Platform engineering team leads are facing a quiet crisis. Your data science teams want Kubeflow for its pipeline orchestration, metadata tracking, and training operators, so you build it for them on Kubernetes. Then day two arrives. Your engineering backlog is swallowed by breaking changes from upstream, Istio configuration complexity, security patching, and storage provisioning bottlenecks. You didn't build an ML platform; you accidentally adopted a full-time infrastructure maintenance program. The Kubeflow operations trap Kubeflow's day-two difficulty has structural roots. It is not a single, cohesive application but a distributed constellation of over a dozen distinct open source microservices, including Katib, Pipelines, Notebooks, and Central Dashboard. Each of these components comes with its own release cycle, dependency graph, and configuration quirks, which means that when platform teams deploy Kubeflow, they are actually signing up for a systems integration job. The friction concentrates in three systemic challenges. Kubeflow leans on Istio for routing, multi-tenancy, and security. Configuring Istio ingress, managing TLS certificates, and debugging broken virtual services can quickly turn into a time sink for senior infrastructure engineers. Kubeflow moves fast. Upgrading from one version to the next rarely involves a simple script, because a single API deprecation in an upstream Kubernetes component can silently break your entire machine learning pipeline orchestration. Machine learning workloads demand dynamic, high-performance storage provisioning and flawless GPU scheduling. Mapping cloud-native storage classes to Kubeflow's persistent volume claims while keeping data access latency low requires constant manual tuning. Managed Kubeflow, zero operational overhead Canonical's new Managed Kubeflow on Microsoft Azure is built to give operations teams their weekends back. It delivers the full power of upstream Kubeflow without the operational burden, and because it is a fully managed service that runs entirely within your own cloud tenancy, no data, no models, and no training workloads are ever sent to Canonical. Compliance teams keep their posture intact while platform teams shed the maintenance burden. The open source management engine is cloud-agnostic. Managed Kubeflow on Azure uses the same architecture as Canonical's on-premises OpenStack integration, and managed services on additional public clouds will follow. The result is environment portability without the operational overhead. Canonical Managed Kubeflow use cases Once teams are freed from infrastructure patching and service mesh debugging, they can concentrate on delivering business value. Kubeflow is a powerhouse when it works, because it provides the framework required to take models from an experimental notebook to high-throughput production. A fully managed platform abstracts the underlying cluster maintenance and turns complex machine learning workflows into repeatable, scalable operations. Here is how a managed, dedicated platform converts the heaviest machine learning workloads from infrastructure burdens into routine production operations. Generative AI: Offloading the compute complexity GenAI workloads push Kubernetes clusters to their limits, and managing the pipelines manually forces platform teams to write fragile, custom automation scripts. Canonical Managed Kubeflow on Azure handles this work natively inside your private Azure cloud tenancy. The generative AI workloads Kubeflow can run include: Distributed pre-training: Clustering multi-node GPU instances requires complex networking, node provisioning, and fault tolerance. Kubeflow orchestrates training jobs across nodes automatically and ties into Azure's low-latency network infrastructure to maximize hardware utilization without manual cluster tuning. Targeted fine-tuning: Data scientists constantly spin up LoRA or PEFT jobs that require immediate, heavy compute, only to leave idled GPUs burning budget later. Kubeflow pipelines automate the entire sequence: ingest data, run the fine-tuning job, and scale capacity back down to zero once the job finishes. Model distillation: Compressing large models into smaller, production-ready versions requires complex teacher-student pipelines. Kubeflow manages these multi-stage workflows, and teams can track training metrics side by side via the integrated MLflow server to validate model performance. Traditional ML: Solid production pipelines. While generative AI takes the spotlight, much core enterprise value still runs on traditional machine learning. Managed Kubeflow keeps these production systems reliably online. Traditional ML workloads include: Predictive maintenance: IoT and time-series data demand continuous updates. Kubeflow can schedule automated retraining pipelines triggered by data drift. This keeps models accurate without platform teams manually monitoring performance pipelines. Fraud detection: Compliance demands a watertight audit trail. The included MLflow server acts as a metadata engine that automatically logs every dataset version, hyperparameter choice, and model version to help assure robust regulatory compliance. Churn and demand forecasting: High-volume batch scoring requires massive, temporary compute scaling. Canonical Managed Kubeflow on Azure can autoscale the underlying infrastructure to process millions of rows, then tear it down cleanly to control cloud spend. Stop maintaining core Kubeflow. Start delivering value. A managed service exists to remove specialized infrastructure overhead without sacrificing data sovereignty. 100 percent in-tenancy: Because the service executes entirely inside your tenancy, your underlying data, source code, and custom weights never leave your perimeter. No hostage to fortune: The service is built on pure upstream Kubeflow, so the pipelines you run on Azure today can also run on Canonical's on-premises OpenStack solution or future cloud releases. Enterprise-grade security: The service integrates with enterprise identity management, including Microsoft Entra ID, and role-based access controls right from launch. Predictable reliability: No more debugging broken operator upgrades. Canonical's experienced managed services team handles backups, upstream fixes, security patches, and version migrations. Deploy in less than 30 minutes You can launch your first production-ready cluster in less than 30 minutes directly from the Azure Marketplace. Give your data scientists the environment they need, and keep full control of your infrastructure. Launch now on Azure. Contributed by Canonical.

Students and faculty at Brown University are worried generative AI could harm learning after an economics professor's take-home exam produced results he said pointed to widespread misuse of the technology. In a report [PDF] published by Brown's Generative AI in Teaching and Learning (GAITL) committee, teaching staff say they fear AI could weaken students' cognitive skills and encourage cheating. The committee's report comes days after Brown's Roberto Serrano warned that society "cannot choose to become idiots" – comments made after he all but proved his economics class was cheating on their midterms. Serrano allowed his students to complete their spring midterms outside class, while keeping it closed-book. After the December 2025 shootings at the university, which killed two and injured nine, the professor gave the nod to a home-based exam – the first in his almost 20 years at the university, he told Inside Higher Ed. For students, this meant effectively unlimited time to complete the test, and for Serrano, an opportunity to ramp up the difficulty. He said scores for these exams typically sit in the 65-80 percent range, but students on this test scored an average of 96 percent. Suspecting AI-assisted cheating had played a role, Serrano did not immediately call for a redo, in the hope that he simply had an especially talented cohort, but promised to void the results if students' scores on the final, which was taken under controlled conditions, did not meet similar standards. The average final score was 48.6 percent, a record low. Three students scored zero. Serrano told students that because of the discrepancy in scores, he was increasing the weighting of the final exam to 80 percent from 50 percent, and voiding the midterm results. "We cannot afford to have a society in which a significant fraction of our best young minds think that cheating is OK," he said. "That leads to a declining society, to a failed society… We cannot choose to become idiots." Brown's GAITL committee found that 56 percent of undergrads use AI tools daily or weekly, rising to 67 percent for graduate and medical students, and 85 percent for master's students. Students most often used AI for tasks that help build advanced cognition, such as explaining solutions to complex problems and debugging code. The committee also cited previous research on the topic in both US and UK higher education, finding that around 25 percent of students were submitting assignments completed with the help of AI tools, and the rate is increasing sharply each year. Students appear willing to adopt AI tools, despite many fearing the technology is making them dimmer. Eighty-eight percent of Brown students and 73 percent of graduate/medical students said they were concerned AI could have negative effects on their cognitive capacities. The same concerns were shared by teaching staff, 95 percent of whom feared for students' long-term learning, while four in five expected cognitive capabilities to decline. Pointing to wider research on AI tools' impact on student cognition, the GAITL committee noted a strong body of evidence to support the view that over-reliance could decrease higher-order thinking and metacognition. The research on the matter is still relatively sparse, but it supports the possibility of AI supplementing student cognition, provided it is used with specific guidance from teachers. Brown has since committed to enacting a range of the committee's recommendations, which cover near, medium, and long-term goals. The first steps will involve publishing university guidelines for using AI, and individual departments and faculties will establish their own standards. Further down the line, Brown will invest in improving AI literacy among teaching staff, which it hopes will improve the way in which AI is used – not abused – across the university, and place informed restrictions on how students can use it without compromising their development. ®

Werkgever Antonello Lovato bood twee jaar geleden geen hulp aan de Indiase arbeider Satnam Singh, die na een werkongeval overleed. De ophefmakende zaak vestigt de aandacht op de wijdverspreide uitbuiting van arbeidsmigranten in de Italiaanse landbouw.

Eind 2016 maakten drie Noorse voetbaltalenten voor de lol een videoclip in een speeltuin. Een van de ‘rappers’: de zestienjarige Erling Braut Haaland, alias ‘Lyng’.

Eén dorp in academisch Nederland houdt dapper stand. Waar anders dan in Leiden, bewaakt één hoogleraar de opleiding van zijn studenten als een leeuw.

Smoke City

WOLFSBURG (ANP/RTR/AFP) - De vakbonden gaan zich fel verzetten tegen de massaontslagen bij Volkswagen Group, waar naast Volkswagen ook merken als Porsche en Audi onder vallen. In verschillende fabrieken werd het werk neergelegd op het moment dat Volkswagen zijn herstructureringsplannen aan de raad van commissarissen voorlegde.

De Duitse autofabrikant met wereldwijd ongeveer 650.000 werknemers zou volgens eerdere berichten de komende jaren 100.000 banen willen schrappen. Daarbij zouden er ook vier fabrieken van het automerk sluiten, in Hannover, Zwickau, Emden en de Audi-fabriek in Neckarsulm. Volkswagen is al langer aan het herstructureren, omdat het bedrijf het lastig heeft door de Amerikaanse importheffingen, zwakke verkopen in China en toenemende concurrentie in Europa.

De vakbonden laten weten dat als de plannen worden doorgezet, de stakingen verder worden opgevoerd. De vakbond IG Metall meldde dat er alleen al in Wolfsburg vierhonderd mensen demonstreerden. In totaal heeft de vakbond opgeroepen op twintig locaties over heel Duitsland te protesteren.

DEN HAAG (ANP) - Luchtvaartmaatschappijen moeten boetes kunnen opleggen aan passagiers die zich misdragen, vindt het kabinet. Het is een van de plannen die minister Vincent Karremans (Infrastructuur, VVD) in een Kamerbrief voorstelt om wangedrag in vliegtuigen tegen te gaan. Als de boetes er komen naast strafrechtelijke vervolging, zou dat als "extra afschrikking" kunnen werken.

De VVD-minister ziet een "zorgwekkende stijging" van verstoringen in vliegtuigen, zoals schelden, bedreigen en fysieke agressie. "Helaas betreft dit een wereldwijde trend." In Nederland zou het dagelijks om twee à drie incidenten gaan. Vaak is er alcohol bij betrokken.

Karremans wil ook dat Europese luchtvaartmaatschappijen hun zwarte lijsten kunnen delen om verstorende passagiers te weren, vindt het kabinet. De bedrijven en vakbonden gaan met hulp van zijn ministerie kijken of het mogelijk is om meer informatie uit te wisselen. Als het plan slaagt, zouden passagiers die zich bij de ene luchtvaartmaatschappij misdragen dus ook niet meer welkom zijn bij de andere maatschappij.

KLM en Transavia delen hun zwarte lijsten al enkele jaren, aldus Karremans. Hij heeft recent in een Europees overleg al gepleit voor de gedeelde lijsten van verstorende passagiers, en zegt daarbij gesteund te zijn door zes andere lidstaten.

JINJIANG (ANP/AFP) - Bij een grote brand in een schoenenfabriek in de Chinese stad Jinjiang zijn zeker 28 doden gevallen, melden staatsmedia. De brand is uitgebroken in een vestiging van de Huiteng Shoes Company in het district Chendai.

De Chinese president Xi Jinping zei donderdag "een zwaar verlies aan mensenlevens" te vrezen. Hij gelastte de slachtoffers te vinden en een grondig onderzoek naar de oorzaak in te stellen.

Volgens Chinese media is de industrie in het voormalige vissersdorp Chendai enorm groot. In de plaats worden volgens de People's Daily Online jaarlijks een miljard paar sneakers gefabriceerd. De schoenenindustrie zou er goed zijn voor meer dan 6 miljard euro aan inkomsten.

Thomas Hawk posted a photo:

Thomas Hawk posted a photo:

ajpscs posted a photo:

looking backward rather than forward

the SQUARE
MUTED
TOKYO DAY WALK
© ajpscs

Do you believe that everybody should have fun or that only a few people should have fun? “This is what it means to be entertained in the United States of America in 2026. Want to have fun? Like, the most fun? Get rich, or die trying.”

De Pindakaasvloer van de pas overleden kunstenaar Wim T. Schippers is weer opgebouwd. Het kunstwerk is vanaf vandaag te zien in het Depot van museum Boijmans van Beuningen in Rotterdam. Bezoekers moeten wel voorzichtig zijn. In de loop der jaren liep de Pindakaasvloer meerdere keren schade op. Soms was het moedwillig, soms per ongeluk.

ajpscs has added a photo to the pool:

looking backward rather than forward

the SQUARE
MUTED
TOKYO DAY WALK
© ajpscs

PhilT_2013 has added a photo to the pool:

gnsk has added a photo to the pool:

Slecht nieuws voor mensen van omtrek de groteske gemeenschap BMI-uitgedaagden mensen met een hoog gewicht: uit een of ander onderzoek (niet gelezen, in plaats daarvan chips gegeten, red.) blijkt dat jullie minder snel uitgenodigd worden voor een sollicitatiegesprek, en als dat toch lukt minder snel aan worden genomen. Ook wordt jullie werk negatiever beoordeeld, verdienen jullie minder geld en maken jullie minder snel promotie, dit omdat (werkelijk geen idee, zit nu dropjes te eten, red.), of iets dergelijks. Daarom, bindend advies aan jonge mensen: vreet niet te veel, want straks heeft solliciteren amper nut, verdien je geen ruk, heb je geen perspectief op verbetering, en kan je eigenlijk alleen nog maar bij GeenStijl terecht. 

Afijn, zin in chips en dropjes nu.