{
  "version": "https://jsonfeed.org/version/1.1",
  "title": "ITANU Magazine",
  "home_page_url": "https://itanu.org/magazine/",
  "feed_url": "https://itanu.org/magazine/feed.json",
  "description": "Technology news in nine languages, three editions a day. Every story names its sources and is checked against them.",
  "language": "en",
  "icon": "https://itanu.org/icon-512.png",
  "authors": [
    {
      "name": "ITANU Magazine newsroom",
      "url": "https://itanu.org/magazine/about/"
    }
  ],
  "items": [
    {
      "id": "https://itanu.org/magazine/rust-parallel-compiler-frontend-nightly-default-f6qmkkta/",
      "url": "https://itanu.org/magazine/rust-parallel-compiler-frontend-nightly-default-f6qmkkta/",
      "title": "Rust turns on parallel compiler frontend by default in nightly builds",
      "summary": "Rust's Compiler Team says nightly builds now run the frontend, which does type checking and borrow checking, in parallel by default. It wants to find remaining issues before it stabilizes the feature.",
      "content_html": "<h3>Key points</h3><ul><li>Rust&apos;s Compiler Team says the parallel frontend has been on by default in nightly builds since nightly-2026-11-14.</li><li>The frontend, which does type checking and borrow checking, has run on a single thread by default; parallel mode has been an opt-in on nightly for three years.</li><li>The team will watch GitHub and Zulip for reported issues over the next few months and plans to stabilize the feature once it is confident it works well.</li></ul><p>Rust&apos;s Compiler Team announced on 16 November 2026 that the Rust compiler&apos;s parallel frontend is now on by default in the nightly channel. The team says the aim is to find any remaining problems before the feature is stabilized. Jakub Beránek wrote the post on behalf of the Compiler Team.</p><p>According to the team, the compiler has used multiple threads for a long time in its backend, which is built on LLVM. The frontend, which covers type checking and borrow checking, has run sequentially on one thread, and the team says this limits the compiler&apos;s performance.</p><p>The post says the frontend has been able to run in parallel for a long time, but various issues kept it from being the default. Three years ago the team made the parallel mode an opt-in on nightly, with a plan to stabilize it in 2024, a goal it did not reach. Last year Vadim Petrochenkov and several other contributors began a coordinated push that fixed many long-standing issues and improved performance, according to the team. The new default has applied since nightly-2026-11-14.</p><p>The team says it especially wants to hear about compiler crashes and about binaries that cannot be reproduced when the parallel frontend is enabled. It also says further performance gains may be possible, because work so far has focused on correctness rather than speed tuning.</p><p>Over the next few months, the team says it will watch GitHub and Zulip for reports about the parallel frontend. If no major issues appear and it is confident the feature works well, it intends to stabilize it.</p><h3>Sources</h3><ol><li><a href=\"https://blog.rust-lang.org/2026/11/16/enabling-parallel-frontend-on-nightly\">Enabling the parallel frontend on nightly</a>, Rust Blog</li></ol><p>ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.</p>",
      "date_published": "2026-11-16T17:37:51.000Z",
      "date_modified": "2026-11-16T17:37:51.000Z",
      "authors": [
        {
          "name": "ITANU Magazine newsroom",
          "url": "https://itanu.org/magazine/about/"
        }
      ],
      "tags": [
        "Software",
        "Open source"
      ],
      "language": "en"
    },
    {
      "id": "https://itanu.org/magazine/nvidia-vera-rubin-mlperf-training-debut-tfqpvtzr/",
      "url": "https://itanu.org/magazine/nvidia-vera-rubin-mlperf-training-debut-tfqpvtzr/",
      "title": "NVIDIA says Vera Rubin systems deliver up to twice GB300 training speed per GPU",
      "summary": "The Vera Rubin NVL72 system made its first appearance in MLPerf Training v6.1. MLCommons, which runs the benchmark, confirmed that this round was the first to include Rubin-generation GPUs.",
      "content_html": "<figure><img src=\"https://itanu.org/assets/magazine-img-a1828ad62e1dda00-960.jpg\" width=\"1166\" height=\"948\" alt=\"A white Nvidia sign with a green logo and street address stands among trees and shrubs on a corporate campus.\"><figcaption>File photo: Nvidia sign at its Santa Clara headquarters campus, 2020 · Will Buckner / Wikimedia Commons · CC BY 2.0</figcaption></figure><h3>Key points</h3><ul><li>NVIDIA says Vera Rubin NVL72 reached up to 2x the per-GPU performance of GB300 NVL72 in MLPerf Training v6.1.</li><li>MLCommons confirmed that NVIDIA Rubin-generation GPUs and Google&apos;s TPUv7 (Ironwood) appeared in the round for the first time.</li><li>Results came from 24 submitting organizations, which entered 153 systems and produced 250 test results.</li><li>NVIDIA&apos;s provisional Vera Rubin results in MLPerf Endpoints are still awaiting MLCommons verification.</li></ul><p>NVIDIA&apos;s Vera Rubin NVL72 system posted its first MLPerf Training results on 16 November 2026. The company says that, measured per GPU, the system reached up to 2x the performance of its GB300 NVL72 system. MLCommons, the group behind the benchmark, said the v6.1 round was the first to include NVIDIA Rubin-generation GPUs.</p><p>MLPerf Training measures how quickly hardware can bring various AI models to a chosen accuracy level. NVIDIA says it swept every benchmark in v6.1.</p><p>NVIDIA also reported results on scaling. In the DeepSeek-V3 benchmark, it says Vera Rubin NVL72 went from 256 to 1,024 GPUs at about 97% efficiency, and four times as many GPUs cut training time from 16 minutes to about four. NVIDIA ran the same benchmark on 10,240 GPUs in GB300 systems, the largest scale in the round, and Microsoft Azure matched it with 10,240 NVIDIA GPUs in GB200 systems.</p><p>According to MLCommons, 24 organizations submitted results, including AMD, Google, Dell, HPE, Oracle, CoreWeave and Crusoe, a first-time entrant. Together they entered 153 systems and produced 250 test results. Cloud-based systems made up 37% of the total, compared with 29% in v6.0, and a dozen entries ran on at least 1,000 accelerators. Hardware came from NVIDIA, AMD and Google, and Google&apos;s TPUv7 (Ironwood) also appeared for the first time.</p><p>Version 6.1 adds two new benchmarks. One is DLRMv4-HSTU, a recommendation test that replaces the earlier DLRM-dcnv2 and trains on the Yambda dataset. The other is Qwen3.5-397B, a large-scale workload for post-training a large language model.</p><p>NVIDIA also entered MLPerf Endpoints for the first time. The figures for Kimi K3, a large open model designed for agent-style workflows, are provisional. NVIDIA says they are pending MLCommons Association verification and may change after peer review.</p><h3>Sources</h3><ol><li><a href=\"https://blogs.nvidia.com/blog/mlperf-training-endpoints-vera-rubin\">NVIDIA Sweeps MLPerf Training and Endpoints as Vera Rubin Debuts With Leading Performance</a>, NVIDIA Blog</li><li><a href=\"https://mlcommons.org/2026/11/mlperf-training-v6-1-results/\">MLPerf Training v6.1 Results: Record Participation</a>, MLCommons</li></ol><p>ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.</p>",
      "image": "https://itanu.org/assets/magazine-img-a1828ad62e1dda00-1166.jpg",
      "date_published": "2026-11-16T17:19:41.000Z",
      "date_modified": "2026-11-16T17:19:41.000Z",
      "authors": [
        {
          "name": "ITANU Magazine newsroom",
          "url": "https://itanu.org/magazine/about/"
        }
      ],
      "tags": [
        "Hardware",
        "Chips",
        "Semiconductors",
        "LLMs"
      ],
      "language": "en"
    },
    {
      "id": "https://itanu.org/magazine/microsoft-months-old-bug-microsoft-365-outage-nkrptpis/",
      "url": "https://itanu.org/magazine/microsoft-months-old-bug-microsoft-365-outage-nkrptpis/",
      "title": "Microsoft blames months-old bug for global Microsoft 365 outage",
      "summary": "A defect shipped in July to Microsoft 365 networking gateway servers combined with extra Teams traffic and reduced capacity. At the worst point, more than 80% of servers stopped responding.",
      "content_html": "<figure><img src=\"https://itanu.org/assets/magazine-img-794e556d61341ab2-960.jpg\" width=\"1600\" height=\"1066\" alt=\"Aerial view of the Microsoft headquarters campus in Redmond, Washington, with construction cranes and Lake Sammamish beyond.\"><figcaption>File photo: Microsoft headquarters campus redevelopment in Redmond, Washington, 2021 · Atomic Taco / Wikimedia Commons · CC BY-SA 2.0</figcaption></figure><h3>Key points</h3><ul><li>Microsoft says a bug in a July update to Microsoft 365 networking gateway servers caused the nearly 24-hour outage on 12 November.</li><li>Extra traffic after a Teams failover and reduced server capacity also contributed, according to a preliminary report.</li><li>At the lowest point, more than 80% of servers were unresponsive.</li><li>A final post-incident report is expected later this week.</li></ul><p>Microsoft says a defect in an update it released in July for its Microsoft 365 networking gateway servers was behind the nearly 24-hour worldwide outage on 12 November, BleepingComputer reported.</p><p>The disruption started at 11:20 ET on Thursday 12 November and ended at 11:10 ET the next day, when Microsoft confirmed that services were fully restored. Outlook, Exchange Online, Teams and Copilot were affected, and users lost access to email or faced lengthy waits for messages to arrive.</p><p>In an initial post-incident report shared through the Microsoft 365 Admin Center, the company named three contributing problems: the previously unnoticed bug, extra traffic following a Teams failover, and reduced server capacity. The update had run for more than three months without trouble, but under certain traffic patterns the bug could overload servers until they stopped responding.</p><p>Microsoft executive Charles Lamanna described the bug as a &quot;latent defect&quot; that caused lock contention, which stopped machines from responding. Lock contention happens when several processes wait on the same lock, so work stalls until it is released. Lamanna said the faulty code came from a fix written specifically for the gateway servers, and that neither the fix nor the underlying defect applies to wider Windows deployments. He said Microsoft first observed the issue during an earlier Microsoft 365 update in July 2026.</p><p>Petri reports that the defect lay dormant since July 2026 and escaped testing because Microsoft&apos;s validation environments mostly used older hardware, while the bug was more likely to surface on newer servers with higher core counts under specific workloads. Microsoft had recently moved Teams traffic onto the gateways that also serve Exchange Online, and a separate issue had already cut available capacity.</p><p>Automated systems then shifted traffic to servers that were still healthy, where the same defect was triggered, according to Petri. Lamanna said the load was spread over ever fewer servers, and within less than two hours the defect had been hit across the fleet; at its lowest point, more than 80% of servers were unresponsive.</p><blockquote><p>We are deeply sorry for the impact to all our customers</p><footer>Charles Lamanna, Microsoft</footer></blockquote><p>Microsoft says its investigation is continuing and that it expects to publish a final post-incident report later this week.</p><h3>Sources</h3><ol><li><a href=\"https://bleepingcomputer.com/news/microsoft/microsoft-says-months-old-bug-caused-global-microsoft-365-outage\">Microsoft says months-old bug caused global Microsoft 365 outage</a>, BleepingComputer</li><li><a href=\"https://www.itpro.com/software/microsoft/microsoft-365-outage-company-exec-lifts-lid-on-the-cascading-failure-that-caused-lengthy-disruption\">Microsoft 365 outage: Company exec lifts lid on the &apos;cascading failure&apos; that caused lengthy disruption</a>, ITPro</li><li><a href=\"https://petri.com/microsoft-explains-microsoft-365-outage/\">Microsoft Explains Causes of Microsoft 365 Outage</a>, Petri</li></ol><p>ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.</p>",
      "image": "https://itanu.org/assets/magazine-img-794e556d61341ab2-1600.jpg",
      "date_published": "2026-11-16T17:11:05.000Z",
      "date_modified": "2026-11-16T17:11:05.000Z",
      "authors": [
        {
          "name": "ITANU Magazine newsroom",
          "url": "https://itanu.org/magazine/about/"
        }
      ],
      "tags": [
        "Software",
        "Cloud"
      ],
      "language": "en"
    },
    {
      "id": "https://itanu.org/magazine/alibaba-qwen4-max-cuts-task-costs-nikpq6e5/",
      "url": "https://itanu.org/magazine/alibaba-qwen4-max-cuts-task-costs-nikpq6e5/",
      "title": "Alibaba says its new Qwen4-Max language model cuts task costs by about 60%",
      "summary": "Alibaba calls Qwen4-Max the most capable model in its Qwen family. The company says it removed offensive attack material from the model's training data and kept defensive security content.",
      "content_html": "<h3>Key points</h3><ul><li>Alibaba released Qwen4-Max on 16 November 2026 and calls it the most capable Qwen model so far.</li><li>On its own Work360 benchmark, Alibaba says the average cost per task is around 60% lower than with Qwen3.8-Max.</li><li>Alibaba says it filtered training data by task intent, removing offensive attack content but keeping defensive material.</li></ul><p>Alibaba released Qwen4-Max on 16 November 2026. The company calls it the most capable model in its Qwen family so far and says its average cost per task is around 60% lower than that of its predecessor, Qwen3.8-Max.</p><p>According to Alibaba, the model has 2.4 trillion parameters in total but activates only 71 billion for each token, using an upgraded sparse Mixture-of-Experts (MoE) architecture. In a mixture-of-experts model, each input is routed to only some of the network&apos;s specialised sub-networks rather than through all of them. Alibaba says this means 25% fewer activated parameters than Qwen3.8-Max, which is the same overall size. Qwen4-Max can also process images natively and has a context window of one million tokens.</p><p>The cost figure comes from Work360, a benchmark that Alibaba owns, which tests difficult industry tasks across 721 specialised fields. The company says Qwen4-Max beat its predecessor overall on that test and used 35% fewer tokens in its High effort reasoning mode. MLQ.ai reports that on Work360 the per-task average went down to $1.67 from $4.43 with the older model, which it describes as a 62% reduction. It also reports that input is now listed at $1.50 per million tokens, down from $2, and output at $4.50 per million, down from $6. MLQ.ai notes that these figures have not been checked independently.</p><p>Alibaba also says it tightened security controls during training. For cybersecurity, it filtered training data by the intent of each task, removing material on offensive attacks while keeping material that helps find and fix vulnerabilities. The company says coding training added security checks alongside functionality tests, and that dedicated safety datasets strengthen protection against chemical, biological, radiological and nuclear risks.</p><p>As an example of long-running agent work, Alibaba says Qwen4 ran for 48 hours without stopping. Over four rounds of a closed-loop workflow, it found a coolant molecule for chip cooling that was then validated by experiment.</p><p>The model is available through the API and Token Plan on QwenCloud, and in Alibaba applications including QwenWork and QwenStudio. CASRAI reports that Alibaba&apos;s changelog lists qwen4-max on 16 November, five days after a preview that was available only through Token Plan. The flagship model was released without downloadable open weights, CASRAI reports. Alibaba says it plans to release smaller Qwen4 models with open weights to developers soon, and CASRAI notes that the announcement gives no date for them.</p><h3>Sources</h3><ol><li><a href=\"https://www.alibabacloud.com/blog/alibaba-unveils-qwen4-max-new-ai-teammate-for-work-creativity-and-discovery_603783\">Alibaba Unveils Qwen4-Max: New AI Teammate for Work, Creativity and Discovery</a>, Alibaba Cloud</li><li><a href=\"https://mlq.ai/news/alibaba-launches-qwen4-max-saying-it-cuts-task-costs-by-about-60-with-tighter-cyber-guardrails/\">Alibaba launches Qwen4-Max, saying it cuts task costs by about 60% with tighter cyber guardrails</a>, MLQ.ai</li><li><a href=\"https://casrai.org/news/alibaba-qwen4-max-general-release-million-token-context-2026\">Alibaba releases Qwen4-Max with 1M context</a>, CASRAI</li></ol><p>ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.</p>",
      "date_published": "2026-11-16T11:22:00.000Z",
      "date_modified": "2026-11-16T11:22:00.000Z",
      "authors": [
        {
          "name": "ITANU Magazine newsroom",
          "url": "https://itanu.org/magazine/about/"
        }
      ],
      "tags": [
        "AI",
        "LLMs",
        "Cybersecurity",
        "AI safety"
      ],
      "language": "en"
    },
    {
      "id": "https://itanu.org/magazine/ula-vulcan-fails-amazon-leo-satellites-k5xvarwh/",
      "url": "https://itanu.org/magazine/ula-vulcan-fails-amazon-leo-satellites-k5xvarwh/",
      "title": "ULA's Vulcan rocket fails in flight and loses 40 Amazon Leo internet satellites",
      "summary": "The launch from Cape Canaveral on 15 November was the first outright failure for United Launch Alliance since it was formed in 2006. It was also the first of 38 Vulcan flights Amazon has bought.",
      "content_html": "<figure><img src=\"https://itanu.org/assets/magazine-img-a89fc46cc8d56515-960.jpg\" width=\"1600\" height=\"1201\" alt=\"A Vulcan Centaur rocket with two solid boosters stands on its mobile launch platform between two lightning towers.\"><figcaption>File photo: Vulcan Centaur rocket on the launch pad at Cape Canaveral, 2024 · NASA/Ben Smegelsky / Wikimedia Commons · Public domain</figcaption></figure><h3>Key points</h3><ul><li>Vulcan suffered an anomaly about 10 minutes into flight and did not put 40 Amazon Leo satellites into orbit.</li><li>The loss is ULA&apos;s first outright failure since the company was formed in 2006.</li><li>The flight was the first of 38 Vulcan launches Amazon has bought for its Leo network.</li><li>The cause was not immediately known, and Amazon Leo must wait for ULA&apos;s failure investigation.</li></ul><p>United Launch Alliance&apos;s Vulcan rocket failed to put 40 Amazon Leo broadband satellites into orbit on Sunday 15 November after an in-flight anomaly shortly after it lifted off from Florida. ULA said the problem struck about 10 minutes into the flight.</p><p>The 214-foot-tall rocket lifted off at 12:40 EST from pad 41 at Florida&apos;s Cape Canaveral Space Force Station. Six solid-fuel strap-on boosters and two methane-burning BE-4 engines powered it on a northeasterly path. Once the stages separated, the hydrogen-fuelled Centaur upper stage took over for a burn planned to last 12 minutes 15 seconds, and ULA&apos;s update said something went wrong during that firing. An upper stage is the part of a rocket that keeps firing after the lower stage drops away and carries the payload the rest of the way to orbit.</p><p>The cause was not immediately known. Earlier attempts had been called off on Friday, when a gaseous nitrogen plant serving several launch providers went down, and on Saturday, because of network problems between the rocket and ground systems.</p><p>In a statement, ULA acknowledged that the flight fell short of its benchmark for mission success.</p><blockquote><p>Today, we did not meet that standard.</p><footer>United Launch Alliance, in a statement referring to its benchmark for mission success</footer></blockquote><p>ULA had never lost a rocket outright since Boeing and Lockheed Martin set it up as a partnership in 2006. The mission was also ULA&apos;s first Vulcan flight in nine months. Vulcan is replacing the company&apos;s Atlas 5 fleet.</p><p>Sunday&apos;s flight was the first of 38 Vulcan launches Amazon has bought for its Leo satellites, which are meant to compete with Starlink and other space-based internet services. ULA has already flown eight Leo missions on Atlas 5 rockets. Amazon has also booked 24 launches on Blue Origin&apos;s New Glenn, 24 on Ariane 6 and 13 on SpaceX&apos;s Falcon 9. The constellation needs more than 3,300 satellites, and this flight would have raised the number launched so far to 436.</p><p>Amazon said the risk that comes with any spaceflight is higher on early missions such as this one, called LV-01, and it stood by its launch provider.</p><blockquote><p>ULA has been a trusted partner for Amazon, and we believe they&apos;ll learn from this mission and deliver on their long-term plans for Vulcan.</p><footer>Amazon Leo</footer></blockquote><p>Amazon Leo now has to wait for ULA&apos;s failure investigation to finish before it flies on Vulcan again. It is also waiting for New Glenn to return to flight after the rocket exploded on the pad last May. ULA has not said what exactly caused the failure or when Vulcan will fly again. The company said it will work with its customer and other relevant organisations to find the cause, decide on fixes and resume flights as soon as it can.</p><h3>Sources</h3><ol><li><a href=\"https://spaceflightnow.com/2026/11/15/ula-suffers-first-total-mission-failure-with-loss-of-vulcan-rocket-carrying-40-amazon-leo-satellites/\">ULA suffers first outright failure with loss of Vulcan rocket carrying 40 Amazon Leo satellites</a>, Spaceflight Now</li><li><a href=\"https://www.aa.com.tr/en/americas/us-rocket-firm-ula-fails-to-deploy-40-amazon-leo-satellites-after-flight-anomaly/4126036\">US rocket firm ULA fails to deploy 40 Amazon Leo satellites after flight anomaly</a>, Anadolu Agency</li></ol><p>ITANU’s newsroom system produced this article with AI and checked it against the sources listed above.</p>",
      "image": "https://itanu.org/assets/magazine-img-a89fc46cc8d56515-1600.jpg",
      "date_published": "2026-11-16T02:19:29.000Z",
      "date_modified": "2026-11-16T02:19:29.000Z",
      "authors": [
        {
          "name": "ITANU Magazine newsroom",
          "url": "https://itanu.org/magazine/about/"
        }
      ],
      "tags": [
        "Space",
        "Satellites"
      ],
      "language": "en"
    }
  ]
}
