LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-1448 Beiträge

Aktualisieren

Pydantic Engineer Duncan McKinnon: Enabling Coding Agents to Query Their Own Traces, Turning Repeated Failures into Verified Skills

Thema · Logfire MCPWesentlich
2026-11-13 17:00 GMT+8

Coding agents read files, query databases, call tools, and edit code in the background, while people outside the terminal only see a line saying "working on it...". Model invocations, tool executions, retries, and dead ends all occur out of sight, without records or the ability for long-term reflection. Pydantic engineer Duncan McKinnon believes that process opacity is one of the major challenges currently facing coding agents. Trace data should not be used solely for passive monitoring; instead, the Logfire MCP server should enable coding agents to query their own historical operation records, forming a closed-loop improvement process from failure evidence to skill creation and finally to experimental verification.

The specific path involves four steps. First, collect evidence from traces. McKinnon found that he repeatedly encountered invalid queries and rework when creating dashboard panels, so he used the Logfire MCP to search coding agent traces, pinpointing specific correction records in each case—for example, a query that double-counted cached input tokens in one instance. Second, identify the correct path with the agent and create skills, encoding patterns from historical successes and failures into reusable workflows. Third, run controlled experiments using the same agent configuration, sessions, and prompts, varying only whether the skill is loaded. Fourth, verify whether the improvement truly solved the problem without introducing new issues.

He selected 12 dashboard panel creation scenarios, which included various errors, and ran them on the same Codex agent both with and without the newly created logfire-dashboard-query skill loaded. Without the skill, 9 out of 12 runs failed, and formatting errors appeared; with the skill, all scenarios succeeded, raising the assertion pass rate from 57/72 to 72/72, with sql_executes improving from 25% to 100%.

The same closed-loop approach was also applied to agent configuration evaluation. In McKinnon's software factory pipeline, the review-fix phase ($43.30) and the review phase ($36.80) together accounted for about 60% of the $132 total cost, indicating that the agent was stuck in a cycle of reviewing and fixing. He collected 5 small PR review cases from traces, built a dataset, and then ran controlled experiments on Claude Code (haiku-5.5), Codex (gpt-6-luna), and Cursor (composer-2.5). Using two evaluation criteria—whether the review was overly nitpicky about format and style, and whether it adhered to the coding conventions specified in the instructions—the results showed that Codex scored highest on every panel (overall 3.93), followed by Cursor (2.80), with Claude Code scoring lowest (2.60).

The boundaries of this method are also clear. It relies on Logfire's ability to capture traces for specific agents and providers, and cost estimates cover only instrumented calls. Experimental conclusions are limited by the specific task types and the design of the evaluation criteria. McKinnon himself stated that these examples are "not comprehensive," but they represent the use cases he has seen having the greatest impact within his team.

Quellen:pydantic.dev

Forschung

CNBC Pro Contributor Michael Khouw: Nvidia CEO Change Raises Implied Volatility, Covered Calls Convert It to Excess Premium

Wesentlich
2026-11-14 03:30 GMT+8

Following the abrupt resignation of Nvidia CEO Jensen Huang, the stock has fallen approximately 30% since October 6 and is currently trading in the $160 range. CNBC Pro contributor Michael Khouw believes this panic creates an opportunity for option sellers: implied volatility has spiked to its highest level since the tariff turmoil in the first half of 2025, and high volatility directly drives up option premiums.

Khouw's core logic is that when a company is in a leadership transition period, market pricing for future uncertainty will keep implied volatility at elevated levels for an extended time. At this point, buying stock and selling out-of-the-money call options (a covered call strategy) ties the collected premiums directly to implied volatility—the deeper the fear, the thicker the premium. His specific trade recommendation is: buy 100 shares of Nvidia and simultaneously sell one call option expiring on December 18 with a strike price of $185.

The risk management logic of this strategy lies in the fact that continuously collected excess premiums systematically lower the holding cost. If the stock trades sideways or recovers slowly during the CEO search, investors earn abnormally high cash yields on a stagnant asset; if the stock falls further, the premium income constitutes a substantial buffer, reducing overall downside risk. Khouw believes that although Huang was the architect of Nvidia's AI dominance, the company's moat—superior GPUs, the CUDA software ecosystem, and networking technology—does not leave with him, so the fundamentals are largely undamaged by the personnel change.

It should be noted that this is a CNBC Pro contributor article explicitly marked as personal opinion, which discloses at the end that Tidal holds all securities mentioned in the text.

Quellen:CNBC

Forschung

Microsoft Blames Network Upgrade for Global Outage, Rules Out AI or Cyberattack

Thema · 微软全球服务中断事件Strukturell
2026-11-14 03:08 GMT+8

US tech giant Microsoft stated on November 13 that its Microsoft 365 cloud services have largely recovered after a widespread outage lasting approximately 24 hours. A company spokesperson clarified that the incident was triggered by an issue during an earlier network capacity upgrade, which surfaced following a recent network change, explicitly ruling out cyberattacks or AI agents as causes.

The disruption began around noon Eastern Time on Thursday, initially affecting North America before spreading to more than 70 countries worldwide. Impacted services included Outlook email, Teams collaboration platform, and the Copilot AI assistant. The outage significantly disrupted critical sectors such as aviation, hospitals, and law enforcement, given the heavy reliance of businesses and governments on these tools.

To manage the backlog of emails sent during the downtime, Microsoft increased the email queue time to 48 hours to ensure delivery of all attempted messages. Independent researcher Jonathan Leitschuh warned that users might experience a flood of notifications and delayed emails over the next few hours as subsystems come back online.

While services are being restored, the episode underscores the risks associated with global digital infrastructure's dependence on a single vendor. Microsoft has promised a full post-incident report after completing its review.

Quellen:WIRED

Forschung

Rodrigo Kede Lima, President of Microsoft Frontier Company: It Is Not Technology That Slows AI Implementation, But Organization

Thema · 微软AI组织落地模式Wesentlich
2026-11-13 01:00 GMT+8

Most business leaders I meet no longer need to be convinced that AI works. They have seen the pilots, the demos, and the copilots. In his November 2026 blog post, Rodrigo Kede Lima, President of Microsoft Frontier Company, notes that the question in boardrooms has shifted: how do we move from AI that impresses to AI that actually changes how the business operates? His assessment is that the technology is ready; what slows down AI is everything surrounding it—value identification, trust and governance, operating model redesign, and enabling people to work in new ways.

The model he proposes has four elements: start with value, not technology; redesign processes, not just optimize individual tasks; iterate quickly alongside customer teams; and embed security, governance, and cost management from day one. The Northern Trust case specifically illustrates how this model works. Onboarding to its Asset Servicing Front Office Solutions platform originally took about two and a half weeks per client. Microsoft engineers collaborated with Northern Trust experts to build an AI-driven onboarding agent within the Excel and SharePoint tools analysts were already using, rather than replacing the existing toolchain. The result was an 83% reduction in data preparation time, dropping from two and a half weeks to three days, with a first-pass quality score of 96.5% in testing. Analysts retained control over decisions while gaining more time for higher-value client service.

Kede Lima cites IDC data to explain why this model matters: 66% of organizations report substantial or transformative results from AI, but only 48% have achieved scaling or sustained value across their organizations. Meanwhile, 71% of organizations expect AI spending to increase by at least 10% in the coming year. The gap between reported outcomes and scaled outcomes is precisely what the model he describes attempts to bridge. It should be noted that the case data (such as the 83% reduction at Northern Trust) comes from Microsoft’s self-reporting on its own consulting projects, presented at Microsoft’s discretion; although the IDC survey data is published by a third-party agency, the report is sponsored by Microsoft. Neither has independent verification.

Quellen:The Official Microsoft Blog

Forschung

Tripadvisor Outgoing CEO Goldberg Says AI Channel Conversion Rates Are Multiple Times Higher Than Site Average, Community Data Is Core Defense

Thema · TripadvisorWesentlich
2026-11-13 18:35 GMT+8

The US online travel platform Tripadvisor is facing a contradictory situation: CEO Matthew Goldberg and CFO Michael Noonan have both announced their departures within the year, while the company reports an 11% year-over-year revenue decline, plans to eliminate approximately 425 positions in its Hotel and Other business divisions, and sees its stock price fall—all while the broader industry continues to grow.

Before his departure, Goldberg made a specific assertion: booking conversion rates via AI channels are significantly higher than the site's average. He stated that Viator’s AI channel conversion rate is close to five times the site average, while Tripadvisor’s is about double; bookings integrated through Model Context Protocol (MCP) partners convert at better-than-average rates with higher average order values. However, he simultaneously acknowledged that AI channels currently account for less than one percent of overall traffic.

What is the defensive logic behind these figures? Goldberg believes Tripadvisor’s marketplace model is resilient against disintermediation by upper-funnel Large Language Models (LLMs), citing the fragmented supply base of small operators, complex transaction management, and customer service needs on both sides of the marketplace. He further noted that the company’s reviews, quality rankings, and availability signals enable AI channels to surface better recommendations, thereby maintaining value within the AI ecosystem.

The underlying asset supporting these judgments is the community. Goldberg emphasized that contributor durability far exceeds overall traffic trends—even as total visits decline, the output of user-generated content has not shrunk at the same pace. He argued that in the AI era, this ongoing human contribution, trusted human perspective, and the resulting accumulated content and data constitute an extraordinarily valuable asset.

This argument comes against the backdrop of AI-driven disintermediation pressure, which is the most common analysis regarding Tripadvisor’s current difficulties. However, peers in the sector have largely viewed AI as a growth engine rather than a threat. Stuart Lauchlan, author at diginomica, agrees with Goldberg’s assessment but points out that he still needs to re-convince Wall Street.

Quellen:diginomica

Forschung

SageMaker HyperPod Data Capture: Turning Post-Deployment Blind Spots into Training Data

Wesentlich
2026-11-14 02:01 GMT+8

Models deployed into production often encounter inputs that were entirely absent from their training sets: proprietary terminology for new products, shifts in user behavior, or unexpected formatting. These issues remain invisible in offline benchmark tests because those benchmarks are frozen in the past. The core argument of this AWS blog is direct: without recording production inference data, teams cannot detect drift, identify failure patterns, or improve models using real interactions. The new Data Capture feature in Amazon SageMaker HyperPod writes every inference input and output to Amazon Simple Storage Service (Amazon S3) across three independent tiers (endpoint, load balancer, and model pod). The authors state that the latency impact is minimal and does not affect model availability.

The article walks through a concrete scenario to demonstrate this closed loop: a home-appliance retailer deployed a support-ticket router using vLLM with guided JSON decoding to classify tickets into six product categories. After going live, the retailer launched its own Arctis series of air conditioners. Because the router lacked a category for these units, it either routed them to 'Other' or incorrectly sent them to ProductFridge. The authors simulated pre- and post-launch traffic using 900 synthetic tickets. By capturing all records via Tier 3 to S3 and querying them with Athena, they found that the share of tickets in the 'Other' category jumped from 4.8% before launch to 36.8% after. A word-frequency analysis showed the term 'arctis' appeared 141 times, precisely pinpointing the blind spot.

After identifying the issue, the authors translated the diagnosis into a SQL labeling policy: tickets matching keywords like 'arctis', 'air condition', 'ac unit', or 'mini split' were relabeled as ProductAC, while others kept their original predictions; mixed signals were flagged for human review. This policy generated 900 training samples from the captured data. Using a second GPU on the same cluster, the HyperPod training operator ran a 60-second LoRA fine-tuning job (rank 16, approximately 29.9 million trainable parameters) and deployed the merged model as a second endpoint. On a held-out evaluation set of 280 tickets, recall for the new ProductAC category improved from 0/40 to 40/40, with no regression observed in the five established categories.

However, the authors highlighted several important limitations. First, the model's self-reported 'confidence' field is not a calibrated probability: among 14 air-conditioner tickets misrouted to ProductFridge, 10 reported confidence scores above 0.9, indicating that filtering by confidence would retain many incorrect labels. Second, because the evaluation and training sets came from the same generator, the 100% accuracy on the held-out set reflected template consistency rather than generalization capability. The authors wrote 24 manual probe tickets to test further; after fine-tuning, the model scored 20/24, exposing two shortcuts learned from the data: equating 'cooling' with air conditioners, and labeling any ticket mentioning Arctis as ProductAC (including stock and price inquiries). Third, the keyword strategy only works when new category boundaries are clear; ambiguous or overlapping categories require routing to Amazon SageMaker Ground Truth for human labeling. Ultimately, this is a vendor-conducted walkthrough using synthetic data, not an independently verified production result.

Quellen:Amazon Web Services

Forschung

EU Parliament Overwhelmed by AI-Generated Content: Fake Legal Citations Cause Administrative Backlog, Complaints Rise 50%

Thema · 欧盟议会AI生成内容危机Wesentlich
2026-11-14 01:21 GMT+8

Despite viewing itself as a global pioneer in artificial intelligence regulation, the European Union's own institutions in Brussels and Strasbourg are grappling with the consequences of uncontrolled generative AI use. According to a report by German tech media heise online citing independent news platform Rapporteur, the EU Parliament and Commission are being overwhelmed by automatically generated content, causing organizational problems that are difficult to manage even with technical assistance.

Alessandro Chiocchetti, Secretary-General of the European Parliament, recently warned of a rapid increase in amendments, parliamentary questions, petitions, and requests for access to files driven by generative AI. Parliament President Roberta Metsola also expressed concern, urging members not to delegate democratic responsibility to algorithms and emphasizing that AI-drafted amendments should not replace genuine dialogue with colleagues or direct contact with voters. On Thursday, the Parliament adopted a resolution on AI in administration and democracy, with new administrative guidelines expected soon.

The impact is visible in daily operations. Interpreters translating documents into 24 official languages using traditional machine translation without AI have encountered "hallucinations" in texts submitted by representatives, including references to non-existent laws. The Parliament's legal service reportedly spends days verifying these AI-generated claims. Observers note that some parliamentarians use AI to simulate activity with minimal effort to improve their standing in internal rankings.

In response, Irish Member of the European Parliament Barry Andrews introduced an internal directive prohibiting his own use of AI in votes, plenary speeches, press releases, and commission questions. Transparently, the directive itself notes it was created with AI assistance. Conversely, Swedish MEP Jörgen Warborn argues that AI significantly boosts productivity, noting that the Industrial Accelerator Act received over 5,000 amendments; while AI should not be used solely to overload the system, he welcomes greater efficiency.

Pressure extends to other bodies. A senior EU official told Rapporteur that complaints to the European Ombudsman have increased by 50% due to AI, threatening to bury the institution in paperwork. Estonian AI expert and Commission advisor Luukas Ilves describes this as a "flood" by autonomous software agents. While the EU is developing its own AI tools to analyze internal documents and keep sensitive data away from US cloud services, administrative inertia makes it challenging to compete with commercial investments from the US.

Quellen:heise online

Forschung

AI Agent Tooling Provider LangChain Secures $300M Investment at $3B Valuation

Wesentlich
2026-11-13 07:24 GMT+8

San Francisco-based startup LangChain has raised $300 million in new funding, bringing its valuation to $3 billion. According to SiliconANGLE, citing The Wall Street Journal, the round was led by Alphabet Inc.'s CapitalG fund, with participation from IVP, Sapphire Ventures, GIC, and DFJ Growth.

LangChain is best known for its open-source framework that simplifies the development of artificial intelligence agents by connecting them to large language models, integrating tools, and managing system prompts. Beyond the core framework, the company offers LangGraph for multi-agent workflows and Deep Agents for complex task automation. These tools have seen widespread adoption, with previous disclosures indicating approximately 90 million monthly downloads.

The company generates revenue through LangSmith, a commercial suite for building, maintaining, and observing AI agents in production. Major tech firms such as Autodesk and Cisco Systems are listed as customers of this platform.

With the new capital, LangChain plans to nearly double its headcount to 850 employees by the end of 2027. The hiring focus will be on infrastructure experts, AI engineers, sales, and marketing personnel to support the expansion of its commercial offerings and technical capabilities.

Quellen:SiliconANGLE

Forschung

Ai2 Releases XYBench to Test If Large Models Can Identify User Misconceptions and Redirect to Real Goals

Thema · XYBenchWesentlich
2026-11-14 01:02 GMT+8

Non-profit research institute Allen Institute for AI (Ai2), in collaboration with researchers from Carnegie Mellon University, NVIDIA, and Johns Hopkins University, has released XYBench, a new benchmark comprising over 8,000 questions designed to test how AI assistants respond to requests based on misconceptions. The benchmark targets the 'XY problem': a user wants to achieve goal X but assumes approach Y will work, thus asking for help with Y while ignoring the actual issue. For example, a user might ask how to keep frosting from sliding off cupcakes, when the real problem is that the cupcakes are too warm to frost.

The research team evaluated models including Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro. The results showed that while these models often suggested solutions, they tended to keep their answers focused on the user's original, mistaken request rather than steering them toward their underlying goal. Compared to human reference answers, all five models were less likely to steer users away from the mistaken approach and toward a solution suited to their goal.

XYBench measures model responses across three dimensions: pragmatic solution (whether it addresses both the stated request and the intended goal), pragmatic emphasis (whether the answer focuses on the stated request or the intended goal), and misconception identification (whether it discourages the mistaken approach and guides toward the correct one). Experiments indicated that even when the user's potential goal was explicitly provided, models still underperformed humans in misconception identification, and sometimes performed worse, suggesting that detecting intent is only part of the challenge; models must also act on it when it diverges from the literal request.

The benchmark won the Best Resource Paper Award at EMNLP 2026. Data, code, and the paper are publicly available for researchers to evaluate whether prompting or training improvements can help models better identify misconceptions and offer more effective alternatives.

Quellen:huggingface.co

Forschung

NVIDIA SOMA Retargeter Pipeline Analysis: Scaling, Inverse Kinematics, and Post-Processing Corrections for Specific Configuration Issues

Wesentlich
2026-11-14 01:00 GMT+8

Humanoid robot researchers face a problem that seems simple but is actually tricky: human demonstrations capture the coordination and rhythm of natural movement, but robots cannot directly copy the same joint motions because their limb proportions, joint structures, and ranges of motion differ. John Malaska, Senior Product Manager at NVIDIA, provides a detailed explanation in a technical blog post about the adaptation pipeline of the open-source tool SOMA Retargeter, describing which stages make up the pipeline, what conflicts each stage addresses, and how configuration options affect the final output.

The pipeline consists of three stages. The first stage is scaling: mapping motion targets on the human skeleton proportionally to the size of the target robot. The second stage is inverse kinematics (IK) solving: using the Newton physics engine, built on NVIDIA Warp, to find a set of joint angles for each frame that allows the robot to get as close as possible to the scaled target positions. The third stage is post-processing correction, most typically foot-plant correction—detecting contact intervals and adjusting foot trajectories to eliminate sliding artifacts. These three stages address problems at different levels: scaling resolves size differences, inverse kinematics resolves joint reachability constraints, and post-processing correction resolves contact consistency.

Malaska uses a source action of a one-handed wave to illustrate the role of reference pose alignment: if the Booster T1's reference pose is not aligned with the SOMA human reference, the robot will extend both arms outward; after alignment, one hand waves beside its head while the other arm rests naturally at its side. The input motion and IK parameters are identical; the only variable is the alignment of the reference pose. This example isolates the effect of the pose-mapping variable, demonstrating that while scaling solves size differences, incorrect pose mapping causes the robot to perform semantically incorrect actions.

When a robot cannot simultaneously satisfy the targets for all body parts, IK weights determine which parts take priority. SOMA Retargeter provides an experimental IK weight optimizer that automatically adjusts weights using a short segment of motion representative of the target task—it repeatedly retargets these motions, checking the tightness of the robot's following of the targets and the smoothness of the motion, and saves the selected weights as a new configuration. Malaska recommends tuning the configuration with representative motions first, then processing larger motion collections through headless batch retargeting.

The output is a CSV file containing the robot's root position, orientation, and joint angles for each frame. These trajectories can be converted into motion library formats such as SONIC or ProtoMotions for use in reinforcement learning policy training. During training, the policy learns how to convert retargeted references into actual control actions; during simulation, balance, ground contact, actuator limits, and tracking accuracy are evaluated.

Quellen:NVIDIA Technical Blog

Forschung

Nvidia CEO Jensen Huang Resigns; AMD Stock Rises 5%

Strukturell
2026-11-13 04:59 GMT+8

U.S. semiconductor giant Nvidia announced on November 12, 2026, that its Chief Executive Officer Jensen Huang has resigned. The news was reported by Business Insider, which cited an extramarital affair scandal and a board investigation into concerns that Huang may have been the subject of a foreign intelligence effort.

Huang's departure sends shockwaves through the AI chip industry. As Nvidia's co-founder and long-time leader, his sudden exit creates a significant power vacuum. According to reports, some industry observers have floated AMD CEO Lisa Su as a possible outside candidate for the role, although this remains speculative.

Market reaction was swift and clear. On the day Nvidia announced Huang's resignation, shares of its main competitor, Advanced Micro Devices (AMD), rose by 5%. Investors appear to view the uncertainty surrounding Nvidia's leadership as an opportunity for AMD to expand its market share.

Su and Huang are not only business rivals but also cousins. Since becoming AMD's CEO in 2014, Su has successfully transformed the company from the brink of collapse into a key player in global high-performance computing. This personnel change complicates the competitive landscape between the two firms, particularly as Nvidia faces a governance crisis while AMD maintains relative stability.

Quellen:Business Insider

Forschung

Malaysia and Singapore Step Up Enforcement on Nvidia Chip Exports: Charges Filed Over 917 Smuggled Servers, Megaspeed Hires Law Firm for Audit

Wesentlich
2026-11-14 00:15 GMT+8

According to Tom's Hardware, Malaysia and Singapore are intensifying scrutiny of AI hardware shipments amid concerns about circumventing U.S. export restrictions on China. Malaysia's Ministry of Investment, Trade and Industry is investigating whether Speedmatrix, the Malaysian subsidiary of Singapore-based Megaspeed, re-exported any AI processors and servers imported for local data centers. Despite months of investigation, authorities have yet to determine if such shipments occurred or violated Malaysian laws. Malaysia introduced restrictions on the export, transit, and trans-shipment of advanced U.S.-developed AI accelerators in July 2025, requiring companies to obtain permits.

Meanwhile, Singaporean authorities have brought 41 criminal charges against Shi Bingzhe, a 34-year-old Chinese citizen who served as director of Superphi and chief operating officer of NJ Global Financial. Prosecutors allege that these two companies purchased 917 servers containing restricted Nvidia hardware from Aperia International in 41 transactions between May 2024 and February 2025, with reason to believe the servers were linked to criminal activity. The case now involves five individuals and six companies. If convicted, Shi faces up to 10 years in prison and a fine of S$500,000 per charge.

In response, Megaspeed announced on November 12 that it had hired a U.S. law firm to conduct a comprehensive inventory of its AI GPUs and independently assess its compliance with U.S. export-control regulations. The company confirmed it had cooperated with inspections by the U.S. Department of Commerce's Bureau of Industry and Security (BIS). Megaspeed CEO James Tan Yong Pong stressed that compliance and transparency are priorities, though the company did not disclose the number or models of GPUs owned, nor the specific findings of BIS inspections.

These developments highlight the critical role of Malaysia and Singapore in the supply chain of restricted AI hardware. While neither investigation has confirmed that hardware ultimately reached China, the enforcement actions illustrate the difficulties authorities face in enforcing export controls across international supply chains.

Quellen:Tom's Hardware

Forschung

Sie sind auf dem aktuellen Stand in dieser Ansicht

Anmelden / Registrieren
LesenForschungRadarAnlageframework

It should be noted that this blog post is a first-party description by an NVIDIA product lead regarding a tool they support. It explains the design logic and configuration options of the pipeline but does not provide independent third-party quantitative assessments of the retargeting output quality, nor does it compare with other retargeting tools. SOMA Retargeter is open source, and its configuration files and reference results can be inspected in the GitHub repository, but the output quality of the pipeline itself—for example, the tracking error of the retargeted motion in physics simulations—still requires verification by users on their own robots and tasks.