LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-0499 posts

Atualizar

Stanford Study: 44% of Widely Cited LLM Claims Expired, Average Lifespan Just 15.7 Months

Tópico · 斯坦福LLM论断寿命研究Material
2026-11-04 03:41 GMT+8

According to a report by technology media outlet The Decoder, Stanford University researchers Owen Queen and James Zou tested 50 widely cited claims about large language models (LLMs) against 20 current mainstream models. The results showed that 44% of the claims had already "expired," while another 34% had significantly weakened. Only 16% still held true. Among the 22 findings with an estimated expiration date, the average lifespan was just 15.7 months.

The study found that claims about what models "can't do" expire the fastest. Previously widely discussed weaknesses, such as counting letters in "strawberry," arithmetic in unfamiliar number systems, and losing information in the middle of long contexts, have largely disappeared in current models. In contrast, claims involving model preferences and behavioral tendencies prove more durable. Features like sycophancy, political bias, and self-preference when acting as evaluators remain significant across multiple models.

The research emphasizes that as AI development accelerates, empirical findings about LLMs should be presented as observations about specific models at a specific point in time, rather than as general truths. This conclusion reminds readers and researchers that many classic conclusions about AI capabilities from the past may no longer apply to current technical realities.

Fontes:The Decoder

Pesquisa

AI Infrastructure Startup Applied Compute Raises $350M at $3.25B Valuation

Material
2026-11-04 09:19 GMT+8

US-based AI infrastructure startup Applied Compute Inc. announced on November 3, 2026, that it has raised $350 million in funding at a $3.25 billion valuation. The round was led by prominent angel investor Elad Gil, with participation from Nvidia Corp., Shopify Ventures, HubSpot Ventures, and over a dozen others.

Applied Compute operates a cloud platform used by developers to train large language models and perform inference. The company stated that its platform uses an efficient training approach called asynchronous reinforcement learning (RL), which allows data generation and weight adjustment components to run simultaneously, speeding up LLM development. Additionally, the platform supports self-distillation, removing the need for teacher models.

Regarding the use of funds, Applied Compute disclosed that its annualized recurring revenue (ARR) has topped $100 million. The company, which launched less than 18 months ago, will use the newly raised capital to grow its market presence and buy more infrastructure. Its customers include major tech firms such as Microsoft Corp. and DoorDash Inc.

Fontes:SiliconANGLE

Pesquisa

OpenAI says ChatGPT handles 20% of Google's search volume; ads haven't hurt usage

Tópico · OpenAIMaterial
2026-11-04 02:30 GMT+8

US artificial intelligence company OpenAI announced at a press event on Monday that the volume of search queries run through ChatGPT now equals 20 percent of the volume run through Google. Adam Fry, OpenAI’s consumer product lead for ChatGPT, stated that the company defines a search as any time a user inputs a query that results in a web search, as opposed to tasks like defining a term or a simple how-to explanation. Fry did not specify the source of the Google search volume number.

While having one-fifth as many searches as Google is still far from overtaking it, Microsoft CEO Satya Nadella testified during Google’s 2023 antitrust trial that Bing had about 3 percent of the search engine market share after trying for more than a decade. During that trial, Google described AI as a potential disruptor, arguing ChatGPT was poised to seriously compete with it — a view OpenAI seems to now validate.

As search queries increasingly route through OpenAI and other AI labs’ services, the company is advancing its monetization path. Fry told reporters that ads have not negatively impacted the product or its usage, adding that “most people do not seem to mind ads in the product.” Last month, OpenAI announced it would begin introducing images of sponsored products and services in ChatGPT rather than just text links. This move is seen as a key step in exploring profitability through advertising as OpenAI approaches an IPO potentially slated for early next year.

Fontes:The Verge

Pesquisa

Belgian Cybersecurity Firm Aikido Acquires Two Startups to Launch Employee AI Security Platform Noon

Visão rápida
2026-11-04 08:00 GMT+8

Belgian cybersecurity company Aikido Security announced on October 20 the launch of Noon, a new employee security platform. This move follows its acquisitions of Polish firm Cyberly and Belgian firm OutKept in September, integrating their respective technologies.

According to venture data platform Dealroom, Aikido completed a $60 million Series B funding round in January 2026, reaching a valuation of $1 billion. The launch marks an expansion of its business from application, code, and cloud security into enterprise workforce protection.

The acquired companies brought distinct capabilities: Cyberly focused on AI attack simulation and risk management for employees using AI tools, while OutKept specialized in phishing simulation and security awareness training. These technologies form the foundation of Noon’s AI tool management and social engineering defense features.

In terms of AI governance, Noon allows enterprises to monitor which AI models, Model Context Protocol (MCP) servers, and other AI tools employees are using. Companies can set policies to restrict sensitive data entry into AI prompts, mitigating leakage risks. The platform also detects malicious emails missed by traditional filters and offers simulations of attacks involving deepfakes.

Fontes:iThome

Pesquisa

AI Data Centers Turn to Retired Jet Engines to Bypass Grid Delays

Material
2026-11-04 06:19 GMT+8

Some AI data centers are generating their own electricity using converted jet engines pulled from retired aircraft, according to a New York Times report. These engines, rebuilt to burn natural gas, are not intended as permanent power plants but highlight how far power supply has fallen behind the AI buildout.

For companies stuck in long queues for grid connections, the main draw of converted jet engines is speed. "It's all about getting power as fast as possible," said Rystad Energy analyst Audun Martinsen. He noted that with jet-powered units, "they can have a data center running in 24 months." By comparison, hooking up 100 megawatts to the grid in Northern Virginia takes about seven years, according to the International Energy Agency. Most new power capacity is expected to come online by 2031, an executive at Mitsubishi Power told Bloomberg last week.

These delays are already costing some operators business. Last month, Microsoft withdrew as the anchor customer for Nscale’s data center in Loughton, England, after the local utility stated it could not deliver sufficient power until the early to mid-2030s. Elon Musk’s xAI is among the companies already using these engines, having installed eight converted jet-engine turbines in Mississippi to help power its Colossus 2 data center in Memphis, citing state permits and satellite imagery. The units come from ProEnergy, which rebuilds GE engines commonly found on Boeing jets.

In 2025, OpenAI followed Musk’s lead with gas turbines at its first Stargate site in Abilene, Texas. Crusoe, a Denver-based startup building AI data centers, is constructing that Abilene campus and ordered 13 of ProEnergy’s converted units in April. In September, Crusoe walked away from a $1.25 billion deal to buy 29 natural gas turbines from Boom Supersonic, a startup developing a supersonic passenger jet that also sells a version of its engine as a power plant for data centers. A Crusoe spokesperson told TechCrunch at the time that the company still plans to use turbines, just not Boom’s.

Currently, the Trump administration is adding pressure by encouraging data centers to generate their own electricity. However, converted jet engines come with trade-offs: they are less efficient and pollute more than turbines designed for conventional gas plants, requiring more units to produce the same amount of power. This may ultimately make them a short-term play, as experts suggest efficiency and environmental concerns could grow in importance within a year or two.

Fontes:TechCrunch

Pesquisa

Collinear AI Releases CWE-bench v2: Covers 91 Vulnerability Classes, Gemini 4 Argon Leads with 67.8% Pass Rate

Tópico · CWE-bench v2Material
2026-11-03 08:00 GMT+8

Collinear AI released CWE-bench v2 on November 3, 2026, a benchmark designed to evaluate the ability of artificial intelligence agents to discover and fix security vulnerabilities in open-source software. The new version opens up 100 public tasks covering 91 different Common Weakness Enumerations (CWEs), spanning a wide range of vulnerability classes. The benchmark requires agents to locate and fix security flaws in a checkout of a real open-source repository given a short instruction, without breaking existing functionality. Scoring is fully programmatic, ensuring deterministic and reproducible results.

According to preliminary results published by Collinear AI, among the 11 models evaluated, Google's Gemini 4 Argon leads with a 67.8% pass@1 accuracy and 81.0% pass@4, slightly ahead of Anthropic's Claude Opus 5.5 (67.3% pass@1). Although 95 tasks were solved by at least one model, proving the tasks are solvable, even the best-performing model failed to solve approximately 19% of the task set on every attempt, showing significant gaps in long-tail vulnerability repair capabilities. In terms of cost efficiency, DeepSeek V4.1 Flash, Claude Opus 5.5, and Gemini 4 Argon form the Pareto frontier, with DeepSeek reaching 57.0% pass@1 at $0.04 per rollout, while Gemini edges ahead at $3.92.

The benchmark dataset and verifiers are open-sourced on Hugging Face and support direct execution via the Harbor framework. All programmatic checks have been calibrated by LLM judge panels to ensure hardness and fairness. Additionally, Collinear AI maintains a held-out set to track progress on tasks no model has seen. This release provides developers and security researchers with a standardized tool to quantitatively assess the actual capability boundaries of frontier models in cybersecurity defense.

Fontes:huggingface.co

Pesquisa

Airbnb Production Benchmark: Fine-Tuning Small Models to Reach Frontier Classification Quality at ~1% Cost

Tópico · Airbnb微调基准Material
2026-11-04 06:31 GMT+8

In November 2026, the Airbnb engineering team released a production benchmark for classification tasks. The core finding was that fine-tuning the Gemma4-E2B model (with 2 billion effective parameters) on hundreds of thousands of domain-specific labeled examples raised its F1 score from 0.563 (unfine-tuned) to 0.902, with a median latency of only 35ms. This performance matched frontier models such as GPT-6.1 Sol and Claude Sonnet 5.5 in quality, while costing roughly one-hundredth of those proprietary alternatives. The test was based on a held-out evaluation set of 10,000 weighted samples assessing the detection of violations of Airbnb’s pricing and payment transparency policies—a real-world task where the model is called thousands of times per minute and must return a score within 200ms.

The authors summarize this result by stating:

"Fine-tuning moves the quality-latency frontier, rather than trading along it."

This means fine-tuning does not involve exchanging quality for speed or speed for quality; instead, it improves both simultaneously. In contrast, even when enabled in low-thinking mode, frontier generative models still exhibit a median latency of about 900ms and a p99 latency of several seconds. While open-weight models are faster, their quality is significantly lower than the fine-tuned results. TypeSafe’s Jev performed best among off-the-shelf models under 100ms but still fell short of the fine-tuned E2B. A detailed prompt (2,700 words) could improve proprietary models by one to three percentage points, yet the gap in latency-constrained models remained significant, and further increasing costs yielded diminishing returns.

The article provides four decision criteria: self-hosted GPU costs fall below frontier API costs when traffic exceeds approximately 30 RPM; when latency budgets are tight, the fine-tuned model’s p99 latency is about 85ms compared to the frontier API’s median of 900ms; when adjustable thresholds are needed rather than fixed verdicts, score outputs allow adjusting the precision-recall tradeoff without rewriting prompts; and when labeled data is available, examples can teach subtle distinctions that instructions struggle to enumerate. The authors acknowledge that low-frequency or weekly-changing tasks remain better suited for frontier models paired with good prompting. This is a first-party reported benchmark from Airbnb, using public list prices and a 71% cache hit rate (higher than actual observed values, favoring hosted models), and has not been independently replicated by a third party.

Fontes:airbnb.tech

Pesquisa

AAIF Executive Director Mazin Gilbert: Standards Bodies Must Move Faster Than Members or Become Irrelevant

Tópico · 智能体AI基金会Material
2026-11-04 06:50 GMT+8

Mazin Gilbert left Google in 2026 to join the newly formed Agentic AI Foundation (AAIF) under the Linux Foundation as executive director. During an interview with The New Stack podcast at the AGNTCon + MCPCon conference in San Jose, he made a core judgment: if the foundation does not move at the speed of light, it will become irrelevant.

The source of this speed pressure is specific. Gilbert says that companies like OpenAI, Anthropic, and Google will build products with standards and also build products without them. Standards bodies are not writing specifications in a vacuum; they are racing against time. Once a company has already deployed its own integration code, a standard that arrives later leaves customers with legacy code that they must maintain themselves.

So what is the mechanism that gets fiercely competing companies to sit down and co-develop protocols? Gilbert’s answer points to infrastructure interdependence: these protocols are the air you breathe, the plumbing. They may not be sexy, but if we don’t get them right, the entire agentic AI space will become a mess, expensive, and basically prohibitive for any company to deploy at scale. Before our recording, he had just met with a group of maintainers from very fiercely competing companies who were working on the protocols together. All of those companies depend on these standards, so they have a reason to get the work done.

Adoption data for MCP seems to support this mechanism. Gilbert notes that MCP downloads grew tenfold in the first 10 months after Anthropic donated the protocol to the foundation last December. Today, fewer than 15 percent of MCP maintainers work at Anthropic, and more than 1,000 organizations contribute 75 percent of pull requests. The July update to the MCP specification made the protocol stateless. Gilbert says every company not only contributed to it but completely updated their entire tech stack to adapt to stateless MCP.

The AAIF currently has 305 members. Gilbert says about one new company joins every business day, calling it the fastest-growing foundation in the Linux Foundation’s history. Forty percent of members are startups, and 19 percent are nonprofits, national labs, and academic institutions. Large companies tend to contribute the protocols themselves, while smaller companies build products on top of them—this is the opposite of the CNCF model, where many projects started at startups.

But Gilbert also admits that not all members are contributing. Some companies are still trying to understand what agents can do for them, and others don’t know how to contribute. He says a third group of companies frankly are panicking because they don’t yet have an agent strategy. In September, the foundation launched its first certification—the Model Context Protocol Associate (MCPA)—to help these companies get started.

An unresolved issue is gateway fragmentation. This year, the AAIF approved two gateway projects: Solo.io’s agentgateway and Agent Router, which migrated from the CNCF Envoy project in September. Gilbert says five or six more companies want to contribute their own gateways. Does the community need six or ten gateways, or do we need to harmonize them together? He doesn’t have an answer, and says the community doesn’t either. The foundation provides a neutral governance space for these companies to work it out together.

Fontes:The New Stack

Pesquisa

US-listed e-commerce giant Coupang Inc. reports Q3 loss: Fire and data breach drive third consecutive quarterly deficit despite revenue recovery

Material
2026-11-04 06:06 GMT+8

Coupang Inc., the US-listed South Korean e-commerce giant, filed its third-quarter financial report with the US Securities and Exchange Commission (SEC) on November 3 local time. The company reported an operating loss of $450 million (approximately 642.3 billion won), swinging from a $162 million profit in the same period last year. This marks the third consecutive quarter of operating losses following the massive personal data breach incident at the end of 2025.

The current quarter's deficit was heavily impacted by one-time non-recurring costs. In addition to the penalty accrued in Q2 from the Personal Information Protection Commission, Q3 reflected approximately $248 million in losses related to a fire at the Incheon logistics center in July. Excluding fire-related expenses, the adjusted operating loss stood at roughly $202 million. Year-to-date, Coupang's cumulative operating losses have reached 1.83 trillion won, effectively erasing the 1.9 trillion won in total profits accumulated between 2023 and 2025.

Despite the significant book losses, core business indicators showed signs of recovery. Total revenue for Q3 was $9.962 billion, up 7% year-over-year (11% on a constant currency basis). The Product Commerce segment, including Rocket Delivery and Marketplace, saw revenue grow by 5%, while active customers rose 2% to 25.2 million. Emerging businesses such as Coupang Play and Coupang Eats grew revenue by 23%, with adjusted EBITDA losses narrowing. Additionally, the company announced a share repurchase of $201 million worth of Class A common stock during the quarter.

Fontes:바이라인네트워크

Pesquisa

Google and Mercor Launch Physical Engineering Benchmark; All Frontier Models Score Below 50%

Material
2026-11-03

On November 4, 2026, the Google AI & Infrastructure team announced the launch of Physical-Engineering's Last Exam, a new benchmark designed to evaluate AI agents' ability to complete tasks in mechanical, electrical, civil, and structural engineering. Developed in collaboration with talent platform Mercor, the initiative includes the release of 29 public tasks, grading tools, and an open competition leaderboard at ple-bench.com.

Existing academic benchmarks often focus on textbook-style questions or single tool calls, failing to reflect how AI agents perform in real-world engineering environments where they must adhere to building codes, manage costs, and produce deliverables subject to licensed engineer review. Physical-Engineering's Last Exam addresses this gap with 85 long-horizon tasks built from realistic project folders containing an average of 61 files, including PDFs, spreadsheets, emails, 2D CAD drawings, and 3D BIM models. Agents must reason through these materials, identify governing standards, run calculations, and generate production-ready outputs within a containerized environment.

The Google team tested 14 frontier models using seven publicly available agent harnesses, such as Antigravity, Claude Code, and Codex. The results revealed that all models scored below 50% on the private test set, with the highest-performing model reaching only 48.3%. Furthermore, while the top configuration satisfied 91% of individual constraints, only 41.3% of its deliverables passed all hard safety and feasibility checks. These findings underscore a substantial gap between current AI capabilities and the demands of complex, high-stakes physical engineering work.

Fontes:Google Cloud Blog

Pesquisa

Speech Recognition Firm Soniox Releases STT v6 with Agent Text Input and Echo Filtering

Visão rápida
Verificado 2026-11-04 06:31 GMT+8

US-based speech recognition company Soniox released Soniox STT v6 on November 3, 2026. This latest generation of its speech-to-text API is designed specifically for voice agents, introducing two key innovations: "Agent Text Input" and "Agent Voice Echo Cancellation."

In traditional cascaded voice agent architectures, the STT model typically hears only the user's audio, lacking context about what the agent just said. This often leads to errors in recognizing short, context-dependent replies like names or numbers. Soniox STT v6 allows developers to stream the agent's outgoing text directly into the STT model, giving it awareness of both sides of the conversation to more accurately transcribe user responses.

Additionally, v6 uses this text context to distinguish between the agent's playback and the user's speech, automatically filtering out leaked agent audio. This mechanism aims to prevent voice agents from interrupting themselves when they hear their own output. However, Soniox documentation notes that this feature does not replace traditional Acoustic Echo Cancellation (AEC) but rather handles residuals left behind by AEC, recommending that both be used together.

Pricing for Soniox STT v6 is approximately $0.12 per hour for real-time streaming and $0.10 per hour for asynchronous file transcription, with speaker diarization and language identification included. The previous v5 models will be deprecated on December 15, 2026.

It is important to note that while Soniox claims v6 is "significantly" more accurate than v5, no independent third-party benchmark results have been published yet. The company's website still displays v5 data from the open-source Pipecat STT benchmark (1.34% semantic WER), leaving v6's actual performance pending developer validation.

Fontes:Soniox | Speech-to-Text AI

Pesquisa

Meta F3 Lightning: After Decoupling Release and Execution, Why CPU Usage Remains High

Material
2026-11-04 01:00 GMT+8

Meta’s ad ranking models require extracting more than 20,000 features per request. These features are not computationally independent: a seven-day click count and a seven-day impression count may read from the same event source and apply the same time-window filter, while a click-through rate feature reuses the results of both. Facebook Feature Framework (F3) compiles all features into a single directed acyclic graph (DAG), merging duplicate nodes so that shared computation runs only once. This drastically reduces execution costs, but the trade-off is that the merged graph also serves as the release unit. If one feature definition contains an error, the entire graph cannot go live, forcing all other features to wait.

As the number of features grew from a few hundred initially to 20,000, this problem became increasingly severe. The F3 team implemented several rounds of improvements to restore release speed, but each gain was eroded by newly deployed features because all features still shipped together in the merged graph. The goal of F3 Lightning was clear: enable independent release for each feature without sacrificing F3’s computational efficiency.

The most direct approach was to decouple only the release process. Each feature would compile into its own independent configuration, with separate validation, deployment, and rollback capabilities. However, early prototypes revealed unexpected costs: when independent configurations arrived at the serving process and executed separately, the repeated reads, filters, and aggregations eliminated by F3’s merged graph returned, causing significant CPU spikes. The issue was not release granularity but execution granularity—decoupling releases simultaneously decoupled shared computation.

The core idea behind F3 Lightning is to bring back shared execution without bringing back the merged graph. Specifically, at service startup, it consolidates the shared work across all currently active feature configurations into a single execution order, rather than merging all features into one graph at compile time. This is achieved in four steps.

Step one is stable identifiers. To recognize that two independently compiled configurations contain the same computation, the runtime must be able to match them. Node numbering fails here because numbers are local to each feature’s graph; the same seven-day filter might be node 2 in one configuration and node 7 in another, and inserting a single node causes all subsequent nodes to be renumbered. F3 Lightning has the compiler derive identifiers from the computation itself: the operation type, parameters, and identifiers of all inputs. Because input identifiers already reflect all upstream work, this identifier captures the entire chain of computation that produced the result. Features defining the same operation on the same inputs receive the same identifier, even if compiled and released weeks apart.

Step two is a single execution plan. At service startup, F3 Lightning performs a topological sort across all active features, placing every node after its dependencies. When multiple features contain nodes with the same identifier, the plan retains only one copy, with all consumers pointing to it. This ensures each computation runs only once. However, the plan remains a graph structure requiring pointer-by-pointer traversal by the executor. A flattening pass during startup converts it into an ordered list, turning each request into a single forward scan along the list instead of traversing a graph.

Fontes:Engineering at Meta

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

Gilbert is frank about the fact that none of us have done this before. He says we are learning as we go. This is both an honest statement of limitation and the fundamental uncertainty facing standard governance in this field: no one knows whether a protocol foundation jointly maintained by competing companies can truly keep up with the iteration speed of their respective products.

At this stage, computations no longer repeat, yet CPU usage remains higher than in traditional F3. The team identified that the remaining cost stems from memory layout: each configuration still carries its own request-time data container, scattering intermediate results across separate memory allocations. Executing the plan sequentially means jumping between different containers, with each jump incurring a cache miss.

Step three is shared memory buffering. F3 Lightning replaces per-feature independent containers with a single buffer per request. At startup, references for each feature are rebound to slots within that buffer. This yields three improvements: adjacent plan steps now touch adjacent memory, restoring the sequential access pattern the merged graph always possessed; a single readiness vector replaces per-feature state tracking, allowing the executor to check if a value is ready without reading the value itself; and consumers no longer receive their own copies—the shared slot holds a reference to the producer’s result, so a computation used by 20 features is stored only once. Scalar values are still copied because copying small values is cheaper than following pointers.

Step four is dense index mapping. Although values are now centralized, locating them remains expensive: every access requires resolving an identifier—hashing a key, probing buckets, and following pointers. At Meta’s serving volume, such lookups are inefficient. The final fix maps identifiers to consecutive integer slots at initialization and rewrites all references in every configuration to point to these slots. Reading a result becomes a direct array index into the shared buffer, eliminating hashing and lookup tables from the request path.

After completing these four steps, F3 Lightning achieves computational efficiency comparable to F3 while allowing each feature to be released independently. The team reports that extraction features reach production approximately 3.6 times faster than with traditional F3, with no increase in infrastructure compute costs.

However, the most counterintuitive conclusion of this project is not the acceleration metrics, but the discovery made in step three. The team had eliminated all redundant calculations, yet CPU usage remained high. The cause was not how many times computations ran, but where results were stored and how many copies existed. Two designs can execute identical operations in the same order, but one scatters results across multiple independent containers while the other concentrates them in a single continuous buffer—the latter behaves completely differently regarding cache performance. On modern CPUs, the physical location of data and the number of copies matter as much as the number of computations. This insight is more transferable than any single-step optimization: it implies that deduplication is merely one layer of efficiency engineering, while memory layout is another, often revealing itself only after deduplication is complete.