaged VCSEL technology with D2D interface technology and high-density multimode fiber connectivity. Lumentum states that this approach demonstrates a potential path toward future co-packaged optics (CPO) and near-packaged optics (NPO) implementations as AI systems demand greater bandwidth density.

This resource is suitable for professional readers interested in the underlying optical interconnect technologies for AI infrastructure. It is important to note that the available information is limited to the introduction of a proof-of-concept demonstration; it does not include specific performance metrics, mass production timelines, or independent benchmark results.

Fontes:youtube.com

Pesquisa

iFlytek Launches Spark X2.6: Trained on Domestic Compute, Cloud Model Priced Below DeepSeek V4.1 Flash

Tópico · 科大讯飞星火X2.6Material
2026-10-22 21:35 GMT+8

On October 22, iFlytek unveiled several new achievements, including the end-cloud integrated large model Spark X2.6, the speech foundation model SparkAudio 1.0, and hardware products such as the iFlytek AI Recording Ring and AI Glasses. The release highlights capabilities upgraded using fully domestic computing power and an end-cloud collaborative architecture tailored for specific industry needs.

The Spark X2.6 model includes a cloud-based 293-A30B model and an edge-side X2.6-4B model, which has been open-sourced. The cloud model supports a million-token context window, with input prices at 0.9 yuan per million tokens and output prices at 3.6 yuan per million tokens. iFlytek claims these prices are lower than those of DeepSeek-V4.1-Flash-1017 for cache misses and outputs. According to official test results, the Spark X2.6-4B outperforms its predecessor Spark-X2.5-4B and competitors like Qwen3.5-4B and Gemma4-E4B in tasks such as information search, document processing, and schedule management.

Technically, iFlytek reported that on the 'Feixing No. 2' domestic supercomputing platform, training time for 2000B data with 256K length was reduced from 4.5 months to 1.2 months, improving ultra-long context training efficiency by 4x. In inference, a four-layer KV Cache architecture on domestic hardware increased the KV Cache hit rate from 63% to 95.7%, boosting inference efficiency by 4.5x compared to pre-optimization levels.

Strategically, iFlytek founder Liu Qingfeng stated that 60% of future tasks will run locally, 30% via cloud collaboration, and 10% independently in the cloud. This vision drives the end-cloud integrated design, aiming to balance performance, cost, and data security. For instance, Spark AStudio can process local emails and reports offline, initiating desensitized cloud collaboration requests only when necessary.

In terms of hardware, the iFlytek AI Recording Ring weighs 35g, supports WiFi 6 high-speed transmission, and is priced at 1,299 yuan. The iFlytek AI Glasses weigh 40g, support translation in over 120 languages, and start at 4,299 yuan. Additionally, iFlytek launched the Lingji N9000 AIPC laptop with up to 160 TOPS of local AI compute and the Spark Office All-in-One machine providing 2,240 TOPS, emphasizing private deployment capabilities.

It should be noted that all performance comparisons and price advantages are based on data and testing environments published by iFlytek, without third-party independent verification.

Fontes:zhidx.com

Pesquisa

Jellyfish Head of AI and Research Nicholas Arcolano: Individual Engineers Drive AI Coding Output More Than Tools

Material
2026-10-23 03:00 GMT+8

This episode of the AIE CODE podcast features Nicholas Arcolano, Head of AI and Research at Jellyfish, addressing a core question: In an era where AI-assisted programming is increasingly prevalent, what truly determines engineering team output? Arcolano cites data from approximately 50,000 engineers across 447 companies.

According to the episode description, Arcolano presents a "natural experiment" case: When Cursor pushed many of its users onto Grok, merged PRs barely moved. This phenomenon leads to his breakdown of driving factors: the tool itself explains under 1% of throughput variance; skill use accounts for about 3% (heavy skill users see roughly a 30% bump); token usage accounts for about 9%; company and team environment account for about 12%; and the individual engineer accounts for most of the rest.

The data also shows that engineers who were top performers before AI are mostly still the top performers now. Arcolano's advice for leaders is to focus on the people, skills, and environment you can control, rather than solely pursuing more powerful models. For professional readers seeking to understand how AI actually impacts software engineering efficiency, this episode offers specific quantitative perspectives and counter-intuitive conclusion cues.

Fontes:youtube.com

Pesquisa

Liquid AI Launches RIPL Engine, Claims Up to 28x Lower Latency Than vLLM for Small Models

Material
2026-10-23 01:33 GMT+8

Liquid AI announced the launch of RIPL (Real-time Inference Pipeline) on October 22, 2026. This low-latency engine is designed to serve Liquid Foundation Models (LFMs) within live request paths, targeting end-to-end response times under 50 milliseconds.

According to benchmarks published by Liquid AI, the LFM2.5-1.2B model running on RIPL demonstrated 14.4x to 28.4x lower p95 latency compared to the Qwen3.5-0.8B model running on vLLM, tested on a single NVIDIA B200 GPU. Specifically, in a workload with 256 input and 128 output tokens, LFM2.5-1.2B achieved a p95 latency of 155.8 ms versus 4,431 ms for Qwen3.5-0.8B. However, the advantage narrowed to 5.6x for the larger 2.6B model, which failed to meet the sub-50 ms target on longer requests.

Liquid AI describes RIPL as optimizing the entire path from request arrival to completed response, rather than relying solely on model architecture. The company presented comparisons using the same model (LFM2.5-350M) on both RIPL and vLLM, indicating that the engine itself contributes significantly to the speedup. RIPL is currently available only for enterprise deployments on NVIDIA H100, H200, B200, and RTX PRO 6000 GPUs.

The announcement also noted that Shopify is deploying LFMs in production for latency-sensitive tasks like search enhancement, with inference times below 20 ms. It is important to note that all performance figures are self-reported by Liquid AI and have not been independently reproduced.

Fontes:MarkTechPost

Pesquisa

Google Cloud Launches Agent Compute Engine: A Secure, Efficient Infrastructure for Hosting AI Agents

Tópico · 智能体沙箱隔离Material
2026-10-23 08:11 GMT+8

On October 22, 2026, Google Cloud, the cloud computing arm of Alphabet Inc., announced the launch of Agent Compute Engine, an infrastructure service specifically optimized for hosting artificial intelligence agents. This service aims to resolve the trade-off between security isolation and resource efficiency that currently plagues AI agent deployments in sandbox environments.

AI agents typically run in technical environments called sandboxes, which host the applications, data, and user instructions agents need to perform their work, along with cybersecurity guardrails to block malicious activity. Google refers to the sandbox implementation in Agent Compute Engine as 'microVMs.' The company states that this technology combines the hardware efficiency of containers with the security of traditional virtualization, notably bringing hypervisor capabilities to isolate VMs and prevent malware spread.

Agent Compute Engine runs microVMs on specially configured stateless appliances that do not permanently retain customer agent files; stored data resides in an immutable file system. Google verifies firmware integrity using custom Titan cybersecurity chips, and the CPUs feature mitigations against side-channel attacks. Each microVM has limited access to storage resources, meaning hackers cannot read or delete data even if they compromise a sandbox. Network traffic must pass through multiple proxies before reaching its destination, further enhancing isolation.

Many AI applications pause inactive agents to save resources, but traditionally, idle sandboxes still consume significant hardware, increasing costs. Google says Agent Compute Engine detects when a sandbox goes idle and shuts it down to prevent unnecessary usage, quickly reactivating it when needed. In a blog post, Google Cloud distinguished engineers Ramanathan Venkatapathy and Rohit Jnagal stated, 'This lets a single physical node host orders of magnitude more living agent microVMs than conventional VMs.'

Additionally, for agents using login credentials to access applications, Agent Compute Engine stores these credentials outside the sandbox. When an agent needs to log in, Google Cloud's Agent Gateway service retrieves and submits the relevant credentials on the agent's behalf, reducing the risk of sensitive data leakage. Google plans to roll out Agent Compute Engine in the coming months alongside a new version of Agent Gateway, which will gain model routing capabilities to process each agent request in the most efficient manner.

Fontes:SiliconANGLE

Pesquisa

Meta Board Approves Class C Non-Voting Stock to Let Zuckerberg Sell Shares for Proposed California Wealth Tax Without Losing Control

Tópico · Meta无投票权新股融资Material
2026-10-23 07:50 GMT+8

Meta’s board of directors has approved a new class of non-voting common stock, known as Class C shares, which will take effect in January. According to SEC filings and media reports, a key motivation for this structural change is to help founder and CEO Mark Zuckerberg address California’s proposed “billionaire tax” (Proposition 40). If passed by voters, the measure would impose a one-time 5% wealth tax on the net worth of several hundred California billionaires primarily residing in the state.

Holders of the new Class C shares possess an ownership stake but no vote on company decisions. Meta states it will use these shares for capital raising, employee compensation, and acquisitions—including its expensive push into artificial intelligence—while maintaining the current voting structure that favors Zuckerberg, who controls approximately 60% of the vote.

The New York Times reported that Zuckerberg pressed Meta’s board to devise a plan to protect him from the tax. SEC documents reveal that discussions began as early as July. The filing describes a mechanism where Meta issues one Class C share for every Class A or B share held, effectively acting as a one-time stock dividend equivalent to a two-for-one split. Crucially, Meta notes that the dividend “will not change the voting power of any shareholder.”

Specifically, the SEC filing states that the dividend “will allow Mr. Zuckerberg to immediately sell or transfer up to 20% of the shares of Class C common stock that he receives... representing 10% of his overall economic stake in the Company... without affecting his majority voting control.” This enables him to pay potential tax liabilities under Proposition 40 if enacted. Estimates suggest Zuckerberg could face a tax bill of up to $11 billion.

In its earnings press release, Meta reiterated that the Special Committee believed it was important for Zuckerberg to continue leading the company from California, citing the need for close coordination with management and proximity to Meta Superintelligence Labs. While Silicon Valley figures like Google co-founder Sergey Brin have actively campaigned against Prop 40, Meta’s strategy illustrates how major tech companies are developing creative mechanisms to navigate political and fiscal pressures without diluting founder control over corporate governance.

Fontes:TechCrunch

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento