<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/"><channel><title>Data Phoenix — Articles &amp; News</title><description>Data Phoenix is a media platform and community for engineers and technical leaders building AI and data solutions.</description><link>https://dataphoenix.info</link><language>en-us</language><image><url>https://dataphoenix.info/favicon-512.png</url><title>Data Phoenix — Articles &amp; News</title><link>https://dataphoenix.info</link></image><item><title>Apple launches U.S. study of conversational health AI</title><link>https://dataphoenix.info/news/apple-health-ai-usage-study-launch</link><guid isPermaLink="true">https://dataphoenix.info/news/apple-health-ai-usage-study-launch</guid><description>Apple launched a virtual U.S. study for up to 200,000 adults to examine how people use and evaluate conversational health AI.</description><pubDate>Mon, 16 Nov 2026 03:41:48 GMT</pubDate><content:encoded>&lt;p&gt;Apple launched a U.S. virtual research study and dedicated app to examine how adults use and evaluate conversational health AI. The project is being conducted with OpenEvidence, whose clinical AI technology underpins the Apple-designed research experience.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://www.apple.com/newsroom/2026/11/apple-launches-research-study-to-explore-the-use-of-ai-for-health&quot; rel=&quot;noopener noreferrer&quot;&gt;Apple Health AI Usage Study&lt;/a&gt; is open to as many as 200,000 eligible U.S. participants. Enrollment requires users to be at least 21 years old and have an Apple Intelligence-enabled iPhone 15 Pro or newer running iOS 26 or later.&lt;/p&gt;
&lt;p&gt;Using the app, participants can ask everyday health questions, receive responses with links to peer-reviewed clinical citations, rate those responses and use tools intended to help them prepare for medical visits. Participants can also choose whether to share HealthKit information such as resting heart rate, activity, sleep or lab results. Researchers will examine whether adding that personal data changes the relevance and clarity of answers.&lt;/p&gt;
&lt;p&gt;Apple said the study will examine trust, response quality, comprehension, everyday health actions and preparation for clinical visits. The company said the work is intended to inform responsible health-AI design at Apple and across the industry, but the opened sources give no timetable for publishing results.&lt;/p&gt;
&lt;p&gt;Apple also said study data will not be used to train AI models and that participants control whether they share HealthKit metrics or survey responses. The company said it has put security measures in place for study data and participant confidentiality. Those claims have not been independently audited in the sources reviewed. The &lt;a href=&quot;https://apps.apple.com/us/app/apple-health-ai-study-app/id6800060365&quot; rel=&quot;noopener noreferrer&quot;&gt;App Store listing&lt;/a&gt; says the app may collect a user identifier linked to identity for analytics and app functionality, along with diagnostics not linked to identity. The privacy label is based on the developer’s disclosure.&lt;/p&gt;
&lt;p&gt;Apple said the app includes guardrails designed to detect certain concerning questions and direct users to a provider or other resources. The company says the app and study do not provide medical advice, diagnosis or treatment and are not a replacement for medical professionals. Apple did not identify the categories covered by those guardrails or publish performance measures for them in its announcement.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>apple</category><category>health-ai</category><category>openevidence</category><category>research-study</category><category>conversational-ai</category></item><item><title>OpenAI says GPT-6 Astra helped develop a candidate lattice proof for the Standard Model</title><link>https://dataphoenix.info/news/openai-astra-lattice-standard-model</link><guid isPermaLink="true">https://dataphoenix.info/news/openai-astra-lattice-standard-model</guid><description>OpenAI published physicist Nathaniel Craig&apos;s candidate lattice construction for the Standard Model and credited GPT-6 Astra with helping develop the detailed proof, while stressing that scientific scrutiny and practical work remain.</description><pubDate>Mon, 16 Nov 2026 01:29:55 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://openai.com/index/step-toward-defining-the-standard-model&quot; rel=&quot;noopener noreferrer&quot;&gt;OpenAI published&lt;/a&gt; a guest research post by UC Santa Barbara physicist Nathaniel Craig on November 15 describing a candidate nonperturbative lattice definition of the Standard Model. The post says GPT-6 Astra helped develop the detailed proof. OpenAI and Craig frame the work as a candidate that still needs expert scientific dialogue, not a settled definition ready for practical simulations.&lt;/p&gt;
&lt;p&gt;The construction proposes a way to represent chiral fermions — particles whose left- and right-handed forms interact differently — on a spacetime lattice while preserving gauge symmetry. A lattice replaces continuous spacetime with a finite grid, allowing physicists to study strongly interacting theories without relying on perturbative approximations. Craig’s &lt;a href=&quot;https://arxiv.org/abs/2610.09093&quot; rel=&quot;noopener noreferrer&quot;&gt;October preprint&lt;/a&gt; describes a gauge-invariant fermion measure at finite lattice spacing for admissible gauge fields in nonabelian chiral gauge theories, including the Standard Model.&lt;/p&gt;
&lt;p&gt;Craig says Astra helped turn his refinement strategy into a concrete construction. By his account, the model proposed a way to carry gauge fields to progressively finer grids, established convergence within the proposed argument and led development of the detailed proof. The available evidence does not include the prompts, full model interaction history or a reproducible audit trail that would independently verify that division of work. For context on another claimed application of the model, DataPhoenix has covered &lt;a href=&quot;https://dataphoenix.info/news/openai-astra-physical-engineering-mazda-research&quot;&gt;Astra’s use in physical-engineering research&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Craig also says lattice theorists Maarten Golterman, Andreas Kronfeld and Yigal Shamir questioned and reviewed the developing argument. The available sources, however, do not include independent endorsements from them.&lt;/p&gt;
&lt;p&gt;The paper reports that 16 Lean 4 modules compiled and that kernel replay accepted the formal proofs under the assumptions recorded in the modules. Craig cautions that those checks establish only the encoded mathematical statements under those assumptions; they do not validate the construction’s physical claims by themselves. The preprint has not been independently validated in the opened sources, which also do not establish peer-review or journal-acceptance status.&lt;/p&gt;
&lt;p&gt;Several scientific and computational steps remain. The sources leave unresolved whether the construction satisfies reflection positivity at finite lattice spacing, a property needed to connect the lattice formulation to a physically consistent quantum theory. Craig also identifies a need for numerical admissibility bounds, realistic lattice simulations and practical implementation methods. He describes the proposal as an answer in principle rather than in practice.&lt;/p&gt;
&lt;p&gt;The construction does not define the interacting Standard Model in continuous spacetime. Craig’s paper leaves that continuum problem open, while the OpenAI post says further scientific dialogue is needed to determine whether the prescription has all the desired properties.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>openai</category><category>gpt-6-astra</category><category>standard-model</category><category>lattice-gauge-theory</category><category>ai-for-science</category></item><item><title>Bessent calls for outside CEO and leadership shakeup at Nvidia</title><link>https://dataphoenix.info/news/bessent-nvidia-leadership-shakeup</link><guid isPermaLink="true">https://dataphoenix.info/news/bessent-nvidia-leadership-shakeup</guid><description>Treasury Secretary Scott Bessent urged Nvidia to name an outsider as its next CEO and sought answers about its compliance culture within 10 days, a narrower recommendation than the company’s announced search.</description><pubDate>Sun, 15 Nov 2026 19:42:05 GMT</pubDate><content:encoded>&lt;p&gt;Treasury Secretary Scott Bessent said Nvidia should hire its next chief executive from outside the company and called for a “dramatic shakeup” in leadership following Jensen Huang’s departure. His recommendation would narrow Nvidia’s announced search, which includes both internal and external candidates.&lt;/p&gt;
&lt;p&gt;Bessent also said the company’s board and outside counsel should provide answers within 10 days about Nvidia’s sales and compliance culture. Nvidia &lt;a href=&quot;https://dataphoenix.info/news/nvidia-jensen-huang-resignation-colette-kress-interim-ceo&quot;&gt;named Colette Kress interim CEO&lt;/a&gt; during the leadership transition.&lt;/p&gt;
&lt;p&gt;Bessent alleged that Nvidia had developed a culture of noncompliance. He also questioned whether apparently illicit China-related sales had boosted revenue and earnings. Those statements are allegations, not established findings. None of the official records reviewed for this article determines that Nvidia violated export controls or improperly inflated its results, and the available sources do not identify a legal or administrative basis for Bessent’s 10-day request.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://www.sec.gov/Archives/edgar/data/1045810/000104581026000082/nvda-20261111.htm&quot; rel=&quot;noopener noreferrer&quot;&gt;Nvidia’s Form 8-K says&lt;/a&gt; the board accepted Huang’s resignation as president, CEO and director on Nov. 11, effective immediately. &lt;a href=&quot;https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition&quot; rel=&quot;noopener noreferrer&quot;&gt;The company said&lt;/a&gt; Kress would retain her finance responsibilities while serving as interim CEO and that the board had started its search for a permanent successor.&lt;/p&gt;
&lt;p&gt;The leadership change is unfolding alongside &lt;a href=&quot;https://dataphoenix.info/news/nvidia-board-foreign-intelligence-investigation&quot;&gt;a separate board inquiry into reported foreign-intelligence targeting of Huang&lt;/a&gt;. &lt;a href=&quot;https://nvidianews.nvidia.com/news/statement-from-nvidia-board-of-directors&quot; rel=&quot;noopener noreferrer&quot;&gt;Nvidia’s board said&lt;/a&gt; it was investigating those reports while saying it had no reason to believe national security or company secrets were compromised. The independent directors later said they had retained outside counsel for that inquiry. The opened Nvidia statements do not disclose findings connecting the board inquiry to the export-control and sales questions raised by Bessent.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>nvidia</category><category>scott-bessent</category><category>jensen-huang</category><category>ceo-transition</category><category>export-controls</category></item><item><title>CMU researchers give everyday objects expressive OATs avatars</title><link>https://dataphoenix.info/news/cmu-oats-object-agent-twins-uist-2026</link><guid isPermaLink="true">https://dataphoenix.info/news/cmu-oats-object-agent-twins-uist-2026</guid><description>OATs, a Carnegie Mellon research prototype, turns live readings from sensor-equipped everyday objects into expressive avatar cues instead of raw measurements.</description><pubDate>Sun, 15 Nov 2026 15:43:59 GMT</pubDate><content:encoded>&lt;p&gt;Carnegie Mellon University researchers Daehwa Kim and Chris Harrison &lt;a href=&quot;https://www.futureinterfaces.com/research/2026/oats&quot; rel=&quot;noopener noreferrer&quot;&gt;presented Object-Agent Twins, or OATs, at UIST 2026&lt;/a&gt; as a prototype toolkit for giving sensor-equipped everyday objects simple character-like states. Instead of showing only raw measurements, the system converts live readings into visible cues.&lt;/p&gt;
&lt;p&gt;OATs maps temperature, humidity, light, motion, occupancy and soil-moisture data to an avatar’s needs, moods and behaviors. The researchers designed the cues to make changes in an object’s condition easier to notice and interpret.&lt;/p&gt;
&lt;p&gt;In the &lt;a href=&quot;https://daehwakim.com/oat/index.html&quot; rel=&quot;noopener noreferrer&quot;&gt;documented browser-based setup flow&lt;/a&gt;, a user names and describes the host object, selects its sensors, and reviews generated positive and negative behaviors. The system prepares emotion assets, then restarts the device with its new character.&lt;/p&gt;
&lt;p&gt;Demonstrations placed OATs on a plant, an office refrigerator and a space heater. The plant character can reflect water, temperature, humidity and light conditions. The refrigerator example uses motion data, while the heater combines temperature and occupancy sensing.&lt;/p&gt;
&lt;p&gt;The prototype supports internal air-temperature and humidity sensing, plus external motion, water-or-soil, light-intensity and occupancy sensors. An OAT updates when its environment changes or once per hour. A web status page explains its current state.&lt;/p&gt;
&lt;p&gt;Public pages maintained by &lt;a href=&quot;https://www.chrisharrison.net/index.php/Research/OATs&quot; rel=&quot;noopener noreferrer&quot;&gt;Harrison&lt;/a&gt; and &lt;a href=&quot;https://daehwakim.com/&quot; rel=&quot;noopener noreferrer&quot;&gt;Kim&lt;/a&gt; list paper and demo honorable-mention recognition for the UIST presentation.&lt;/p&gt;
&lt;p&gt;The official materials call OATs a toolkit and a first prototype, not a consumer product release. They list no retail price, availability date or manufacturing plan. According to the project page, the first prototype runs for about three to four days without the sound-level sensor and roughly one day when that sensor is active.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>oats</category><category>object-agent-twins</category><category>carnegie-mellon-university</category><category>uist-2026</category><category>human-computer-interaction</category><category>sensor-avatars</category></item><item><title>South Korea tax-exam top scorer admits using AI to cheat</title><link>https://dataphoenix.info/news/south-korea-tax-exam-ai-cheating-admission</link><guid isPermaLink="true">https://dataphoenix.info/news/south-korea-tax-exam-ai-cheating-admission</guid><description>South Korea’s top scorer in the national tax-accountant exam admitted using a miniature camera, a smartphone and generative AI to cheat, prompting a police investigation into the qualification test.</description><pubDate>Sun, 15 Nov 2026 15:38:41 GMT</pubDate><content:encoded>&lt;p&gt;South Korea’s top scorer in the national tax-accountant qualification exam admitted using a miniature camera, a smartphone and generative AI to cheat across both stages of the test. The examinee submitted a written confession through a lawyer on Nov. 15, and the Seoul Metropolitan Police Agency assigned the case to a specialized investigative unit.&lt;/p&gt;
&lt;p&gt;By the examinee’s account, he photographed questions with the camera, sent the images to a generative AI system and viewed its answers on a smartphone. The available evidence does not identify the camera or AI service, explain how the phone was concealed or establish the data path. For now, the described workflow remains an admission rather than a police-reconstructed finding.&lt;/p&gt;
&lt;p&gt;The examinee answered 159 of 160 questions correctly in the April first-stage test, receiving 397.5 points out of 400. He scored 663 out of 800 in the written second stage in July and ranked first overall. The Human Resources Development Service of Korea administers the national qualification on the government’s behalf.&lt;/p&gt;
&lt;p&gt;Two days before the admission, &lt;a href=&quot;https://webzine.hrdkorea.or.kr/section/press/view?id=13802&amp;amp;page=1&quot; rel=&quot;noopener noreferrer&quot;&gt;HRD Korea said its review had found no concrete evidence supporting suspicions involving smart glasses&lt;/a&gt;. The agency said all four proctors who supervised the examinee across the two stages reported no suspicious conduct or irregularity. Each examination room had two proctors. The examinee sat in the center of the second row in the first stage and the center of the front row in the second, but the rented school classroom had no CCTV footage available for review.&lt;/p&gt;
&lt;p&gt;HRD Korea said smartphones, smartwatches and other electronic or communications devices were prohibited. It also said identity checks, device checks, announcements and proctor training had been conducted normally. The statement did not establish which check failed, whether radio-frequency detection was used in either room or whether the admitted setup would have been detected by measures the agency was preparing.&lt;/p&gt;
&lt;p&gt;The agency had already formed an AI-cheating response task force and &lt;a href=&quot;https://webzine.hrdkorea.or.kr/section/press/view?id=13745&quot; rel=&quot;noopener noreferrer&quot;&gt;conducted a simulation involving smart-glasses cheating on Sept. 29&lt;/a&gt;. The drill included examiner training, belongings inspections and radio-frequency detection in an exam-like setting. On Nov. 13, HRD Korea said it planned to expand its 529 installed CCTV units, introduce radio-frequency detectors after field validation and strengthen checks and training for emerging electronic devices.&lt;/p&gt;
&lt;p&gt;Police said the investigation remained active and would proceed under established procedures. The examinee’s admission had not been tested in court or confirmed in a completed police finding.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>south-korea</category><category>exam-security</category><category>ai-cheating</category><category>generative-ai</category><category>professional-qualifications</category></item><item><title>Italy fines Meta €805 million over Facebook Android gallery access</title><link>https://dataphoenix.info/news/italy-meta-facebook-android-gallery-fine</link><guid isPermaLink="true">https://dataphoenix.info/news/italy-meta-facebook-android-gallery-fine</guid><description>Italy’s privacy regulator ordered Meta to obtain separate consent and keep a Facebook Android gallery-suggestion feature off by default after imposing an €805 million fine.</description><pubDate>Sun, 15 Nov 2026 11:30:35 GMT</pubDate><content:encoded>&lt;p&gt;Italy’s data protection authority &lt;a href=&quot;https://garanteprivacy.it/home/docweb/-/docweb-display/docweb/10321080&quot; rel=&quot;noopener noreferrer&quot;&gt;fined Meta Platforms Ireland €805 million&lt;/a&gt; after finding that Facebook’s Android app accessed photos, videos and related metadata on Italian users’ phones without specific consent or adequate disclosure.&lt;/p&gt;
&lt;p&gt;Meta has 60 days to introduce a separate, feature-specific consent request explaining the feature’s purpose, its access to the phone gallery and how users can disable it. The feature must remain off by default, including for people who previously granted the Android system permission. Meta must tell the regulator within 30 days which initiatives it has begun and within 90 days which measures it has adopted.&lt;/p&gt;
&lt;p&gt;The case centers on “Get suggestions from gallery when browsing Facebook,” a feature available since 2023 that offers photos and videos from a phone as potential Facebook posts while a user browses the app. The authority opened its investigation in July 2025 after a report from the Bolzano Consumer Protection Centre, then conducted a technical analysis.&lt;/p&gt;
&lt;p&gt;Investigators found that Facebook sought gallery access through a single permission covering the whole app and several functions with different purposes. The request followed a generic notice about enabling other features, but the notice did not tell users they were also authorizing gallery suggestions. The suggestion feature was enabled by default.&lt;/p&gt;
&lt;p&gt;When granted in full, the permission can let the app read the device’s entire photo archive on an ongoing basis, the authority said. Neither users nor the regulator could verify which information was actually used. The decision does not establish that every item in a user’s gallery was used.&lt;/p&gt;
&lt;p&gt;The regulator said the conduct violated Article 122 of Italy’s Privacy Code, which implements the EU ePrivacy Directive. Accessing information stored on a smartphone requires free and informed consent regardless of whether the data is transmitted or harm occurs, it said. The authority also rejected Meta’s argument that the Irish regulator had jurisdiction, saying the ePrivacy framework does not include the GDPR’s one-stop-shop mechanism.&lt;/p&gt;
&lt;p&gt;About 20 million Android users in Italy were involved, according to the authority. It said the penalty was about 0.47% of Meta’s global revenue and cited the number of people affected, the duration of the infringement, a 2019 measure against the company and limited cooperation during the investigation. Changes made after the complaint counted as mitigation, but the regulator said they came late and benefited only users who updated the app.&lt;/p&gt;
&lt;p&gt;The proceeding did not assess whether Meta used gallery content to train artificial intelligence. Meta denies doing so. The regulator said it may separately examine a gallery cloud-processing feature launched in 2026.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>meta</category><category>facebook</category><category>android</category><category>privacy</category><category>italy</category><category>data-consent</category></item><item><title>Perplexity releases Decider v1.3 models with locate extraction</title><link>https://dataphoenix.info/news/perplexity-decider-v1-3-release</link><guid isPermaLink="true">https://dataphoenix.info/news/perplexity-decider-v1-3-release</guid><description>Perplexity released Decider v1.3 and Decider Flash v1.3, adding a locate function that returns matching text spans while retaining the models’ structured decision outputs.</description><pubDate>Sun, 15 Nov 2026 09:29:32 GMT</pubDate><content:encoded>&lt;p&gt;Perplexity &lt;a href=&quot;https://www.perplexity.ai/hub/blog/perplexity-decider&quot; rel=&quot;noopener noreferrer&quot;&gt;released Decider v1.3 and Decider Flash v1.3&lt;/a&gt; on Nov. 12 with a locate function that extracts matching passages from an input. The decision models return structured probabilities rather than free-form text, giving developers another option for classification, routing, moderation, ranking and evaluation.&lt;/p&gt;
&lt;p&gt;The models are available through Perplexity’s &lt;a href=&quot;https://dataphoenix.info/news/perplexity-ultrafast-decisions-api-connector-controls&quot;&gt;Decisions API&lt;/a&gt; and OpenRouter. Perplexity says their downloadable weights are available on Hugging Face under the Apache 2.0 license.&lt;/p&gt;
&lt;p&gt;The new locate question type returns verbatim text with start and end character offsets and a score. When it finds no match, it returns an empty span list. Locate can be combined with yes-or-no, multiple-choice and ordered-score questions about the same input, allowing an application to classify content and extract supporting text in one request. Perplexity’s &lt;a href=&quot;https://docs.perplexity.ai/docs/decisions/quickstart&quot; rel=&quot;noopener noreferrer&quot;&gt;API documentation&lt;/a&gt; says the feature is limited to the v1.3 models, accepts text state without images and has been evaluated on inputs up to 32,000 tokens, though the API accepts longer inputs.&lt;/p&gt;
&lt;p&gt;Applications send content and one or more questions, then a Decider model returns probabilities that software can compare against thresholds. The &lt;a href=&quot;https://huggingface.co/perplexity-ai/pplx-decider-v1.3&quot; rel=&quot;noopener noreferrer&quot;&gt;Decider v1.3 model card&lt;/a&gt; and &lt;a href=&quot;https://huggingface.co/perplexity-ai/pplx-decider-flash-v1.3&quot; rel=&quot;noopener noreferrer&quot;&gt;Flash model card&lt;/a&gt; display an Apache 2.0 license but say repository access requires authentication.&lt;/p&gt;
&lt;p&gt;Perplexity says it trained the models by converting examples from hundreds of public datasets into a common format of state, questions, options and targets. The three-stage process uses generalist full fine-tuning, targeted low-rank adaptation and a vision-specific low-rank adaptation step. For Flash, the company used the larger model’s predicted probability distribution as the training target. It said this distillation method generalized better to held-out data than cross-entropy training on the same data.&lt;/p&gt;
&lt;p&gt;V1.3 also uses bidirectional full-attention layers and an ‘isolated questions’ pattern, under which each question attends to the shared state and its own tokens. Perplexity said it chose that pattern for performance and answer stability. Locate adds a special token, span-prediction head and separate adapter while leaving the original backbone and decision head frozen. The company said it kept the locate adapter unmerged because merging it degraded ordinary decision requests.&lt;/p&gt;
&lt;p&gt;Perplexity said it removed training samples that could overlap with its evaluations and conducted bias audits and removals. Those controls, the training corpus and the reported results have not been independently audited or replicated.&lt;/p&gt;
&lt;p&gt;The company evaluated the models with external leaderboards and an internal suite of more than 200 zero-shot tasks. It reported 91.6% mean accuracy for Decider v1.3 across 63 classification tasks whose reference answers were relabeled by a frontier model, 6.54 percentage points above its production classifiers. Across 17 out-of-domain locate tasks, Perplexity reported average scores of 69.0 for Decider v1.3 and 63.6 for Flash, compared with 43.3 for Liquid d1, 41.5 for Clairic and 33.6 for GLiNER large v2.5; each task used its own primary metric.&lt;/p&gt;
&lt;p&gt;Perplexity also tested answer stability on 372 multiple-choice questions. It said Decider v1.3 changed its semantic answer on 0.27% of questions under each of three option or label transformations, compared with 2.69% to 4.84% for Jev 1.14, and showed no changes when identical inputs were repeated.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>perplexity</category><category>decider-v1-3</category><category>decision-models</category><category>text-extraction</category><category>open-weight-models</category></item><item><title>Cohere adds third-party open models to Model Vault</title><link>https://dataphoenix.info/news/cohere-model-vault-open-models</link><guid isPermaLink="true">https://dataphoenix.info/news/cohere-model-vault-open-models</guid><description>Cohere has added outside open-model families to Model Vault, with a broader standard catalog across three regions and a smaller encrypted catalog currently limited to the United States.</description><pubDate>Sun, 15 Nov 2026 08:35:00 GMT</pubDate><content:encoded>&lt;p&gt;Cohere has &lt;a href=&quot;https://cohere.com/blog/the-worlds-leading-open-models-fully-under-your-control&quot; rel=&quot;noopener noreferrer&quot;&gt;added third-party open models to Model Vault&lt;/a&gt;, expanding its managed inference service beyond Cohere’s own models.&lt;/p&gt;
&lt;p&gt;At the November 3 launch, Cohere’s standard Model Vault matrix listed DeepSeek, GLM, MiMo, Gemma, Nemotron, GPT-OSS, Phi and Tencent’s Hy3 as available. The Model Vault Encrypted column listed only Gemma, GPT-OSS and Phi among those third-party families, making the encrypted offering a narrower catalog rather than a security mode available for every standard-vault model.&lt;/p&gt;
&lt;p&gt;The catalogs are evolving. &lt;a href=&quot;https://docs.cohere.com/docs/model-vault/standard/supported-models&quot; rel=&quot;noopener noreferrer&quot;&gt;Cohere’s current standard-vault documentation&lt;/a&gt; lists DeepSeek, GLM, MiMo, Nemotron, Gemma, GPT-OSS, Phi, Qwen and the Mistral AI families Ministral, Devstral and Leanstral. &lt;a href=&quot;https://docs.cohere.com/docs/model-vault/encrypted/supported-models&quot; rel=&quot;noopener noreferrer&quot;&gt;Cohere’s current encrypted-vault list&lt;/a&gt; includes Gemma, GPT-OSS, Phi, Qwen, Ministral and Devstral. Cohere says exact model and GPU availability can vary by region and change over time.&lt;/p&gt;
&lt;p&gt;Standard Model Vault is &lt;a href=&quot;https://docs.cohere.com/v2/docs/model-vault/regions&quot; rel=&quot;noopener noreferrer&quot;&gt;available in the United States, European Union and Canada&lt;/a&gt;, while Model Vault Encrypted currently runs only in the United States; Cohere describes EU and Canadian encrypted serving as coming soon. Cohere says customers choose a region when creating a vault, every model and inference request stays in that region, and the region cannot later be changed. The company presents that setup as helping customers address residency and regulatory requirements, not as a compliance guarantee.&lt;/p&gt;
&lt;p&gt;Cohere says a &lt;a href=&quot;https://docs.cohere.com/docs/model-vault/standard&quot; rel=&quot;noopener noreferrer&quot;&gt;standard vault&lt;/a&gt; uses dedicated, logically isolated infrastructure, protects data in transit and at rest, and has zero data retention enabled by default. It says &lt;a href=&quot;https://docs.cohere.com/docs/model-vault/encrypted&quot; rel=&quot;noopener noreferrer&quot;&gt;Model Vault Encrypted&lt;/a&gt; adds end-to-end encryption, confidential computing, hardware-backed trusted execution environments and remote attestation so prompts and responses are exposed in plaintext only inside the protected environment, not to Cohere or the cloud operator. Third-party models are served through an OpenAI-compatible Chat Completions API at the customer’s vault endpoint.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>cohere</category><category>model-vault</category><category>open-models</category><category>managed-inference</category><category>confidential-computing</category></item><item><title>Octen expands agent search with Pulse and a $20 plan</title><link>https://dataphoenix.info/news/octen-agent-search-pulse-20-plan</link><guid isPermaLink="true">https://dataphoenix.info/news/octen-agent-search-pulse-20-plan</guid><description>Octen expanded its search service for AI agents, adding the Pulse page-finding API and image search alongside a $20 monthly unlimited Web Search plan for personal accounts.</description><pubDate>Sun, 15 Nov 2026 08:30:46 GMT</pubDate><content:encoded>&lt;p&gt;Octen expanded its search service for AI agents, adding the Pulse page-finding API and image search alongside a $20-a-month unlimited Web Search plan. For personal accounts, the additions pair a flat-price live-web option with a lower-cost retrieval system built to route queries to specific pages.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://octen.ai/pricing&quot; rel=&quot;noopener noreferrer&quot;&gt;Standard Search Plan&lt;/a&gt; includes unlimited Web Search with no overage fees and 1,000 Image Search calls a month. Octen says requests run on shared capacity and may queue at busy times. Full content, additional image searches and other APIs remain pay-as-you-go charges. Customers that need guaranteed throughput must use a separate QPS plan.&lt;/p&gt;
&lt;p&gt;Octen made &lt;a href=&quot;https://octen.ai/blog/introducing-octen-pulse-search&quot; rel=&quot;noopener noreferrer&quot;&gt;Pulse Search&lt;/a&gt; generally available on October 28, with lite-mode pricing starting at $0.10 per 1,000 calls. Lite mode returns a URL and title, while standard mode can add metadata, relevant passages, full content and images.&lt;/p&gt;
&lt;p&gt;Octen’s roughly 10-millisecond figure is the median server-side processing time, or P50, for Pulse lite mode on previously unseen queries. It is not an end-to-end performance measure. Network transit, content retrieval and any answering model add time of their own.&lt;/p&gt;
&lt;p&gt;Pulse uses generative retrieval: its model emits a two-token identifier restricted to a table of existing URLs, then maps the identifier to a page. Octen positions the system for navigational searches for familiar destinations. Pages added after training and rarely visited pages remain limitations, so the company directs fresher or broader queries to its live-index Web Search product.&lt;/p&gt;
&lt;p&gt;In &lt;a href=&quot;https://arxiv.org/abs/2610.35256&quot; rel=&quot;noopener noreferrer&quot;&gt;Octen’s technical paper&lt;/a&gt;, the company reports 59.6% top-1 accuracy across its navigational-traffic evaluation, versus 46.5% for a click-enhanced dense-retrieval baseline. The paper also says embedding retrieval performs better on rarely visited pages. These are Octen’s results, not an independent benchmark.&lt;/p&gt;
&lt;p&gt;Image Search accepts text or a reference image. Outside plan allowances, Octen lists pay-as-you-go pricing of $2 per 1,000 calls.&lt;/p&gt;
&lt;p&gt;The hosted Pulse API and the downloadable model come with different usage terms. Octen allows commercial use through the hosted API, but the &lt;a href=&quot;https://huggingface.co/Octen/Octen-WebGR-4B&quot; rel=&quot;noopener noreferrer&quot;&gt;Octen-WebGR-4B model card&lt;/a&gt; licenses the downloadable weights and URL table under CC BY-NC 4.0 for noncommercial use. The accompanying inference code uses the Apache 2.0 license.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>octen</category><category>pulse-search</category><category>agent-search</category><category>web-search</category><category>image-search</category></item><item><title>World Labs releases Frameboy preview for real-time Atlas exploration</title><link>https://dataphoenix.info/news/world-labs-frameboy-atlas-realtime-preview</link><guid isPermaLink="true">https://dataphoenix.info/news/world-labs-frameboy-atlas-realtime-preview</guid><description>World Labs released Frameboy, a research preview that streams frames from an optimized Atlas model on cloud GPUs to a browser for interactive exploration.</description><pubDate>Sun, 15 Nov 2026 05:36:12 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://www.worldlabs.ai/blog/frameboy-atlas-in-real-time&quot; rel=&quot;noopener noreferrer&quot;&gt;World Labs released Frameboy&lt;/a&gt; on Nov. 5 as a research preview that lets users explore Atlas-generated worlds interactively in a browser. The demo streams frames from cloud GPUs as the user controls the camera, turning the company’s world-model output into a continuously navigable scene.&lt;/p&gt;
&lt;p&gt;The company says Frameboy can stream at up to 15 frames per second. That is a company performance claim, not an independently reproduced benchmark. The announcement did not identify the cloud GPU, resolution, scene complexity or measurement method behind it.&lt;/p&gt;
&lt;p&gt;Frameboy runs an Atlas variant that World Labs says was optimized for inference speed and efficiency while retaining the model’s ability to synthesize new views. According to the company, Atlas can generate a scene from as little as one image or a text caption. The preview can also navigate real-world scenes built from images captured from multiple viewpoints.&lt;/p&gt;
&lt;p&gt;World Labs also added Frameboy’s streaming mode to Viewfinder, its browser-based scene-building tool. A user can build a scene frame by frame in Viewfinder and then launch a live Frameboy session as an overlay. The company says Frameboy offers a curated set of scenes, while Viewfinder accepts a user’s own inputs. Viewfinder was in closed beta when the preview was announced.&lt;/p&gt;
&lt;p&gt;That speed comes with a constraint. World Labs says Frameboy requires a powerful cloud GPU server and is not cheap compared with local rendering. The company did not provide a price, hardware configuration or cost-per-frame comparison, so the claim cannot be quantified from the announcement.&lt;/p&gt;
&lt;p&gt;Frameboy is separate from the broader Atlas release that followed one week later. A separate &lt;a href=&quot;https://dataphoenix.info/news/world-labs-atlas-tools-api-release&quot;&gt;Atlas tools and developer APIs release&lt;/a&gt; expanded World Labs’ model-and-product offering beyond the real-time preview. The &lt;a href=&quot;https://worldlabs.ai/blog&quot; rel=&quot;noopener noreferrer&quot;&gt;World Labs research index&lt;/a&gt; dates “Atlas: Releasing our next generation world model” to Nov. 12, and the &lt;a href=&quot;https://worldlabs.ai/atlas&quot; rel=&quot;noopener noreferrer&quot;&gt;Atlas product page&lt;/a&gt; covers world generation and reconstruction alongside Viewfinder, the Chisel scene-building tool, splat-to-mesh conversion, relighting, developer APIs and integrations including Blender. The Nov. 5 Frameboy item is specifically a real-time research-preview interface powered by an optimized Atlas variant.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>world-labs</category><category>frameboy</category><category>atlas</category><category>world-models</category><category>generative-ai</category></item><item><title>OpenAI&apos;s BioBasicsBench tests wet-lab AI agents</title><link>https://dataphoenix.info/news/openai-biobasicsbench-wet-lab-benchmark</link><guid isPermaLink="true">https://dataphoenix.info/news/openai-biobasicsbench-wet-lab-benchmark</guid><description>OpenAI&apos;s 100-task BioBasicsBench tests whether AI agents can interpret raw wet-lab evidence, catch experimental errors and recognize when the data cannot support a conclusion.</description><pubDate>Sun, 15 Nov 2026 05:32:14 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://openai.com/index/measuring-how-ai-agents-handle-wet-lab-science&quot; rel=&quot;noopener noreferrer&quot;&gt;OpenAI introduced BioBasicsBench&lt;/a&gt;, a 100-task benchmark developed with Ligo Biosciences to test how AI agents interpret raw evidence from routine wet-lab work. OpenAI says the evaluation measures whether agents can check controls, identify experimental errors and recognize when the evidence is insufficient for a conclusion.&lt;/p&gt;
&lt;p&gt;The tasks cover protein methods, imaging and cytometry, quantification and assays, nucleic acid methods, and molecular cloning. Inputs range from gels and blots to protocols and notes, instrument exports, bench photographs, microscopy, sequencing files, videos, colony plates and reference sheets.&lt;/p&gt;
&lt;p&gt;OpenAI says scientists drew the questions from their own experiments and paired each one with an expert-written rubric. Every task was reviewed by another Ligo scientist, an external PhD-level wet-lab scientist and, finally, an internal reviewer.&lt;/p&gt;
&lt;p&gt;The benchmark targets a gap between published science and laboratory work in progress. Paper-ready figures are selected for communication, so they may omit failed runs, ambiguous images, conflicting observations and the troubleshooting clues found in raw evidence. An agent working with that material must determine not only what a measurement appears to show, but whether the controls and experimental conditions make the conclusion defensible.&lt;/p&gt;
&lt;p&gt;BioBasicsBench focuses on laboratory evidence, while a separate &lt;a href=&quot;https://dataphoenix.info/news/openai-lifeval-benchmark&quot;&gt;OpenAI LifeVal benchmark covers everyday AI tasks&lt;/a&gt;. OpenAI awards criterion-level partial credit for conclusions, measurements, reasoning and recognition of insufficient evidence. Its reported mean reward is the average share of available rubric credit, not the percentage of tasks solved completely.&lt;/p&gt;
&lt;p&gt;The company evaluated seven OpenAI models across 37 combinations of model and reasoning effort. With five attempts on every task, the evaluation produced 18,500 graded results. At xhigh reasoning effort, OpenAI reports that GPT-6 Astra received 66.7% mean reward, with a 95% confidence interval of 60.3% to 73.1%. GPT-5.6 Sol received 43.8%, with an interval of 37.6% to 50.0%.&lt;/p&gt;
&lt;p&gt;These are provider-run results, not an independent replication. The opened evidence does not include the complete tasks, raw assets, rubrics, scoring implementation or per-attempt outputs needed to reproduce them. The comparison is also limited to OpenAI models and includes no human wet-lab baseline.&lt;/p&gt;
&lt;p&gt;OpenAI’s examples show improvement alongside persistent failure. Astra recognized that overlapping fluorescence channels could not support a cell-death conclusion, spotted a negative-control plate using the wrong antibiotic and identified a microscopy sequence as a z-stack rather than a biological disappearance.&lt;/p&gt;
&lt;p&gt;OpenAI also found Astra overriding controls in a colony-PCR task, discounting contradictory microscopy observations and overcounting colonies by about 12%. In the colony-PCR example, the model grouped bands visually but reversed their experimental identities. It selected colonies with the wrong plasmid and would have sent the incorrect plasmid into downstream purification. OpenAI cautions that higher benchmark scores do not make current models reliable lab judges.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>openai</category><category>biobasicsbench</category><category>ai-benchmarks</category><category>wet-lab-science</category><category>scientific-ai</category></item><item><title>AWS adds NVIDIA Nemotron 3.5 Super VL to SageMaker JumpStart</title><link>https://dataphoenix.info/news/nvidia-nemotron-3-5-super-vl-sagemaker-jumpstart</link><guid isPermaLink="true">https://dataphoenix.info/news/nvidia-nemotron-3-5-super-vl-sagemaker-jumpstart</guid><description>AWS has added NVIDIA Nemotron 3.5 Super VL to SageMaker JumpStart, giving customers a managed path to deploy NVFP4 and BF16 variants with fixed instance mappings and serving limits.</description><pubDate>Sun, 15 Nov 2026 04:43:42 GMT</pubDate><content:encoded>&lt;p&gt;AWS has &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/nvidia-nemotron-3-5-super-vl-is-now-available-in-amazon-sagemaker-jumpstart&quot; rel=&quot;noopener noreferrer&quot;&gt;made NVIDIA Nemotron 3.5 Super VL available through Amazon SageMaker JumpStart&lt;/a&gt;, giving customers a managed way to deploy the multimodal model from the JumpStart catalog in SageMaker Studio or with SageMaker Python SDK v3. The SDK route requires version 3.25 or later. The catalog also hosts an &lt;a href=&quot;https://dataphoenix.info/news/aws-openfold3-sagemaker-jumpstart-preview&quot;&gt;OpenFold3 preview in SageMaker JumpStart&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;For NVFP4, JumpStart defaults to &lt;code&gt;ml.g7.24xlarge&lt;/code&gt; and also supports &lt;code&gt;ml.g7.48xlarge&lt;/code&gt;; both use NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. The BF16 checkpoint runs on &lt;code&gt;ml.p5en.48xlarge&lt;/code&gt; with NVIDIA H200 GPUs. AWS lists the optimized NVFP4 checkpoint at about 75 GB and the full-precision BF16 reference weights at about 249 GB, positioning BF16 for customization and evaluation.&lt;/p&gt;
&lt;p&gt;Nemotron 3.5 Super VL is a 120-billion-parameter hybrid mixture-of-experts vision-language model with 12 billion active parameters. It accepts text, images and video and returns text. The JumpStart addition is a deployment and distribution update, not the model’s debut: &lt;a href=&quot;https://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning&quot; rel=&quot;noopener noreferrer&quot;&gt;NVIDIA introduced the model and its agent-focused multimodal capabilities&lt;/a&gt; on October 15, before AWS added the managed endpoint path on November 9.&lt;/p&gt;
&lt;p&gt;JumpStart’s serving presets are narrower than the model’s underlying capability. NVIDIA’s &lt;a href=&quot;https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16&quot; rel=&quot;noopener noreferrer&quot;&gt;BF16 model card&lt;/a&gt; and &lt;a href=&quot;https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4&quot; rel=&quot;noopener noreferrer&quot;&gt;NVFP4 model card&lt;/a&gt; describe context support up to 1 million tokens. AWS, however, caps the combined prompt and completion length at 520 tokens for &lt;code&gt;interact&lt;/code&gt;, 2,275 for &lt;code&gt;generate&lt;/code&gt;, 13,650 for &lt;code&gt;summarize&lt;/code&gt; and 15,600 for &lt;code&gt;modify&lt;/code&gt;. Images and video also consume that budget.&lt;/p&gt;
&lt;p&gt;A request can contain no more than one image and one video. Raw PDF files are not accepted, so users must render a PDF page as an image and send one page per request, or stitch multiple pages into a single image. Video must fit the endpoint payload limit. Frame sampling is fixed when an endpoint starts, and the service rejects per-request &lt;code&gt;media_io_kwargs&lt;/code&gt;. AWS did not specify the payload limit in bytes.&lt;/p&gt;
&lt;p&gt;The endpoint serves base weights only and does not load LoRA adapters for either precision variant. Although the model has a native Multi-Token Prediction head, speculative decoding is disabled in the JumpStart configuration. Reasoning is enabled by default and can be changed per request with &lt;code&gt;enable_thinking&lt;/code&gt;; tool calling and JSON-schema output are supported.&lt;/p&gt;
&lt;p&gt;In a test using 250 input tokens, 1,500 output tokens and five concurrent streaming requests, AWS reported mean per-request output throughput of 63.07 tokens per second on &lt;code&gt;ml.g7.24xlarge&lt;/code&gt; NVFP4, 64.01 on &lt;code&gt;ml.g7.48xlarge&lt;/code&gt; NVFP4 and 140.85 on &lt;code&gt;ml.p5en.48xlarge&lt;/code&gt; BF16. Mean time to first token ranged from 0.45 to 0.53 seconds. AWS cautioned that the tests used different GPU systems and configurations, so the results do not isolate the effect of BF16 versus NVFP4 precision.&lt;/p&gt;
&lt;p&gt;New AWS accounts may need to request a quota increase before deploying the G7 configurations because SageMaker endpoint quotas for &lt;code&gt;ml.g7&lt;/code&gt; instances start at zero. AWS did not give regional availability or endpoint pricing in the announcement.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>aws</category><category>nvidia</category><category>sagemaker-jumpstart</category><category>nemotron-3-5-super-vl</category><category>multimodal-ai</category></item><item><title>AssemblyAI runs speech inference on EC2 Spot GPUs with queued retries</title><link>https://dataphoenix.info/news/assemblyai-ec2-spot-speech-inference</link><guid isPermaLink="true">https://dataphoenix.info/news/assemblyai-ec2-spot-speech-inference</guid><description>AssemblyAI runs speech-to-text inference across EC2 Spot GPU fleets, with Amazon SQS retries moving interrupted jobs to healthy workers.</description><pubDate>Sun, 15 Nov 2026 03:56:26 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/cost-optimizing-assemblyai-speech-to-text-inference-on-aws-with-amazon-ec2-spot/&quot; rel=&quot;noopener noreferrer&quot;&gt;AssemblyAI runs its speech-to-text inference platform&lt;/a&gt; on a mix of Amazon EC2 Spot and On-Demand GPU capacity, according to an AWS case study. Amazon SQS holds transcription jobs for workers, allowing interrupted work to return to the queue for another healthy worker instead of tying a request to one interruptible instance.&lt;/p&gt;
&lt;p&gt;When EC2 prepares to reclaim a Spot instance, an EventBridge event invokes a Lambda interruption handler. Jobs that need more time than the warning window return to the main SQS queue, where another healthy GPU worker can pick them up on a different instance type or in another Availability Zone. &lt;a href=&quot;https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot-instance-termination-notices.html&quot; rel=&quot;noopener noreferrer&quot;&gt;EC2 normally issues the warning two minutes before it stops or terminates an instance&lt;/a&gt;, although AWS says notices are best-effort and hibernation does not receive the two-minute lead time.&lt;/p&gt;
&lt;p&gt;AssemblyAI uses On-Demand instances as an always-on baseline and Spot capacity for demand spikes. Its Spot pool spans multiple Availability Zones and the g4dn, g5 and g6 GPU families, including g4dn.xlarge, g4dn.2xlarge, g5.xlarge, g5.4xlarge and g6.2xlarge instances. Amazon ECS managed scaling adds GPU instances as pending inference work grows. The deployment assigns one inference model to each GPU across what AWS described as thousands of instances.&lt;/p&gt;
&lt;p&gt;Each request enters an SQS queue before a GPU-backed ECS task retrieves the audio from Amazon S3, runs the inference pipeline and writes the completed transcription to a results store. AWS said individual jobs usually last from 15 seconds to several minutes. Work that can finish within the interruption window completes normally, while longer work returns to the queue. &lt;a href=&quot;https://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSDeveloperGuide/sqs-visibility-timeout.html&quot; rel=&quot;noopener noreferrer&quot;&gt;SQS makes an unfinished message available for another attempt&lt;/a&gt; after its visibility timeout expires; setting the timeout to zero makes it immediately available to other consumers. Jobs that exhaust their retries go to a dead-letter queue for investigation.&lt;/p&gt;
&lt;p&gt;AWS reported that the Spot design cut AssemblyAI’s inference-compute cost by 70% compared with On-Demand pricing while compute demand grew nearly fourfold year over year. It also said customers see a small increase in processing time after an interruption while the transcription still completes, and that the design maintains AssemblyAI’s service-level commitments. The case study did not provide the methodology or raw measurements behind the cost, growth, latency or availability claims.&lt;/p&gt;
&lt;p&gt;AssemblyAI operates independent deployments in AWS’s US West (Oregon) and Europe (London) regions. Each region has separate SQS queues, ECS clusters, capacity providers and GPU fleets, which AWS said supports customers’ data-residency requirements.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>industry-and-adoption</category><category>assemblyai</category><category>ec2-spot</category><category>speech-inference</category><category>amazon-sqs</category><category>gpu-compute</category></item><item><title>Databricks adds push-based CDC path into Unity Catalog</title><link>https://dataphoenix.info/news/databricks-push-cdc-unity-catalog</link><guid isPermaLink="true">https://dataphoenix.info/news/databricks-push-cdc-unity-catalog</guid><description>Databricks has linked Debezium Server with Zerobus Ingest, creating a push-based route for continuously landing database changes in Unity Catalog-managed Delta tables.</description><pubDate>Sun, 15 Nov 2026 03:51:54 GMT</pubDate><content:encoded>&lt;p&gt;Databricks has &lt;a href=&quot;https://www.databricks.com/blog/push-based-cdc-unity-catalog-zerobus-ingest-and-debezium&quot; rel=&quot;noopener noreferrer&quot;&gt;added a push-based change-data-capture path&lt;/a&gt; that sends Debezium events through Zerobus Ingest into Unity Catalog-managed Delta tables.&lt;/p&gt;
&lt;p&gt;Database operators can use the integration to continuously move transaction-log changes into Databricks without relying on Kafka as the transport layer between capture and the lakehouse. Delivery is at least once, so consumers still need to deduplicate events rather than assume exactly-once processing.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://debezium.io/blog/2026/09/29/debezium-3-7-final-released/&quot; rel=&quot;noopener noreferrer&quot;&gt;Debezium Server 3.7&lt;/a&gt;, released September 29, includes the Databricks Zerobus Ingest sink adapter. Debezium Server runs alongside the source database, reads its transaction log and turns changes into events. The sink sends those events to a Unity Catalog-managed Delta table through the &lt;a href=&quot;https://docs.databricks.com/aws/en/ingestion/zerobus-overview&quot; rel=&quot;noopener noreferrer&quot;&gt;serverless Zerobus API&lt;/a&gt;. The target table must be in the same region as the Databricks workspace.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://debezium.io/documentation/reference/operations/debezium-server.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Debezium Server documentation&lt;/a&gt; describes three delivery routes: the Zerobus Java SDK over gRPC, a per-record REST option for environments that cannot maintain persistent connections, and a Kafka-protocol option targeting Zerobus’s Kafka-compatible endpoint.&lt;/p&gt;
&lt;p&gt;Teams can choose typed mode, which maps event fields to target-table columns, or envelope mode, which retains the CDC envelope for downstream casting and materialization. The table schema must match the selected mode.&lt;/p&gt;
&lt;p&gt;Databricks says the source-side design can suit restricted networks because database owners control capture and make outbound connections instead of allowing an external pull connector to connect inbound. The documentation establishes the architecture but does not independently test that operational claim.&lt;/p&gt;
&lt;p&gt;The sink advances source offsets only after Zerobus acknowledges a batch. Even so, a durable batch can be replayed after a later failure, making a deterministic idempotency key or equivalent source-position data necessary for deduplication. Landing is append-only; building a current-state Type 1 table still requires downstream processing such as &lt;a href=&quot;https://dataphoenix.info/news/databricks-spark-43-runtime-19&quot;&gt;AUTO CDC&lt;/a&gt; or a merge sequenced by source position.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>databricks</category><category>debezium</category><category>zerobus-ingest</category><category>unity-catalog</category><category>change-data-capture</category></item><item><title>AWS releases catalog of 23 reusable generative AI assets</title><link>https://dataphoenix.info/news/aws-reusable-generative-ai-assets</link><guid isPermaLink="true">https://dataphoenix.info/news/aws-reusable-generative-ai-assets</guid><description>AWS released a catalog of 23 reusable generative AI assets spanning document processing, contact centers, agents, retrieval, development and operations.</description><pubDate>Sun, 15 Nov 2026 03:47:24 GMT</pubDate><content:encoded>&lt;p&gt;AWS has &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/accelerating-generative-ai-delivery-with-reusable-assets&quot; rel=&quot;noopener noreferrer&quot;&gt;published a catalog of 23 reusable generative AI assets&lt;/a&gt;, giving teams deployable code and implementation guidance for document processing, customer support, AI agents, retrieval, development and cloud operations.&lt;/p&gt;
&lt;p&gt;The release packages implementation work that teams would otherwise have to assemble themselves. Five assets cover document processing and extraction, five cover contact centers and voice, four focus on agentic AI and chatbots, four on retrieval-augmented generation and knowledge bases, three on development and migration, and two on infrastructure and operations.&lt;/p&gt;
&lt;p&gt;The entries combine source code, deployment documentation and infrastructure-as-code. Use cases range from document ingestion and extraction to voice and text agents, full-stack agent templates, reranking, semantic caching, model migration, Amazon Bedrock logging and AI-assisted business intelligence.&lt;/p&gt;
&lt;p&gt;AWS describes all 23 assets as open source, production-ready, security-reviewed, validated, maintained and tested in production environments. The public repositories inspected for this story include security policies and detailed deployment material, but no common audit showing that every asset received the same review or maintenance treatment. The open-source claim was not checked license by license across the catalog; two inspected agentic repositories use Apache-2.0 and MIT-0 licenses.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://github.com/aws-samples/contact-center-genai-agent&quot; rel=&quot;noopener noreferrer&quot;&gt;Contact Center GenAI Agent&lt;/a&gt; combines Amazon Connect, Amazon Lex, &lt;a href=&quot;https://dataphoenix.info/news/aws-bedrock-retrieval-modes&quot;&gt;Amazon Bedrock Knowledge Bases&lt;/a&gt; and a retrieval-augmented generation stack for voice and text support. Its repository uses AWS CloudFormation and includes optional asynchronous hallucination detection.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://github.com/aws-solutions-library-samples/accelerated-intelligent-document-processing-on-aws&quot; rel=&quot;noopener noreferrer&quot;&gt;accelerated intelligent document processing repository&lt;/a&gt; implements a serverless optical-character-recognition and generative AI pipeline with two processing modes, queuing, monitoring, configuration profiles, human review and evaluation features. It supports CloudFormation with AWS CDK, along with experimental Terraform and OpenTofu deployment options.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://github.com/awslabs/fullstack-solution-template-for-agentcore&quot; rel=&quot;noopener noreferrer&quot;&gt;Fullstack AgentCore Solution Template&lt;/a&gt; pairs a React interface and Amazon Cognito authentication with an Amazon Bedrock AgentCore backend. It supports multiple agent frameworks and CDK or Terraform infrastructure. The &lt;a href=&quot;https://github.com/aws-samples/sample-agentic-chatbot-accelerator&quot; rel=&quot;noopener noreferrer&quot;&gt;Agentic Chatbot Accelerator&lt;/a&gt; adds text and voice interfaces, four agent patterns, Model Context Protocol server and knowledge-base integrations, evaluation, feedback and observability, with CDK and Terraform deployment options.&lt;/p&gt;
&lt;p&gt;AWS says the full-stack template can deploy in under 30 minutes and the contact-center walkthrough can be completed in under an hour. Those timings were not independently reproduced and depend on prerequisites, account configuration and permissions.&lt;/p&gt;
&lt;p&gt;An &lt;a href=&quot;https://aws.amazon.com/solutions/case-studies/doordash-bedrock-case-study/&quot; rel=&quot;noopener noreferrer&quot;&gt;AWS case study on DoorDash&lt;/a&gt; says its contact-center work was ready for live testing in two months, cut application development time by 50%, kept response latency at 2.5 seconds or less and increased automated testing capacity 50-fold. DoorDash and AWS reported the figures; they were not independently measured for this story.&lt;/p&gt;
&lt;p&gt;A &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/how-ricoh-built-a-scalable-intelligent-document-processing-solution-on-aws/&quot; rel=&quot;noopener noreferrer&quot;&gt;case study co-written by AWS and Ricoh&lt;/a&gt; says the document-processing system reduced onboarding from four to six weeks to two to three days and lowered engineering effort from about 80 hours to under five per deployment. The November catalog describes a sevenfold throughput increase from 10,000 to 70,000 documents per month. The March case study’s prose instead says current processing exceeds 10,000 documents per month and that the infrastructure can scale to 70,000, while its table lists more than 70,000 as the after-state.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>aws</category><category>generative-ai</category><category>amazon-bedrock</category><category>agentic-ai</category><category>rag</category><category>developer-tools</category></item><item><title>GS Caltex puts SageMaker Canvas models into production pipelines</title><link>https://dataphoenix.info/news/gs-caltex-sagemaker-canvas-production-pipelines</link><guid isPermaLink="true">https://dataphoenix.info/news/gs-caltex-sagemaker-canvas-production-pipelines</guid><description>GS Caltex moved no-code models built by process engineers in SageMaker Canvas into governed production workflows that run scheduled batch inference, AWS said.</description><pubDate>Sun, 15 Nov 2026 03:33:25 GMT</pubDate><content:encoded>&lt;p&gt;GS Caltex &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/production-mlops-at-gs-caltex-with-amazon-sagemaker-canvas-part-1&quot; rel=&quot;noopener noreferrer&quot;&gt;connected models built by process engineers in Amazon SageMaker Canvas to managed production workflows in SageMaker Pipelines&lt;/a&gt;, AWS said in a case study published Nov. 13. The arrangement lets process engineers build and validate models without code, while ML engineers control production configuration, approval, deployment and scheduled inference.&lt;/p&gt;
&lt;p&gt;GS Caltex runs batch inference every 30 minutes, according to AWS. In the architecture AWS described, an Amazon EventBridge rule starts a SageMaker Pipeline. The pipeline reads and writes data in Amazon S3 and invokes SageMaker Processing and Batch Transform jobs. AWS presented these scheduled jobs, which release compute after each run, as an alternative to a continuously running inference endpoint.&lt;/p&gt;
&lt;p&gt;This is not a one-click handoff from Canvas. ML engineers set production data paths, provide model configuration, connect preprocessing and inference jobs, and write the code that ties approval and scheduling to the workflow. The pipeline reuses the Data Wrangler preprocessing flow and trained model that process engineers validated in Canvas, with engineers adapting environment-specific paths and settings. Learned preprocessing values are reused rather than fitted again for each inference batch.&lt;/p&gt;
&lt;p&gt;The inference pipeline follows five ordered steps: prepare the input, run the Canvas preprocessing flow, remove the target column and CSV header, execute Batch Transform with the approved model, and combine the predictions with the preprocessed data.&lt;/p&gt;
&lt;p&gt;Validated Canvas model versions go into SageMaker Model Registry with model files, inference configuration and evaluation results for ML-engineer review. AWS documentation &lt;a href=&quot;https://docs.aws.amazon.com/sagemaker/latest/dg/canvas-mlops.html&quot; rel=&quot;noopener noreferrer&quot;&gt;allows Canvas-built models to be registered for production MLOps&lt;/a&gt;, and an earlier AWS guide &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/operationalize-ml-models-built-in-amazon-sagemaker-canvas-to-production-using-the-amazon-sagemaker-model-registry/&quot; rel=&quot;noopener noreferrer&quot;&gt;shows how to move Canvas models through Model Registry&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The GS Caltex reference implementation checks that a model package is complete and approved before creating the SageMaker model resource. AWS cautioned that Model Registry records approval status but does not enforce the check for every inference request, leaving that safeguard to deployment code. ML engineers can inspect pipeline-step status and CloudWatch logs. The case study also recommends alerts for job failures and delays, along with checks for schema mismatches, row alignment, overlapping runs, runtime and cost before the schedule is enabled.&lt;/p&gt;
&lt;p&gt;GS Caltex said the workflow cut model deployment time for ML engineers to a few hours from about a week. The company also reported meaningful inference-cost savings from scheduled batch jobs, but the AWS case study supplied no amount, percentage, baseline or measured usage data. The public reference code uses a sample loan-classification model, not a disclosed copy of the company’s production workload. AWS did not identify GS Caltex’s industrial process, model type, workload size, facilities or model-quality results.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>industry-and-adoption</category><category>gs-caltex</category><category>amazon-sagemaker</category><category>sagemaker-canvas</category><category>mlops</category><category>batch-inference</category></item><item><title>IBM expands Apptio with AI cost and value-management tools</title><link>https://dataphoenix.info/news/ibm-apptio-ai-spending-portfolio</link><guid isPermaLink="true">https://dataphoenix.info/news/ibm-apptio-ai-spending-portfolio</guid><description>IBM added generally available AI cost connections and value-analysis tools to Apptio, while placing MCP servers, planning features and further unit-economics support on a staged roadmap.</description><pubDate>Sun, 15 Nov 2026 02:54:35 GMT</pubDate><content:encoded>&lt;p&gt;IBM has &lt;a href=&quot;https://newsroom.ibm.com/2026-11-09-apptio-expands-portfolio-to-help-enterprises-connect-ai-spending-to-business-value&quot; rel=&quot;noopener noreferrer&quot;&gt;expanded its Apptio portfolio&lt;/a&gt; with generally available connections for AI cost and usage data, along with tools intended to connect that spending with budgets and business outcomes.&lt;/p&gt;
&lt;p&gt;Enterprises can use the connections to bring model-consumption data into the technology financial-management workflows that already cover cloud and other IT costs. IBM said an &lt;a href=&quot;https://newsroom.ibm.com/2026-06-08-new-ibm-study-finds-cios-and-ctos-face-growing-ai-control-gap-as-enterprise-deployment-scales&quot; rel=&quot;noopener noreferrer&quot;&gt;IBM Institute for Business Value study&lt;/a&gt; found that 85% of 2,000 surveyed technology executives lacked full visibility into real-time AI spending.&lt;/p&gt;
&lt;p&gt;Available now are direct, native connectors for what IBM calls leading frontier AI systems. IBM did not name the systems covered. The company said connectors for additional data sources are planned for the fourth quarter of 2026.&lt;/p&gt;
&lt;p&gt;IBM Cloudability now generally supports AI Unit Economics and financial planning for AI services. Its &lt;a href=&quot;https://www.ibm.com/products/cloudability/finops-for-ai&quot; rel=&quot;noopener noreferrer&quot;&gt;FinOps for AI product page&lt;/a&gt; says token and workload telemetry can allocate costs among teams, projects and business units. The service also supports forecasting and budget adjustments for AI spending.&lt;/p&gt;
&lt;p&gt;Within Apptio, AI Assistant, AI Lens and AI Data Assistant are generally available, according to IBM. AI Value &amp;amp; ROI and enhancements covering AI spend, total cost of ownership and usage became generally available on November 6. IBM describes &lt;a href=&quot;https://newsroom.ibm.com/2026-08-06-IBM-Introduces-Apptio-AI-Value-ROI-to-Close-the-Gap-Between-AI-Spend-and-Business-Results&quot; rel=&quot;noopener noreferrer&quot;&gt;AI Value &amp;amp; ROI&lt;/a&gt; as linking each AI initiative to customer-selected proof metrics, then combining business results with modeled technology, token, usage, labor and other cost inputs.&lt;/p&gt;
&lt;p&gt;IBM says the portfolio now links model and infrastructure consumption with cost allocation, budgeting, forecasting and measurement against selected business outcomes. The opened sources do not provide measured customer outcomes for the newly available features.&lt;/p&gt;
&lt;p&gt;IBM also made expanded Kubecost network-cost visibility generally available across all product tiers. The added detail covers AWS NAT Gateway ingress and egress, cross-zone traffic and cross-region traffic, allowing charges to be traced to Kubernetes resources and traffic patterns.&lt;/p&gt;
&lt;p&gt;IBM’s availability list labels the Kubecost Model Context Protocol server “GA: Q3.” The company gives fourth-quarter general-availability targets for MCP servers for Cloudability, Apptio and Targetprocess. Those servers are designed to connect Apptio portfolio data with external AI tools. Driver-Based Planning and AI Scenario Planning are also listed for the fourth quarter. Further AI Unit Economics enhancements are planned for the fourth quarter of 2026, followed by incremental support in the first and second quarters of 2027.&lt;/p&gt;
&lt;p&gt;The announcement also says Cloudability support for AI Commitments was “available in Q3,” with more support planned for the first two quarters of 2027. It does not provide a specific launch date or define the scope of the Q3 support. IBM cautioned that future availability dates and planned functionality are goals that may change or be withdrawn.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>industry-and-adoption</category><category>ibm</category><category>apptio</category><category>ai-spending</category><category>cloudability</category><category>finops</category><category>unit-economics</category></item><item><title>Microsoft winds down Oracle AI cloud capacity deal</title><link>https://dataphoenix.info/news/microsoft-oracle-ai-capacity-wind-down</link><guid isPermaLink="true">https://dataphoenix.info/news/microsoft-oracle-ai-capacity-wind-down</guid><description>Microsoft declined to renew an Oracle AI-compute rental arrangement, Bloomberg exclusively reported, after older Nvidia hardware no longer met its specifications. The scope and financial terms remain undisclosed.</description><pubDate>Sun, 15 Nov 2026 02:50:21 GMT</pubDate><content:encoded>&lt;p&gt;Microsoft declined to renew an Oracle AI-compute rental arrangement, Bloomberg exclusively reported, and has been turning off the rented capacity as part of a broader reassessment of its cloud partnerships.&lt;/p&gt;
&lt;p&gt;The wind-down removes at least some externally rented infrastructure from Microsoft’s AI capacity mix. Bloomberg’s sources said the Oracle systems used older-generation Nvidia chips that had become less useful because they no longer met Microsoft’s specifications. The reporting did not identify the GPU generation or models.&lt;/p&gt;
&lt;p&gt;That broader partnership review also includes &lt;a href=&quot;https://dataphoenix.info/news/microsoft-nscale-texas-expansion-option&quot;&gt;Microsoft’s separate Nscale capacity decision&lt;/a&gt;. Available evidence does not show whether the two decisions share contractual or technical terms.&lt;/p&gt;
&lt;p&gt;The scale and economics of the Oracle change remain unclear. It is not known how much capacity Microsoft rented, how much has been shut off, whether any Oracle capacity remains active or when the wind-down will finish. The contract’s value, pricing and any wind-down costs have not been disclosed. Neither company has publicly said whether the reported non-renewal covers the entire arrangement or only part of it.&lt;/p&gt;
&lt;p&gt;Oracle &lt;a href=&quot;https://www.prnewswire.com/news-releases/oracle-cloud-infrastructure-utilized-by-microsoft-for-bing-conversational-search-301979431.html&quot; rel=&quot;noopener noreferrer&quot;&gt;announced the original arrangement&lt;/a&gt; in November 2023 as a multi-year agreement. It said Microsoft would use Oracle Cloud Infrastructure, or OCI, alongside Azure infrastructure for the inference work behind Bing conversational search. Inference is the process of running a trained AI model to produce answers.&lt;/p&gt;
&lt;p&gt;Oracle said Microsoft connected OCI to Azure through Oracle Interconnect for Microsoft Azure and used Azure Kubernetes Service to direct workloads to OCI Compute. That setup let Microsoft manage computing in Oracle’s cloud through tools in its Azure environment. The announcement did not state the agreement’s financial terms or quantify the rented capacity.&lt;/p&gt;
&lt;p&gt;An Oracle spokesperson said demand for OCI AI infrastructure remained strong, including for earlier-generation GPUs, without confirming or denying the Microsoft non-renewal or saying whether the affected capacity had been reassigned. In a portfolio-wide update, &lt;a href=&quot;https://blogs.oracle.com/ceo/from-the-q4-earnings-call&quot; rel=&quot;noopener noreferrer&quot;&gt;Oracle said&lt;/a&gt; 92% of 35,000 GPUs up for renewal in its fiscal fourth quarter were renewed and global GPU utilization was 97.5%. Those company-reported figures cover Oracle’s broader portfolio and do not establish the status or financial effect of Microsoft’s capacity.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>infrastructure-and-compute</category><category>microsoft</category><category>oracle</category><category>ai-infrastructure</category><category>cloud-computing</category><category>nvidia-gpus</category><category>compute-capacity</category></item><item><title>AWS adds single-shot and retrieval-only modes to Bedrock knowledge bases</title><link>https://dataphoenix.info/news/aws-bedrock-retrieval-modes</link><guid isPermaLink="true">https://dataphoenix.info/news/aws-bedrock-retrieval-modes</guid><description>For each AgenticRetrieveStream request, Amazon Bedrock developers can choose retrieval-only, single-shot RAG or multi-step agentic retrieval.</description><pubDate>Sun, 15 Nov 2026 02:37:05 GMT</pubDate><content:encoded>&lt;p&gt;AWS expanded AgenticRetrieveStream in Amazon Bedrock Managed Knowledge Base with &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/single-shot-and-agentic-retrieval-with-amazon-bedrock-managed-knowledge-base&quot; rel=&quot;noopener noreferrer&quot;&gt;retrieval-only and single-shot retrieval-augmented generation&lt;/a&gt;. Developers can now return passages, retrieve once and generate an answer, or use multi-step retrieval without switching APIs.&lt;/p&gt;
&lt;p&gt;Applications choose a mode for each request by setting &lt;code&gt;maxAgentIteration&lt;/code&gt; from zero through five and selecting whether to generate a response. The setting does not automatically route requests between modes.&lt;/p&gt;
&lt;p&gt;At zero, the service performs one retrieval and returns passages without invoking a foundation model. Setting &lt;code&gt;generateResponse=true&lt;/code&gt; with zero is invalid and returns a 4xx error, according to AWS. AWS describes this mode as functionally equivalent to calling Retrieve directly, although the AgenticRetrieveStream result omits the score and location fields included in a Retrieve response.&lt;/p&gt;
&lt;p&gt;At one, the service performs one retrieval and one answer-generation call without a planning round or query rewrite. AWS says this can reduce latency and model usage compared with agentic retrieval, but the company published no workload-specific measurements for latency, usage, cost or answer quality.&lt;/p&gt;
&lt;p&gt;Values from two through five retain the existing agentic behavior. With response generation enabled, a value of &lt;code&gt;n&lt;/code&gt; permits as many as &lt;code&gt;n - 1&lt;/code&gt; planning rounds before the final generation call. &lt;a href=&quot;https://docs.aws.amazon.com/bedrock/latest/userguide/kb-test-agentic-retrieve.html&quot; rel=&quot;noopener noreferrer&quot;&gt;AWS counts each foundation-model call as one iteration&lt;/a&gt;; an initial speculative retrieval runs before planning and does not count toward the limit.&lt;/p&gt;
&lt;p&gt;All three modes can search multiple knowledge bases, and one request can specify up to five retrievers. The stream returns retrieval results and trace events. When generation is enabled, it also streams response text and includes the generated answer and citations in the final result.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>amazon-bedrock</category><category>rag</category><category>agentic-retrieval</category><category>developer-tools</category></item><item><title>Confluent launches Streamhouse and makes Lightning Tables available on AWS</title><link>https://dataphoenix.info/news/confluent-streamhouse-ai-context</link><guid isPermaLink="true">https://dataphoenix.info/news/confluent-streamhouse-ai-context</guid><description>Confluent introduced Streamhouse to serve current business context and made its Lightning Tables query layer generally available in supported AWS environments.</description><pubDate>Sun, 15 Nov 2026 02:30:51 GMT</pubDate><content:encoded>&lt;p&gt;Confluent introduced its &lt;a href=&quot;https://newsroom.ibm.com/2026-11-04-confluent-introduces-its-streamhouse-architecture-to-power-ai-with-real-time-business-context&quot; rel=&quot;noopener noreferrer&quot;&gt;Streamhouse architecture&lt;/a&gt; on November 4, making Lightning Tables generally available on AWS and adding Query as the fifth pillar of its Data Streaming Platform. The other four pillars are Stream, Connect, Process and Govern.&lt;/p&gt;
&lt;p&gt;Streamhouse is designed to continuously capture, process, govern and serve current business context for AI agents, applications and analytics. Its query layer, &lt;a href=&quot;https://docs.confluent.io/cloud/current/topics/lightning-tables.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Lightning Tables&lt;/a&gt;, is a managed, read-only service that continuously ingests Kafka topic data and materializes it as tables reflecting the latest state.&lt;/p&gt;
&lt;p&gt;That availability has limits. Confluent’s documentation says Lightning Tables must run in a supported AWS region on a Basic, Standard, Enterprise or Dedicated Confluent Cloud cluster. The source topic must use a registered Avro, Protobuf or JSON Schema, and enabling the service requires at least the CloudClusterAdmin role.&lt;/p&gt;
&lt;p&gt;Customers can query Lightning Tables through the Confluent Cloud console, command-line interface, API and JDBC. Confluent also says AI agents can access the tables over MCP through its Real-Time Context Engine.&lt;/p&gt;
&lt;p&gt;Confluent says internal tests measured about 100 milliseconds from new-data creation to serving, query latency of 10 to 100 milliseconds and throughput above 1,000 queries per second. The figures have not been independently benchmarked.&lt;/p&gt;
&lt;p&gt;The company positions Streamhouse as a real-time operational layer that works alongside data warehouses and lakehouses rather than replacing them. The opened materials do not enumerate every supported AWS region or give general-availability dates for Lightning Tables on Azure or Google Cloud.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>confluent</category><category>streamhouse</category><category>lightning-tables</category><category>aws</category><category>data-streaming</category><category>ai-agents</category></item><item><title>Hon Hai&apos;s October revenue rises 46% to NT$1.305 trillion</title><link>https://dataphoenix.info/news/hon-hai-october-2026-revenue-ai-servers</link><guid isPermaLink="true">https://dataphoenix.info/news/hon-hai-october-2026-revenue-ai-servers</guid><description>Hon Hai&apos;s unaudited October revenue reached NT$1.305 trillion, up 45.73% year over year, with cloud and networking products leading growth across its major segments.</description><pubDate>Sun, 15 Nov 2026 01:49:20 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://www.honhai.com/en-us/press-center/press-releases/latest-news/2147&quot; rel=&quot;noopener noreferrer&quot;&gt;Hon Hai reported&lt;/a&gt; unaudited consolidated revenue of NT$1.305 trillion (NT$1,305,320 million) for October 2026. Revenue was 45.73% higher than a year earlier and 12.66% above September. The October result follows the company’s separate September revenue result of NT$1.159 trillion.&lt;/p&gt;
&lt;p&gt;The result lifted Hon Hai’s January-through-October revenue to NT$8.975 trillion, 40.4% above the same period in 2025, according to the company’s &lt;a href=&quot;https://www.honhai.com.tw/en-us/investor-relations/financial-information/reports?category=monthly&quot; rel=&quot;noopener noreferrer&quot;&gt;monthly revenue table&lt;/a&gt;. The company labels the monthly figures unaudited.&lt;/p&gt;
&lt;p&gt;Hon Hai said all four major product segments grew strongly from a year earlier, led by cloud and networking products. The company did not disclose revenue by product or customer in its monthly table.&lt;/p&gt;
&lt;p&gt;Cloud and networking products declined slightly from September despite continued demand tied to AI applications. Hon Hai attributed the sequential decline to consignment-model changes, supply constraints and the early production ramp-up of new models.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>infrastructure-and-compute</category><category>hon-hai</category><category>monthly-revenue</category><category>cloud-infrastructure</category><category>ai-applications</category><category>foxconn</category></item><item><title>NVIDIA says open 5G and 6G tools surpassed one million downloads</title><link>https://dataphoenix.info/news/nvidia-open-source-6g-tools-one-million-downloads</link><guid isPermaLink="true">https://dataphoenix.info/news/nvidia-open-source-6g-tools-one-million-downloads</guid><description>NVIDIA says downloads of its open-source 5G and 6G research tools have passed one million as researchers apply the software to semantic communications, neural receivers and energy management.</description><pubDate>Sun, 15 Nov 2026 01:41:34 GMT</pubDate><content:encoded>&lt;p&gt;NVIDIA said its open-source 5G and 6G research tools have surpassed one million downloads, a company-reported adoption milestone for software used to test AI-native wireless systems.&lt;/p&gt;
&lt;p&gt;The company said global developer adoption of NVIDIA AI Aerial and Sionna grew tenfold over the past year and that more than 8,000 research and academic institutions use the platforms. NVIDIA did not disclose how it counted downloads, adoption or institutions, and the supporting studies do not independently verify those figures.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://blogs.nvidia.com/blog/telco-open-source-tools-research-6g-networks&quot; rel=&quot;noopener noreferrer&quot;&gt;NVIDIA disclosed the figures&lt;/a&gt; while highlighting projects in semantic communications, learned signal reception and network energy management. The work runs on controlled 5G systems, testbeds or simulations and is framed as research relevant to prospective 6G networks. None of the cited projects demonstrates deployment on a commercial 6G network.&lt;/p&gt;
&lt;p&gt;A &lt;a href=&quot;https://arxiv.org/html/2609.15895&quot; rel=&quot;noopener noreferrer&quot;&gt;King’s College London-led preprint&lt;/a&gt; describes a 5G testbed linking a camera-equipped robot arm to a Jetson AGX Orin edge server. Instead of periodically sending raw images, the system transmitted compact, task-relevant representations such as 3D bounding boxes and 2D or 3D scene graphs. This approach, called goal-oriented semantic communication, aims to send the information needed for a task rather than all captured data.&lt;/p&gt;
&lt;p&gt;The researchers evaluated fruit-packing and cube-stacking tasks over 20 independent runs per framework and introduced faults including dropped objects and incorrect placements. They reported reductions of up to 52.6% in task-completion time and improvements of up to 45% in task-success probability compared with periodic raw-image transmission. The paper is an arXiv preprint, and the opened record does not establish peer review or production deployment.&lt;/p&gt;
&lt;p&gt;In a separate test, &lt;a href=&quot;https://www.softbank.jp/en/corp/technology/research/topics/238/&quot; rel=&quot;noopener noreferrer&quot;&gt;SoftBank integrated a Transformer-based neural receiver&lt;/a&gt; into NVIDIA Aerial’s 5G Layer 1 and evaluated it over the air with a commercial smartphone in a shield room. A neural receiver uses a trained model for signal-processing steps that conventional systems handle with fixed algorithms. SoftBank’s model replaced channel estimation, interpolation, equalization and soft demapping, while both receiver paths used the same downstream descrambling, rate-recovery and LDPC-decoding stages.&lt;/p&gt;
&lt;p&gt;SoftBank said mean uplink Layer 1 throughput was higher for the neural receiver at all 15 interference settings tested. At one setting, it reported 47.9 Mbit/s versus 29.9 Mbit/s for the conventional receiver, a 60.5% gain. The result came from a single-layer uplink test in a shield room and does not establish performance in field deployments or across all radio conditions.&lt;/p&gt;
&lt;p&gt;An &lt;a href=&quot;https://arxiv.org/html/2609.04004&quot; rel=&quot;noopener noreferrer&quot;&gt;ETH Zurich and NVIDIA preprint&lt;/a&gt; also tested whether receiver models could be adapted to particular indoor sites. The authors reported that fine-tuning a model-driven neural receiver reduced dataset block-error rate from 0.351 to 0.237 in a small laboratory and from 0.463 to 0.240 on an office floor. The paper identifies one author with NVIDIA and acknowledges company sponsorship. It also says the reference receiver achieved a lower block-error rate than the fine-tuned model-driven receiver in both environments.&lt;/p&gt;
&lt;p&gt;A &lt;a href=&quot;https://eprints.whiterose.ac.uk/id/eprint/242057/1/Accepted_Network_in_a_box__ICC_2026_Yejing.pdf&quot; rel=&quot;noopener noreferrer&quot;&gt;University of Leeds-led paper&lt;/a&gt; applied AI to energy management. Its system used a long short-term memory model to predict radio-unit load 10 minutes ahead, classified expected demand with confidence estimates and guided base-station sleep decisions. A digital twin provided a repeatable environment for testing and safety validation before control actions were exercised on a 5G network-in-a-box testbed.&lt;/p&gt;
&lt;p&gt;In a controlled four-base-station urban scenario, the authors reported a 45.6% reduction in radio access network power, equal to 6.03 kWh per day, with zero outage events. The result is specific to the study’s digital-twin and testbed framework and does not demonstrate savings on a production network.&lt;/p&gt;
&lt;p&gt;The public &lt;a href=&quot;https://github.com/NVIDIA/aerial-cuda-accelerated-ran&quot; rel=&quot;noopener noreferrer&quot;&gt;Aerial repository&lt;/a&gt; provides GPU-accelerated components for building 5G and prospective 6G radio access network software. &lt;a href=&quot;https://github.com/NVlabs/sionna&quot; rel=&quot;noopener noreferrer&quot;&gt;Sionna&lt;/a&gt; supplies open-source simulation tools for communications research. Both repositories list the Apache License 2.0.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>infrastructure-and-compute</category><category>nvidia</category><category>aerial</category><category>sionna</category><category>5g</category><category>6g</category><category>open-source</category><category>neural-receivers</category><category>semantic-communications</category><category>open-ran</category></item><item><title>Kiro Crew launches persistent bots for recurring development work</title><link>https://dataphoenix.info/news/kiro-crew-bots-general-availability</link><guid isPermaLink="true">https://dataphoenix.info/news/kiro-crew-bots-general-availability</guid><description>Kiro Crew bots turn recurring development responsibilities into persistent, scheduled workflows while leaving goals, permissions, and review boundaries with users.</description><pubDate>Sun, 15 Nov 2026 01:37:20 GMT</pubDate><content:encoded>&lt;p&gt;Kiro made bots generally available in Kiro Crew on Nov. 5, introducing named, long-running agents that users configure for recurring development responsibilities. Unlike isolated coding sessions, the &lt;a href=&quot;https://kiro.dev/blog/crew-bots/&quot; rel=&quot;noopener noreferrer&quot;&gt;Kiro Crew bots&lt;/a&gt; retain an ongoing conversation, notes, and a dedicated workspace so work and context can carry across separate runs.&lt;/p&gt;
&lt;p&gt;According to Kiro, a bot can start multiple worker sessions in parallel, monitor their progress, redirect work that drifts from its plan, and return the results to its main conversation. If one branch needs a human decision, the bot can ask the user while continuing work that does not depend on the answer. Users can also create schedules through the conversation; Kiro cites overnight build-failure checks and weekly release-note preparation as examples.&lt;/p&gt;
&lt;p&gt;That persistence does not give bots unrestricted authority. Users define each bot’s goal, inputs, completion criteria, review boundaries, escalation points, and schedule. Tool access and permissions are set per bot. Kiro describes a research bot that can read files but not change them, a deployment bot that requires approval before making changes, and a code-review bot that can comment or request changes but cannot approve or merge code. A bot may request another tool, but its access stays unchanged until the user approves the request.&lt;/p&gt;
&lt;p&gt;Each responsibility also keeps its own memory. Kiro says users can inspect, edit, or delete what a bot remembers, revise its instructions, and adjust its permissions as the workflow changes. A bot can maintain a customizable dashboard and work log, giving users a view of scheduled activity, open items, and underlying sessions without combining every responsibility in one context.&lt;/p&gt;
&lt;p&gt;Kiro also says users may choose different models for a coordinating bot and its worker agents. Crew bots can run through the product on a laptop, a Mac Mini, or an Amazon EC2 instance.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>kiro</category><category>kiro-crew</category><category>coding-agents</category><category>developer-tools</category><category>automation</category></item><item><title>Tesla invites select customers to buy Cybercab for personal use</title><link>https://dataphoenix.info/news/tesla-cybercab-personal-use-limited-offer</link><guid isPermaLink="true">https://dataphoenix.info/news/tesla-cybercab-personal-use-limited-offer</guid><description>Tesla’s invitation-only Cybercab offer requires a paid monthly autonomy subscription and confines operation to Tesla-enabled geofences.</description><pubDate>Sun, 15 Nov 2026 00:31:04 GMT</pubDate><content:encoded>&lt;p&gt;Tesla invited select customers on November 13 to buy Cybercab for personal use, creating a limited path to individual ownership of a vehicle previously deployed through its robotaxi service. The invitation carries a reported $29,990 purchase price.&lt;/p&gt;
&lt;p&gt;The vehicle will operate only with an active Autonomy Subscription. The subscription is reported at $299 a month. Tesla’s &lt;a href=&quot;https://www.tesla.com/configurator/api/v4/terms?category=autonomy_subs&amp;amp;locale=en_US&amp;amp;model=cc&amp;amp;saleType=Sale&quot; rel=&quot;noopener noreferrer&quot;&gt;Cybercab Autonomy Subscription Agreement&lt;/a&gt; confirms the recurring requirement but does not list either amount.&lt;/p&gt;
&lt;p&gt;The agreement limits Cybercab to personal, non-commercial use. Owners may not list the vehicle on Tesla’s Robotaxi Network or on third-party ride-hailing, delivery, rental or similar platforms unless Tesla authorizes or enables that use.&lt;/p&gt;
&lt;p&gt;Operation is geographically constrained as well. Tesla says Cybercab works only within Robotaxi geofences and other operating limits that the company enables. A privately owned vehicle’s supported area may differ from the commercial Robotaxi service area, and Tesla can change those areas over time.&lt;/p&gt;
&lt;p&gt;Broader availability remains conditional. Tesla says feature activation and use depend on continuing development and regulatory approval, which varies by jurisdiction. The current invitation is not a nationwide retail launch.&lt;/p&gt;
&lt;p&gt;Cybercab’s touchscreen manual controls can move the vehicle only short distances at speeds of up to 2 mph. Tesla says the controls are not intended for public roads and require the user to hold a valid U.S. driver’s license.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>tesla</category><category>cybercab</category><category>autonomous-vehicles</category><category>robotaxi</category><category>personal-use</category></item><item><title>A&amp;O Shearman values AI commitment at $1.1 billion, mostly in lawyer time</title><link>https://dataphoenix.info/news/ao-shearman-ai-commitment</link><guid isPermaLink="true">https://dataphoenix.info/news/ao-shearman-ai-commitment</guid><description>A&amp;O Shearman&apos;s five-year AI commitment derives most of its $1.1 billion value from lawyer time, while its technology spending remains unspecified.</description><pubDate>Sat, 14 Nov 2026 23:43:59 GMT</pubDate><content:encoded>&lt;p&gt;A&amp;amp;O Shearman has committed $1.1 billion over five years to its artificial intelligence strategy. Most of the stated value comes from lawyer time rather than a cash technology outlay. In its &lt;a href=&quot;https://www.aoshearman.com/en/news/ao-shearman-accelerates-ai-strategy-with-a-usd11bn-lawyer-led-commitment&quot; rel=&quot;noopener noreferrer&quot;&gt;announcement of the program&lt;/a&gt;, the firm said every legal practice must devote at least 3% of its time to developing specialist AI tools and services for its area of law.&lt;/p&gt;
&lt;p&gt;That structure makes the headline figure a firm-assigned value for a combined work and technology program. It is not a disclosure that A&amp;amp;O Shearman will spend $1.1 billion in cash on software, infrastructure or vendors. The program also puts a firmwide time mandate behind &lt;a href=&quot;https://dataphoenix.info/news/openai-legora-legal-work-adoption&quot;&gt;legal AI adoption&lt;/a&gt;, rather than limiting the effort to product procurement.&lt;/p&gt;
&lt;p&gt;The firm said lawyer time represents the majority of the commitment. It has not disclosed how it valued that time, the size of the cash portion or how technology funding will be divided among products, vendors and other costs.&lt;/p&gt;
&lt;p&gt;A&amp;amp;O Shearman said its approach brings legal specialists together with data scientists and developers to create tools for specific areas of law. Its ContractMatrix product suite was developed with Microsoft and legal AI company Harvey. The firm said it is delivering the software to clients across several jurisdictions, in some cases using new pricing structures.&lt;/p&gt;
&lt;p&gt;The firm also reported increased use of its existing AI products. It said client uptake across ContractMatrix rose ninefold over the preceding 12 months. In September, 96% of its partners and 93% of its lawyers used its AI tools, generating more than 558,000 activities across ContractMatrix and Harvey workflow products. Those are company-reported metrics; the opened evidence did not include an independent audit or underlying dataset.&lt;/p&gt;
&lt;p&gt;A&amp;amp;O Shearman said a qualified lawyer remains responsible for the firm’s advice regardless of AI’s role in producing it. The firm said its AI governance committees oversee deployments across practices, sectors and offices.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>industry-and-adoption</category><category>ao-shearman</category><category>legal-ai</category><category>lawyer-time</category><category>contractmatrix</category><category>enterprise-ai</category></item><item><title>Adobe previews 10 creative interaction experiments at MAX 2026</title><link>https://dataphoenix.info/news/adobe-max-2026-ai-sneaks</link><guid isPermaLink="true">https://dataphoenix.info/news/adobe-max-2026-ai-sneaks</guid><description>Adobe previewed 10 research projects at MAX 2026 that explore conversational, spatial and direct-manipulation interfaces across video, audio, images and design, without committing to ship them in Creative Cloud or Firefly.</description><pubDate>Sat, 14 Nov 2026 23:40:59 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://blog.adobe.com/en/publish/2026/11/11/adobe-max-sneaks&quot; rel=&quot;noopener noreferrer&quot;&gt;Adobe previewed 10 research projects at MAX 2026&lt;/a&gt; spanning video, audio, imaging and design. The Sneaks explore new ways to direct scenes, develop characters, navigate images and manipulate creative work, but Adobe presented them as research innovations.&lt;/p&gt;
&lt;p&gt;This was not a product launch. Adobe gave no release dates, pricing, performance benchmarks or commitments to bring the projects to &lt;a href=&quot;https://dataphoenix.info/news/adobe-max-agentic-creative-workflows&quot;&gt;Creative Cloud&lt;/a&gt; or Firefly. Of the 10 Sneaks, only Project Deep Brush and Project Hands On were available as early research previews in Adobe Labs Playground. Project Custom Filters was also previewed there, but Adobe listed it separately from the 10 Sneaks.&lt;/p&gt;
&lt;p&gt;Four video experiments center on context, conversation and timing. Adobe says Project Scene Sweep recognizes characters, locations and scenes across a sequence, allowing an edit to carry across cuts and camera angles. Project Screen Test lets creators define a fictional character’s appearance, voice and personality, then use natural-language dialogue to explore the character and possible responses in different scenes.&lt;/p&gt;
&lt;p&gt;According to Adobe, Project Live Set explores changing a subject, environment or camera direction while video continues to play. Project On Cue adds an editable timing signal for generated-video events; Adobe says creators can draw the signal or extract it from video or music.&lt;/p&gt;
&lt;p&gt;For audio, Project Sound Bender is designed to isolate a described sound from a finished mix and arrange variations on an interactive canvas. Adobe says creators can shape the sound in real time, match its movement to video or generate a new sound from a description.&lt;/p&gt;
&lt;p&gt;The image and design projects put controls directly on the work. Adobe says Project Deep Brush interprets an image’s shape, color and lighting so brushstrokes can follow surfaces, preserve shadows during recoloring or add light to a scene. The company describes Project Sight Span as turning a single image into a navigable environment for finding new viewpoints and composing beyond the original frame. Project Hands On uses object segmentation and vision-language models to identify image content and add on-canvas controls, according to Adobe.&lt;/p&gt;
&lt;p&gt;Adobe says Project Rinse Repeat captures edits from one part of an image, adapts them to another and recreates them with native Photoshop tools, keeping the result editable. The company says Project Link Up treats a drag gesture as design intent, adjusting an element’s perspective, scale and style while returning editable vector artwork.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>adobe</category><category>adobe-max</category><category>adobe-sneaks</category><category>creative-ai</category><category>interaction-design</category></item><item><title>Q-CTRL raises $290 million in Salesforce Ventures-led Series C</title><link>https://dataphoenix.info/news/q-ctrl-290-million-series-c</link><guid isPermaLink="true">https://dataphoenix.info/news/q-ctrl-290-million-series-c</guid><description>Q-CTRL raised $290 million in a Salesforce Ventures-led Series C and plans to expand its AI-driven quantum-control and navigation programs.</description><pubDate>Sat, 14 Nov 2026 23:31:13 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://q-ctrl.com/blog/q-ctrl-secures-us-290-million-series-c-funding-led-by-salesforce-ventures-to-accelerate-the-convergence-of-quantum-and-ai&quot; rel=&quot;noopener noreferrer&quot;&gt;Q-CTRL raised US$290 million in a Series C led by Salesforce Ventures&lt;/a&gt;, securing new capital to expand its products, scale globally and continue its autonomy and aerospace programs.&lt;/p&gt;
&lt;p&gt;The company named Affinity Partners, AMD Ventures, EDGE Group, MPower Partners, Australia’s National Reconstruction Fund Corporation and SPARX Asset Management as new investors. Returning backers included Alpha Edison, Alumni Ventures, Bullhound Capital, DCVC, In-Q-Tel, Lockheed Martin Ventures, Main Sequence, Ridgeline, Salus Ventures and Square Peg. Q-CTRL called itself a newly minted unicorn but did not disclose an exact valuation or financing terms.&lt;/p&gt;
&lt;p&gt;Q-CTRL develops software designed to reduce errors in quantum hardware. &lt;a href=&quot;https://quantum.cloud.ibm.com/functions?id=q-ctrl-dynamics-simulator&quot; rel=&quot;noopener noreferrer&quot;&gt;IBM’s Quantum Platform describes Q-CTRL’s tools&lt;/a&gt; as providing automated, AI-driven error suppression and mapping time-evolution workloads to hardware while counteracting device noise.&lt;/p&gt;
&lt;p&gt;In May, &lt;a href=&quot;https://research.ibm.com/blog/ibm-think-2026-modeling-the-universe-with-quantum-computing&quot; rel=&quot;noopener noreferrer&quot;&gt;IBM Research said Q-CTRL used IBM quantum hardware&lt;/a&gt; for a materials simulation involving up to 60 interacting electrons. IBM attributed the comparison to Q-CTRL: the workflow took about two minutes instead of 100 hours and produced results within 1% of a leading classical method. The opened sources did not independently reproduce those figures.&lt;/p&gt;
&lt;p&gt;Q-CTRL is also developing &lt;a href=&quot;https://q-ctrl.com/ironstone-opal&quot; rel=&quot;noopener noreferrer&quot;&gt;Ironstone Opal&lt;/a&gt;, a navigation system that measures variations in Earth’s magnetic and gravitational fields. The company says AI-powered software filters platform interference, compares live readings with geophysical maps and combines quantum sensors with an inertial-navigation unit to determine position without GPS.&lt;/p&gt;
&lt;p&gt;Q-CTRL claims the system improves positioning by more than 100 times, meets a 0.3-nautical-mile navigation requirement with 95% confidence and is 20 times more resilient in motion. The opened sources did not include an independent test report validating those figures.&lt;/p&gt;
&lt;p&gt;The financing announcement said the navigation technology is being deployed on commercial aircraft and autonomous vehicles. Q-CTRL’s product page, however, describes Ironstone Opal as available for presale and refers to selected partners conducting field trials and integration demonstrations. The scale and commercial status of those deployments remain unclear. Airbus listed Q-CTRL’s sensor-based navigation technology among notable outcomes from its startup engagement in its 2025 board report.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>funding-and-ma</category><category>q-ctrl</category><category>quantum-computing</category><category>quantum-control</category><category>funding</category><category>salesforce-ventures</category><category>quantum-navigation</category></item><item><title>AWS maps a six-agent product development workflow in Quick</title><link>https://dataphoenix.info/news/aws-quick-ai-pdlc-agents</link><guid isPermaLink="true">https://dataphoenix.info/news/aws-quick-ai-pdlc-agents</guid><description>AWS has detailed a configurable Quick workflow in which an orchestrator coordinates five specialist agents, with a person approving each stage before the pipeline advances.</description><pubDate>Sat, 14 Nov 2026 22:27:10 GMT</pubDate><content:encoded>&lt;p&gt;AWS has &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/accelerating-product-development-with-ai-driven-lifecycle-management-ai-pdlc-in-aws-quick&quot; rel=&quot;noopener noreferrer&quot;&gt;detailed a configurable product development workflow in Quick&lt;/a&gt; that pairs an orchestrator with five specialist agents to move from customer evidence to a clickable prototype. A person must approve each stage before the pipeline moves forward.&lt;/p&gt;
&lt;p&gt;The workflow spans the Discover, Decide, Design and Prototype phases. Outputs from each phase stay in a Quick Space, while a shared audit file links source material and recorded decisions to later specifications and prototype assets.&lt;/p&gt;
&lt;p&gt;Users interact with the Orchestrator agent, which loads the specialists’ question frameworks, asks the relevant questions and delegates artifact generation. The Discovery agent takes answers, URLs and connected enterprise data, scores pain points for severity and frequency, and creates a Pain Point Board. The Prioritize agent builds a weighted Scoring Matrix, and the Working Backwards agent produces a PR/FAQ for the selected use case.&lt;/p&gt;
&lt;p&gt;Once approved, the Spec Generator turns the validated use case into a Design Handoff Brief with acceptance criteria, user flow, constraints and source-lineage metadata. The Designer then uses the upstream artifacts to create a journey map, wireframes and a clickable HTML prototype for stakeholder validation. Specialists record their decisions, and the Orchestrator validates and adds them to the shared audit trail.&lt;/p&gt;
&lt;p&gt;The pipeline’s stage gates apply to the product lifecycle, a different scope from &lt;a href=&quot;https://dataphoenix.info/news/aws-amazon-quick-smart-approvals&quot;&gt;Amazon Quick’s Smart Approvals controls&lt;/a&gt; for eligible agent actions.&lt;/p&gt;
&lt;p&gt;AWS offers two pacing options. Deep Discovery asks eight to 12 questions per phase. Fast Track asks three to five essential questions and suggests defaults that users can confirm or override. The &lt;a href=&quot;https://github.com/aws-samples/sample-AI-PDLC-using-AmazonQuick&quot; rel=&quot;noopener noreferrer&quot;&gt;reference package&lt;/a&gt; contains agent definitions and prompts rather than application code; its full JSON agent package is the configuration source for all six agents.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>aws</category><category>aws-quick</category><category>ai-agents</category><category>product-development</category><category>human-in-the-loop</category></item><item><title>Oracle sets 12-month window for AI Lakehouse semantic layer</title><link>https://dataphoenix.info/news/oracle-ai-lakehouse-semantic-layer-12-month-roadmap</link><guid isPermaLink="true">https://dataphoenix.info/news/oracle-ai-lakehouse-semantic-layer-12-month-roadmap</guid><description>Oracle announced new Autonomous AI Lakehouse capabilities, but its Business Semantic Layer and the AI Data Catalog’s agentic features remain roadmap items due within 12 months.</description><pubDate>Sat, 14 Nov 2026 21:27:20 GMT</pubDate><content:encoded>&lt;p&gt;On Oct. 26, Oracle &lt;a href=&quot;https://www.oracle.com/news/announcement/ai-world-oracle-advances-autonomous-ai-lakehouse-with-business-semantic-layer-and-expanded-support-for-iceberg-data-2026-10-26/&quot; rel=&quot;noopener noreferrer&quot;&gt;announced new capabilities for Autonomous AI Lakehouse&lt;/a&gt;, including a Business Semantic Layer, agentic metadata enrichment for AI Data Catalog, expanded Apache Iceberg support and new governance controls. Oracle said the catalog’s agentic capabilities and the semantic-layer capabilities will become available within 12 months; it did not present either as generally available at launch.&lt;/p&gt;
&lt;p&gt;The underlying &lt;a href=&quot;https://www.oracle.com/autonomous-database/autonomous-ai-lakehouse/&quot; rel=&quot;noopener noreferrer&quot;&gt;Autonomous AI Lakehouse&lt;/a&gt; is available now on Oracle Cloud Infrastructure, Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Exadata Cloud@Customer and OCI Dedicated Region, according to Oracle. The existing infrastructure lets customers query supported operational databases, object stores and lakehouse tables without first moving all of the data into a separate system. Oracle intends the roadmap additions to put shared business definitions and policy controls over those distributed sources.&lt;/p&gt;
&lt;p&gt;Oracle described the Business Semantic Layer as a reusable definition system for dimensions, measures, metrics, hierarchies, calculations and key performance indicators. The system would give analysts, applications and AI agents the same vocabulary for interpreting fields and calculating business measures. Oracle said the layer would draw on catalog annotations, imported Oracle AI Database SQL annotations and AI-suggested analytic views to support natural-language-to-SQL, retrieval and analytics.&lt;/p&gt;
&lt;p&gt;The AI Data Catalog can mount external catalogs, Oracle and non-Oracle databases, Iceberg tables, applications and object-store data without duplicating them, Oracle said. The planned agentic features would continuously enrich metadata with business meaning, lineage, workload patterns, relationships and retrieval signals. Oracle said those additions are designed to narrow searches to relevant assets and improve discovery and natural-language analytics. It did not give a date more precise than the 12-month window.&lt;/p&gt;
&lt;p&gt;Oracle also &lt;a href=&quot;https://blogs.oracle.com/database/your-data-stays-put-ai-comes-to-it-whats-new-in-oracle-autonomous-ai-lakehouse&quot; rel=&quot;noopener noreferrer&quot;&gt;detailed expanded native Iceberg support&lt;/a&gt;, including the ability to create, populate, update and delete Iceberg tables with Oracle SQL, plus time travel for auditing and recovery. Oracle’s release grouped those functions with the new and expanded capabilities but did not assign them an explicit availability date. Oracle, however, said Data Studio now offers an AI-assisted workspace through an OCI account, connects to more than 100 data sources and accepts natural language, SQL or Python.&lt;/p&gt;
&lt;p&gt;For governance, Oracle said AI Data Catalog provides role-based, object-level controls across supported Oracle, non-Oracle and Iceberg data. Oracle’s technical post says those policies work alongside, rather than replace, controls enforced by the underlying source systems. Oracle also claimed Data Lake Accelerator enhancements ran queries up to three times faster on a 100 TB TPC-DS data set and increased maximum cluster size from 32 to 100 nodes; the supplied sources did not independently validate that benchmark.&lt;/p&gt;
&lt;p&gt;Oracle did not disclose exact general-availability dates or pricing for the catalog’s agentic capabilities or the Business Semantic Layer. Its future-product disclaimer says the description is not a delivery commitment and that development, release, timing and pricing may change at Oracle’s discretion.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>oracle</category><category>autonomous-ai-lakehouse</category><category>business-semantic-layer</category><category>ai-data-catalog</category><category>apache-iceberg</category></item><item><title>llama.cpp and ggml reach v1.0.0, set rules for breaking changes</title><link>https://dataphoenix.info/news/llama-cpp-ggml-v1-semantic-versioning</link><guid isPermaLink="true">https://dataphoenix.info/news/llama-cpp-ggml-v1-semantic-versioning</guid><description>The ggml project released llama.cpp and ggml libraries at version 1.0.0, committing to major-version increases when future changes break API or ABI compatibility.</description><pubDate>Sat, 14 Nov 2026 21:00:06 GMT</pubDate><content:encoded>&lt;p&gt;The ggml project &lt;a href=&quot;https://huggingface.co/blog/ggml-org/llama-cpp-v1&quot; rel=&quot;noopener noreferrer&quot;&gt;released llama.cpp and ggml libraries at version 1.0.0&lt;/a&gt; on November 13, pairing the milestone with a commitment to increase their major version when future changes break API or ABI compatibility.&lt;/p&gt;
&lt;p&gt;For developers who depend on the libraries, the policy turns the version number into an explicit signal for incompatible interface changes. Under the project’s rules, a major increase marks a breaking change to the public C API. A minor increase covers a backward-compatible feature, backend update, model-support change or API addition. A patch increase covers a bug fix with no API change. Those categories follow the &lt;a href=&quot;https://semver.org/&quot; rel=&quot;noopener noreferrer&quot;&gt;Semantic Versioning 2.0.0 specification&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;The commitment covers libllama’s public headers in &lt;code&gt;llama.cpp/include/&lt;/code&gt; and libggml’s public headers in &lt;code&gt;ggml/include/&lt;/code&gt;. The announcement says future ABI incompatibilities will also require a major-version increase.&lt;/p&gt;
&lt;p&gt;Version 1.0.0 includes breaking cleanup. The &lt;a href=&quot;https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0&quot; rel=&quot;noopener noreferrer&quot;&gt;llama.cpp release&lt;/a&gt; removes deprecated public API, including 35 deprecated functions. The &lt;a href=&quot;https://github.com/ggml-org/ggml/releases/tag/v1.0.0&quot; rel=&quot;noopener noreferrer&quot;&gt;ggml release&lt;/a&gt; prunes deprecated public API and removes legacy &lt;code&gt;DIAG_MASK&lt;/code&gt; operations.&lt;/p&gt;
&lt;p&gt;Both releases add features alongside the compatibility changes. llama.cpp adds OCP FP8 key-value caches, a host-memory mixture-of-experts cache with a configurable CPU share, external PLE shard loading, additional model architectures, server hardening and a speech endpoint. ggml adds configurable YaRN truncation and FP8 support, with performance work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends.&lt;/p&gt;
&lt;p&gt;In a &lt;a href=&quot;https://github.com/ggml-org/llama.cpp/discussions/30566&quot; rel=&quot;noopener noreferrer&quot;&gt;release-preparation discussion&lt;/a&gt;, ggerganov said major releases would likely occur more often than the interval between the pre-1.0 line and v1.0.0 because incompatible API or ABI changes will now be versioned more strictly. The published material does not define a separate ABI-compatibility test. It also does not promise a deprecation window, long-term-support branch or migration period before a future major release.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>llama-cpp</category><category>ggml</category><category>semantic-versioning</category><category>local-ai</category><category>developer-tools</category></item><item><title>AMD introduces Versal devices with silicon-backside attack protection</title><link>https://dataphoenix.info/news/amd-versal-silicon-backside-attack-protection</link><guid isPermaLink="true">https://dataphoenix.info/news/amd-versal-silicon-backside-attack-protection</guid><description>AMD introduced three Versal tamper-resistant adaptive SoCs with a bonded protection die intended to detect specified silicon-backside attacks, but the opened evidence does not independently validate their security performance.</description><pubDate>Sat, 14 Nov 2026 20:54:36 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://newsroom.amd.com/news/amd-advances-built-in-protection-physical-attacks-security-critical-systems&quot; rel=&quot;noopener noreferrer&quot;&gt;AMD introduced&lt;/a&gt; three Versal tamper-resistant adaptive system-on-chips designed to detect specified physical attacks on the silicon backside. The package-level defense targets security-critical embedded and AI systems, though the opened evidence contains no product-specific independent security test results.&lt;/p&gt;
&lt;p&gt;The 2VE3358TR, 2VE3558TR and 2VE3858TR pair a compute die with a sensor-based protection die through 3D active-on-active packaging. AMD says this added die covers the compute silicon’s backside, detects covered attack methods and prevents the system from booting if the protection die is removed or tampered with. The company cautions that the devices are tamper-resistant, not tamper-proof: protection is limited to specified physical attack methods under a product threat model and other limitations.&lt;/p&gt;
&lt;p&gt;AMD describes the products as the first adaptive SoC or FPGA family with built-in protection against silicon-backside physical attacks. That industry-first assertion relies on AMD’s own competitive analysis and was not independently established by the opened sources. The evidence also contains no product-specific third-party penetration test, certification, threat-model document or attack-coverage matrix.&lt;/p&gt;
&lt;p&gt;A &lt;a href=&quot;https://arxiv.org/abs/2611.17065&quot; rel=&quot;noopener noreferrer&quot;&gt;contemporaneous research paper&lt;/a&gt; argues that FPGA anti-tamper sensor detection and response mechanisms can be bypassed or affected by temperature, aging and power-delivery behavior. It does not test Versal TR devices or AMD’s backside protection die, so it cannot establish how these products will perform against physical attacks.&lt;/p&gt;
&lt;p&gt;The chips also provide hardware acceleration for the ML-DSA and SLH-DSA post-quantum signature algorithms. The &lt;a href=&quot;https://csrc.nist.gov/Projects/digital-signatures&quot; rel=&quot;noopener noreferrer&quot;&gt;National Institute of Standards and Technology&lt;/a&gt; identifies those algorithms with FIPS 204 and FIPS 205, respectively. The available sources, however, do not show certification or independent validation of AMD’s implementation.&lt;/p&gt;
&lt;p&gt;The devices use AMD’s Versal AI Edge Series Gen 2 computing architecture. AMD advertises up to 229,000 DMIPS, 543,000 lookup tables and 369 INT8 TOPS. Its &lt;a href=&quot;https://www.amd.com/en/products/adaptive-socs-and-fpgas/versal/gen2/ai-edge-series.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Gen 2 specifications&lt;/a&gt; identify 369 TOPS as a maximum-sparsity figure and list 184 INT8 TOPS for dense workloads on the 2VE3858. The research record includes no independent benchmark for the tamper-resistant devices.&lt;/p&gt;
&lt;p&gt;AMD says each TR model has a footprint-compatible standard Versal AI Edge Gen 2 counterpart, allowing a design to move to the security-focused part without repeating synthesis or place-and-route. AMD expects first samples in the second half of 2027 and plans production for the first half of 2028. Those dates are company roadmap targets, not completed availability milestones.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>amd</category><category>versal</category><category>hardware-security</category><category>tamper-resistance</category><category>adaptive-soc</category><category>post-quantum-cryptography</category></item><item><title>Nebuly raises $21 million Series A to measure enterprise AI returns</title><link>https://dataphoenix.info/news/nebuly-series-a-enterprise-ai-returns</link><guid isPermaLink="true">https://dataphoenix.info/news/nebuly-series-a-enterprise-ai-returns</guid><description>Nebuly raised a $21 million Series A led by Planven to expand a platform that it says measures the business value of enterprise AI agents.</description><pubDate>Sat, 14 Nov 2026 20:43:21 GMT</pubDate><content:encoded>&lt;p&gt;Nebuly raised $21 million in a Series A led by Planven to expand a platform that it says measures the business value generated by enterprise AI agents. In its &lt;a href=&quot;https://www.nebuly.com/blog/nebuly-raises-21m-series-a&quot; rel=&quot;noopener noreferrer&quot;&gt;funding announcement&lt;/a&gt;, the company said the capital will support further product development and growth of its go-to-market team in the United States and Europe.&lt;/p&gt;
&lt;p&gt;The round also included Exor Ventures, Tenacity Capital, TGVG Capital and Fastweb+Vodafone through Founders Factory, according to Nebuly. Returning investors and angel investors also participated, the company said, while CDP Venture Capital SGR supported the financing through its Artificial Intelligence Fund and Digital Transition Fund. Nebuly did not disclose its valuation, the round’s structure or its total funding to date.&lt;/p&gt;
&lt;p&gt;Another recent financing in the category was &lt;a href=&quot;https://dataphoenix.info/news/larridin-30-million-series-a&quot;&gt;Larridin’s $30 million Series A&lt;/a&gt;, which backed a separate platform for tying enterprise AI spending and adoption to workflow output.&lt;/p&gt;
&lt;p&gt;Nebuly describes its software as a measurement layer for employee-facing and customer-facing AI tools. For internal agents, the company says the platform estimates hours saved, capacity freed and work that teams could not previously perform. For customer-facing assistants, Nebuly says it tracks revenue influence and surfaces churn-risk and upsell signals. The company also says the assumptions behind those calculations remain visible and adjustable, allowing finance and leadership teams to inspect how the figures were produced.&lt;/p&gt;
&lt;p&gt;Nebuly said the platform connects to custom agents and assistants built on Claude, ChatGPT and Microsoft Copilot. Its models can run on premises or in a customer’s private cloud, which the company says allows enterprise data to be analyzed without sending it to external model APIs.&lt;/p&gt;
&lt;p&gt;Nebuly said it has signed multiple six-figure enterprise contracts and named Oura, Iveco Group, Fastweb+Vodafone, CNH and Zeiss among its customers. Those commercial claims were not independently audited in the opened sources, and the available evidence does not establish quantified productivity or revenue gains realized by customers.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>funding-and-ma</category><category>nebuly</category><category>series-a</category><category>enterprise-ai</category><category>ai-agents</category><category>planven</category></item><item><title>Apple researchers propose weighted state souping for long-context retrieval</title><link>https://dataphoenix.info/news/apple-weighted-state-souping</link><guid isPermaLink="true">https://dataphoenix.info/news/apple-weighted-state-souping</guid><description>Apple researchers proposed query-dependent weighting for cached state-space-model encodings, aiming to preserve reusable context as retrieval deteriorates across growing document and tool collections.</description><pubDate>Sat, 14 Nov 2026 20:30:06 GMT</pubDate><content:encoded>&lt;p&gt;Apple researchers have proposed weighted state souping, a method meant to preserve reusable, precomputed state-space-model representations while limiting the retrieval decline seen in crowded long contexts.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://arxiv.org/abs/2610.25621&quot; rel=&quot;noopener noreferrer&quot;&gt;paper&lt;/a&gt; builds on state souping, which stores the internal states created when a recurrent model encodes separate chunks. Rather than average every cached state equally, the new method derives a key from each chunk and a vector from the user’s query, then initializes the model with a softmax-weighted combination of those states. Each chunk gets one scalar weight, shared across layers and heads, that remains fixed while the model processes the query and generates an answer.&lt;/p&gt;
&lt;p&gt;The authors tested 1.5-billion-parameter &lt;a href=&quot;https://dataphoenix.info/news/apple-attention-to-mamba-distillation-research&quot;&gt;Mamba-2 models&lt;/a&gt; with mean, weighted and sparse souping against sequential Mamba-2 and a parameter-matched Transformer. The evaluations covered document question answering, RULER retrieval and variable-tracking tasks, and tool-use tests based on the Berkeley Function Calling Leaderboard.&lt;/p&gt;
&lt;p&gt;According to the paper, weighted souping maintained roughly stable performance as document or tool pools grew to 128 chunks, despite training on pools of no more than 24. For RULER, the authors capped training contexts at 2,048 tokens and evaluated lengths from 1,000 to 128,000 tokens, reporting robust results beyond the training length. The sources reviewed do not include an independent replication, and the paper presents task-specific curves rather than a single aggregate accuracy improvement.&lt;/p&gt;
&lt;p&gt;A sparse version retains only the highest-scoring chunks before applying softmax. The experiments used the top four chunks for document question answering and tool-use tests, and the top 20 for RULER. The authors reported that sparse souping matched or improved on the full weighted method in their tests, particularly on HotpotQA and MuSiQue. They also warned that a cutoff set too low can remove information needed to answer a question.&lt;/p&gt;
&lt;p&gt;The efficiency case hinges on where cached states are stored. The paper reports that sparse-soup encoding time can stay effectively constant as context grows when the number of fetched states remains fixed, though scoring still grows linearly with the number of chunks. Each cached state in the tested Mamba-2 setup occupied 33.6 MB in bfloat16. Moving states from system memory or disk can narrow the advantage, and the authors say retrieval can be slower than re-encoding in the worst case.&lt;/p&gt;
&lt;p&gt;The method requires training or additional fine-tuning; it is not an inference-only change. It is also sensitive to chunk boundaries, while independently encoded chunks cannot form cross-chunk representations during encoding. &lt;a href=&quot;https://machinelearning.apple.com/research/weighted-state-souping&quot; rel=&quot;noopener noreferrer&quot;&gt;Apple says&lt;/a&gt; the work was accepted at the NeurIPS 2026 workshop on personalized, aligned and long-term memory for AI systems. The sources reviewed do not describe a product release or deployment.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>apple</category><category>weighted-state-souping</category><category>long-context-retrieval</category><category>state-space-models</category><category>mamba-2</category></item><item><title>European Commission proposes Digital Fairness Act and direct enforcement powers</title><link>https://dataphoenix.info/news/eu-digital-fairness-act-proposal</link><guid isPermaLink="true">https://dataphoenix.info/news/eu-digital-fairness-act-proposal</guid><description>The European Commission&apos;s proposed Digital Fairness Act would restrict manipulative interfaces, some personalised advertising and pricing, and addictive design, alongside a plan to expand cross-border consumer enforcement.</description><pubDate>Sat, 14 Nov 2026 19:41:16 GMT</pubDate><content:encoded>&lt;p&gt;The European Commission has put forward a Digital Fairness Act that would restrict manipulative online interfaces, some profiling-based advertising and pricing, and digital features designed to encourage compulsive use. A companion proposal would give the Commission direct powers to investigate and enforce certain cross-border consumer-law cases.&lt;/p&gt;
&lt;p&gt;The measures &lt;a href=&quot;https://commission.europa.eu/law/law-topic/consumer-protection-law/digital-fairness-act_en&quot; rel=&quot;noopener noreferrer&quot;&gt;adopted by the Commission on 11 November&lt;/a&gt; are proposals, not law. The European Parliament and the Council of the European Union must examine them under the ordinary legislative procedure and may amend or reject them.&lt;/p&gt;
&lt;p&gt;Under &lt;a href=&quot;https://commission.europa.eu/document/download/0575d59f-554f-4a9e-b54b-e6b717175809_en?filename=Proposal+for+a+Digital+Fairness+Act_0.pdf&quot; rel=&quot;noopener noreferrer&quot;&gt;the proposed Digital Fairness Act&lt;/a&gt;, traders could not design or operate visual, auditory or other elements of online interfaces in ways likely to deceive or manipulate consumers or otherwise materially distort their economic behaviour. An annex would identify dark patterns considered unfair in all circumstances, while delegated powers would let the Commission update that list in response to technology, market developments, consumer research and enforcement experience.&lt;/p&gt;
&lt;p&gt;The proposal also reaches some personalised commercial practices. Profiling-based advertising could not use a consumer’s personal data when a trader knows with reasonable certainty that the person is a minor, or use special-category personal data protected under the General Data Protection Regulation. Traders would not have to process additional personal data solely to assess age, and online platforms already covered by the relevant section of the Digital Services Act would be excluded from that provision.&lt;/p&gt;
&lt;p&gt;Profiling-based personalised pricing would be prohibited when it is reasonably likely to harm a consumer and the trader knows with reasonable certainty that the person is a minor or is vulnerable because of age, credulity, disability or a serious circumstance that impairs judgment.&lt;/p&gt;
&lt;p&gt;For digital products intended, or reasonably foreseeable, to encourage compulsive or excessive use, traders would have to offer accessible controls that disable specified features separately or together. The proposal names autoplay and uninterrupted feeds, non-interaction notifications, engagement streaks or penalties, and persistent prompts that simulate interpersonal relationships.&lt;/p&gt;
&lt;p&gt;Other provisions would require traders to disclose the monetary value of purchases made with paid virtual currencies and provide controls to disable paid variable-reward systems. Traders would have to disclose reward probabilities, while paid variable-reward, pay-to-progress and pay-to-win mechanisms for minors would be barred subject to proposed carve-outs tied to a separate EU KIDS Act.&lt;/p&gt;
&lt;p&gt;Digital contracts are also covered. The draft would require reminders and active acceptance before a free trial converts to a paid service, reminders before automatic renewals, an online contract-termination facility, access to a human customer-service representative, and consumer control over the main parameters of automated contracting systems.&lt;/p&gt;
&lt;p&gt;The proposed regulation would repeal and replace the Unfair Commercial Practices Directive with a directly applicable regulation while preserving the core framework for unfair practices.&lt;/p&gt;
&lt;p&gt;A &lt;a href=&quot;https://commission.europa.eu/document/download/9c63e001-7dae-4d07-90ff-3efb4961490b_en?filename=Proposal+for+the+revision+of+the+CPC+Regulation.pdf&quot; rel=&quot;noopener noreferrer&quot;&gt;separate revision of the Consumer Protection Cooperation Regulation&lt;/a&gt; would let the Commission open proceedings when it reasonably suspects a widespread infringement with an EU-wide dimension under covered consumer laws. The Commission says the current system allows it to coordinate national authorities but not enforce consumer law directly, and that coordinated cases and penalties can move slowly.&lt;/p&gt;
&lt;p&gt;That enforcement proposal would set deadlines for key steps in coordinated national actions, strengthen the coordinated imposition and collection of fines, and require traders not established in the European Union that target EU consumers to appoint an EU legal representative able to cooperate with authorities. For Digital Fairness Act breaches investigated directly under the proposed framework, the Commission could impose a fine of up to 6% of a trader’s total worldwide annual turnover. That power is not currently in force.&lt;/p&gt;
&lt;p&gt;The Digital Fairness Act file is identified as 2026/0356 (COD). Under &lt;a href=&quot;https://www.consilium.europa.eu/en/council-eu/decision-making/ordinary-legislative-procedure/&quot; rel=&quot;noopener noreferrer&quot;&gt;the ordinary legislative procedure&lt;/a&gt;, Parliament and the Council must agree on the same text before adoption. The Commission draft says the regulation would enter into force 20 days after publication in the Official Journal and apply 12 months later, but those terms can change during negotiations.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>digital-fairness-act</category><category>dark-patterns</category><category>consumer-protection</category><category>european-union</category><category>online-platforms</category></item><item><title>Pure DC secures $780 million facility for Europe and Middle East build-out</title><link>https://dataphoenix.info/news/pure-dc-secures-780-million-growth-facility</link><guid isPermaLink="true">https://dataphoenix.info/news/pure-dc-secures-780-million-growth-facility</guid><description>Pure Data Centres Group secured a $780 million corporate growth facility for development and future site acquisitions, while a separate $1.1 billion accordion remains an optional expansion mechanism.</description><pubDate>Sat, 14 Nov 2026 19:35:32 GMT</pubDate><content:encoded>&lt;p&gt;Pure Data Centres Group said it has secured a $780 million platform growth facility to support its data-center development pipeline and future site acquisitions across Europe and the Middle East.&lt;/p&gt;
&lt;p&gt;The corporate-level financing gives Pure DC capital for its regional build-out, but the company has not identified the sites, capacity or construction timetable it will fund. The &lt;a href=&quot;https://via.ritzau.dk/pressemeddelelse/15269830/pure-data-centres?publisherId=90446&quot; rel=&quot;noopener noreferrer&quot;&gt;company’s announcement&lt;/a&gt; also did not disclose the facility’s maturity, pricing, security package, covenants or amount drawn at closing.&lt;/p&gt;
&lt;p&gt;A separate $1.1 billion accordion could expand the financing framework to about $1.9 billion if additional capital is committed. The reviewed records do not establish that the accordion is currently committed or drawn, so it is distinct from the $780 million facility already secured.&lt;/p&gt;
&lt;p&gt;Citigroup Global Markets, First Abu Dhabi Bank and Infranity joined ABN AMRO, Allianz Global Investors and SMBC in the six-member lender group. &lt;a href=&quot;https://www.cahill.com/news/firm-news/2026-11-05-cahills-digital-infrastructure-finance-group-represents-pure-data-centres-group-in-19-billion-platform-growth-facility&quot; rel=&quot;noopener noreferrer&quot;&gt;Pure DC’s transaction counsel&lt;/a&gt; said the corporate facility complements financing arranged for individual campuses.&lt;/p&gt;
&lt;p&gt;Chief Financial Officer Mike Schwartz said the facility strengthens Pure DC’s corporate funding base and provides a framework for raising additional capital. Pure DC said its announced financing over the preceding 12 months exceeds $6.2 billion; that figure is a company total rather than an independently audited measure in the reviewed records.&lt;/p&gt;
&lt;p&gt;Pure DC said its strategy is to secure sites with available power before developing capacity for AI training, inference and hyperscale cloud workloads. The reviewed materials do not independently establish future site delivery, capacity additions or demand outcomes.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>funding-and-ma</category><category>pure-dc</category><category>data-centers</category><category>infrastructure-financing</category><category>europe</category><category>middle-east</category></item><item><title>Hugging Face trains multi-vector retrieval for token pooling</title><link>https://dataphoenix.info/news/hugging-face-token-pooling-aware-training</link><guid isPermaLink="true">https://dataphoenix.info/news/hugging-face-token-pooling-aware-training</guid><description>Hugging Face published a method that trains retrieval models on full and pooled document representations, aiming to preserve more search quality while storing fewer vectors.</description><pubDate>Sat, 14 Nov 2026 19:30:16 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/token-pooling&quot; rel=&quot;noopener noreferrer&quot;&gt;Hugging Face published&lt;/a&gt; a pooling-aware training method on November 11, 2026, that optimizes a multi-vector retrieval model against both full and compressed document representations. In the author’s NanoBEIR@13 evaluation, it retained 97.8% of its own unpooled nDCG@10 score while storing 74.8% fewer document vectors with factor-four pooling. No independent replication was found in the opened evidence.&lt;/p&gt;
&lt;p&gt;The method targets the storage burden of multi-vector retrieval. These systems keep contextual vectors for individual document tokens and score them through late interaction, rather than compressing each document into one vector. Token pooling reduces the count by merging similar token vectors. Fewer vectors do not necessarily translate into an equivalent reduction in bytes, latency or cost, which also depend on vector dimensions, quantization, index structure and hardware.&lt;/p&gt;
&lt;p&gt;The pooler described in the post uses Ward hierarchical clustering. It repeatedly merges the pair of token-vector groups that produces the smallest increase in within-group squared distance, then replaces each resulting group with a normalized mean vector. With pool_factor=4, documents retain about one quarter of their original vectors by default, while query embeddings remain unpooled.&lt;/p&gt;
&lt;p&gt;Pooling-aware training adds the compressed representation to the objective instead of applying pooling only after ordinary training. The demonstrated configuration gives equal weight to retrieval losses from unpooled documents and factor-four pooled documents, with both losses updating the same encoder. &lt;a href=&quot;https://sbert.net/docs/package_reference/multi_vector_encoder/losses.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Sentence Transformers documentation&lt;/a&gt; says hierarchical pooling passes gradients through the cluster means with a straight-through estimator, while the discrete cluster assignments themselves are not differentiated.&lt;/p&gt;
&lt;p&gt;For the comparison, the author said both models started from ModernBERT-base and used one epoch over the same 120,000 MS MARCO triplets, seed 42, and matching initialization and optimization settings; only the loss changed. The &lt;a href=&quot;https://github.com/huggingface/sentence-transformers/blob/main/examples/multi_vector_encoder/training/token_pooling/training_cached_contrastive.py&quot; rel=&quot;noopener noreferrer&quot;&gt;public training script&lt;/a&gt; reflects those settings and creates a separate 1,000-example held-out split. The author also said no hyperparameter search was run.&lt;/p&gt;
&lt;p&gt;In the author-run NanoBEIR@13 benchmark, the ordinarily trained model scored 51.84 mean nDCG@10 without pooling and 49.03 with factor-four pooling, or 94.6% retention. The pooling-aware model scored 52.39 unpooled and 51.23 when pooled, or 97.8% retention. Its unpooled score was also 0.55 points higher, so the comparison does not isolate compression robustness from every possible change in overall model quality.&lt;/p&gt;
&lt;p&gt;The accompanying &lt;a href=&quot;https://github.com/huggingface/sentence-transformers/blob/main/examples/multi_vector_encoder/training/token_pooling/evaluate_token_pooling.py&quot; rel=&quot;noopener noreferrer&quot;&gt;evaluation script&lt;/a&gt; tests unpooled retrieval, fixed caps of 64, 32, 16, 8 and 4 document vectors, and pool factors from two through six. It reports mean nDCG@10, allowing the same model to be compared across several storage budgets.&lt;/p&gt;
&lt;p&gt;The method builds on earlier work rather than introducing token pooling itself. A &lt;a href=&quot;https://arxiv.org/abs/2409.14683&quot; rel=&quot;noopener noreferrer&quot;&gt;2024 paper by Benjamin Clavié, Antoine Chaffin and Griffin Adams&lt;/a&gt; described clustering-based token pooling for ColBERT-like indexes without architectural changes or query-time processing. Those findings do not independently validate Hugging Face’s 2026 training results, which remain limited to the disclosed models, training data and NanoBEIR setup.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>hugging-face</category><category>token-pooling</category><category>multi-vector-retrieval</category><category>sentence-transformers</category><category>information-retrieval</category><category>model-compression</category></item><item><title>Oumi expands Frontier AI Engineer into agent development</title><link>https://dataphoenix.info/news/oumi-frontier-ai-engineer-agent-development</link><guid isPermaLink="true">https://dataphoenix.info/news/oumi-frontier-ai-engineer-agent-development</guid><description>Oumi has expanded Frontier AI Engineer beyond model engineering to cover end-to-end agent development, while GRPO remains in private preview.</description><pubDate>Sat, 14 Nov 2026 18:44:26 GMT</pubDate><content:encoded>&lt;p&gt;Oumi has expanded &lt;a href=&quot;https://oumi.ai/press-release/agent-development&quot; rel=&quot;noopener noreferrer&quot;&gt;Frontier AI Engineer&lt;/a&gt; beyond model-focused engineering into an end-to-end workflow for AI agents. The updated product covers building, evaluating, training, deploying, monitoring and continuously improving those agents.&lt;/p&gt;
&lt;p&gt;The November 4 launch ties agent failures and production interactions to evaluation and post-training. Oumi says teams can turn those interactions into evaluation sets and training data, then use Frontier AI Engineer, or FAiE, to improve the models behind deployed agents.&lt;/p&gt;
&lt;p&gt;Oumi says FAiE can help teams define an agent’s objectives, tools and operating environment; design evaluations; analyze failures; curate data; select models and training strategies; and run experiments. The workflow is meant to identify whether a failure stems from prompts, tools, orchestration or the underlying model before a team chooses a remedy.&lt;/p&gt;
&lt;p&gt;Bring Your Own Harness is another part of the launch. Oumi says customers can keep their existing orchestration, tools, memory, environments and application logic while connecting those agents to its platform for evaluation, data curation and model training. Companies can therefore use the new training loop without rebuilding the software surrounding an agent.&lt;/p&gt;
&lt;p&gt;For post-training, Oumi added direct preference optimization, or DPO, alongside its existing supervised fine-tuning support. DPO trains a model from examples of preferred behavior. Oumi also says teams can chain training methods in multistage workflows, with FAiE helping select strategies, design rewards, monitor training and diagnose problems.&lt;/p&gt;
&lt;p&gt;Reinforcement learning through &lt;a href=&quot;https://dataphoenix.info/news/apple-theoretical-limits-language-model-alignment&quot;&gt;group relative policy optimization&lt;/a&gt;, or GRPO, is explicitly in private preview, and prospective users must request access. Oumi presents end-to-end agent development, Bring Your Own Harness, DPO and multistage training as launch capabilities without the preview qualifier. The company did not, however, provide separate availability dates or expressly label those features generally available. The announcement also did not disclose pricing, customer adoption figures or independent benchmark results.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>oumi</category><category>frontier-ai-engineer</category><category>ai-agents</category><category>model-training</category><category>product-launch</category></item><item><title>Hugging Face adds decision-model support to Transformers</title><link>https://dataphoenix.info/news/hugging-face-transformers-decision-model-support</link><guid isPermaLink="true">https://dataphoenix.info/news/hugging-face-transformers-decision-model-support</guid><description>Hugging Face added experimental Transformers interfaces that return choices, ordered scores and yes-or-no probabilities without generating response text.</description><pubDate>Sat, 14 Nov 2026 18:41:26 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://huggingface.co/blog/decision&quot; rel=&quot;noopener noreferrer&quot;&gt;Hugging Face has added decision-model support to Transformers&lt;/a&gt;, giving developers a common interface for asking compatible models for typed choices, ordered scores and yes-or-no probabilities without generating response text. The release adds &lt;code&gt;DecisionPipeline&lt;/code&gt;, &lt;code&gt;AutoModelForDecision&lt;/code&gt; and &lt;code&gt;DecisionConfig&lt;/code&gt;, together with a serving route and a Hub tag for finding compatible models.&lt;/p&gt;
&lt;p&gt;A request combines a shared &lt;code&gt;state&lt;/code&gt; with one or more named questions. The interface returns structured answers and probability distributions over the allowed responses. For compatible autoregressive language models, Transformers builds the decision prompts, evaluates answer-label logits in batched forward passes and applies softmax, rather than decoding an open-ended answer token by token.&lt;/p&gt;
&lt;p&gt;Each of the three question primitives serves a different need. A &lt;code&gt;choice&lt;/code&gt; question selects among named alternatives and returns the probability for each option plus a confidence value. A &lt;code&gt;score&lt;/code&gt; question represents an ordered scale and reports the probability-weighted expected value of the level index, alongside the distribution, legend and confidence. A &lt;code&gt;noul&lt;/code&gt; question handles a binary statement by returning the probability that it is true; the application decides what action, if any, to take from that probability.&lt;/p&gt;
&lt;p&gt;Developers can invoke the feature through &lt;code&gt;pipeline(&quot;decision&quot;, model=...)&lt;/code&gt; in &lt;a href=&quot;https://dataphoenix.info/news/hugging-face-transformers-model-integration-contracts&quot;&gt;Transformers&lt;/a&gt;, or load a model and its configuration with &lt;code&gt;AutoModelForDecision&lt;/code&gt; and &lt;code&gt;DecisionConfig&lt;/code&gt;. &lt;a href=&quot;https://huggingface.co/docs/transformers/main/en/serve-cli/serving&quot; rel=&quot;noopener noreferrer&quot;&gt;Transformers Serve exposes the same request format at &lt;code&gt;POST /v1/systemone&lt;/code&gt;&lt;/a&gt;. The endpoint wraps the served model in &lt;code&gt;DecisionPipeline&lt;/code&gt;, so the pipeline and HTTP route build the same prompts and return the same answer types. According to the serving documentation, it supports autoregressive models that have a chat template.&lt;/p&gt;
&lt;p&gt;Hugging Face’s published support-ticket example asks which department should handle a case, rates its urgency and estimates whether a refund is being requested. The response contains all three typed answers and their probabilities while reporting zero output tokens, reflecting that the model scored allowed answers instead of generating response text.&lt;/p&gt;
&lt;p&gt;The interface is not tied to a single architecture. Hugging Face says it can cover models with dedicated decision heads as well as systems built from generative-model backbones. A dedicated &lt;a href=&quot;https://huggingface.co/models?other=decision-model&amp;amp;sort=trending&quot; rel=&quot;noopener noreferrer&quot;&gt;&lt;code&gt;decision-model&lt;/code&gt; filter on the Hugging Face Hub&lt;/a&gt; gives developers a way to find compatible models.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://huggingface.co/docs/transformers/main/main_classes/pipelines&quot; rel=&quot;noopener noreferrer&quot;&gt;Transformers pipeline documentation&lt;/a&gt; lists &lt;code&gt;decision&lt;/code&gt; as a supported pipeline task but labels decision models, &lt;code&gt;DecisionPipeline&lt;/code&gt; and &lt;code&gt;DecisionConfig&lt;/code&gt; experimental. It says those APIs may change at any time without deprecation.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>hugging-face</category><category>transformers</category><category>decision-models</category><category>structured-inference</category></item><item><title>Perplexity releases BrowseSafe++ benchmark for browser-agent prompt injection</title><link>https://dataphoenix.info/news/perplexity-browsesafe-plus-plus-release</link><guid isPermaLink="true">https://dataphoenix.info/news/perplexity-browsesafe-plus-plus-release</guid><description>Perplexity released a 38,675-pair benchmark, evaluation tooling and two described detectors for testing whether browser agents can identify prompt injections embedded in webpages.</description><pubDate>Sat, 14 Nov 2026 18:35:56 GMT</pubDate><content:encoded>&lt;p&gt;Perplexity released &lt;a href=&quot;https://www.perplexity.ai/en-GB/hub/blog/browsesafe-plus-plus&quot; rel=&quot;noopener noreferrer&quot;&gt;BrowseSafe++&lt;/a&gt;, a benchmark of 38,675 matched attack-and-control webpage pairs designed to help developers test prompt-injection defenses in browser agents. The package includes a gated dataset, an evaluation workflow, a research SDK, a paper and a public leaderboard. Perplexity also described two BrowseSafe-Guard detectors, although the reviewed release materials did not provide downloadable model weights.&lt;/p&gt;
&lt;p&gt;The benchmark is split into 29,500 labelled training pairs and 9,175 test pairs, or 77,350 pages in total. Public test inputs contain one label-free page from each pair. Results are scored through the leaderboard instead of distributing the withheld labels.&lt;/p&gt;
&lt;p&gt;Indirect prompt injection occurs when instructions embedded in content read by an agent try to override the user’s task or redirect the agent’s actions. In BrowseSafe++, each attacked page is matched with a control that removes the injection while retaining the rest of the page. The records also include a legitimate user task and its expected outcome.&lt;/p&gt;
&lt;p&gt;Each pair is available as HTML, Markdown, an accessibility tree and a full-page screenshot. The &lt;a href=&quot;https://arxiv.org/abs/2610.17457&quot; rel=&quot;noopener noreferrer&quot;&gt;paper&lt;/a&gt; and &lt;a href=&quot;https://github.com/perplexityai/browsesafe-plus-plus&quot; rel=&quot;noopener noreferrer&quot;&gt;SDK documentation&lt;/a&gt; say the corpus spans 12 page types, 24 attack types, 18 placements, six attacker goals and 80 languages across page content and injection text.&lt;/p&gt;
&lt;p&gt;Perplexity and the paper’s authors say 32,187 pairs retain injections observed on production webpages, while 6,488 come from the earlier BrowseSafe benchmark. They describe collecting pages read by a deployed browser agent between Sept. 1, 2025, and Sept. 21, 2026. They then deduplicated the pages, masked personal information, filtered candidates and rewrote surrounding content before reinserting the original injection into the attacked copy. That production provenance has not been independently audited.&lt;/p&gt;
&lt;p&gt;The paper describes BrowseSafe-Guard as two binary classifiers fine-tuned from Qwen3.5-0.8B: one for HTML, Markdown or accessibility-tree text, and a multimodal model for screenshots. In the authors’ evaluation, the text detector reached F1 scores of 98.29 on HTML, 93.26 on Markdown and 92.98 on accessibility trees. They reported 93.32 F1 and a 6.08% false-positive rate for the screenshot detector. Those results have not been independently replicated.&lt;/p&gt;
&lt;p&gt;The authors estimate that the text detector costs about $0.18 per 1,000 pages using measured rented-H200 GPU time. They compare that with $18.95 to $180.61 for the frontier API models they tested at listed prices. The comparison is an issuer estimate rather than an independent cost audit.&lt;/p&gt;
&lt;p&gt;In a separate adaptive-attack experiment covering 1,024 tasks and seven target agents, the authors reported that original injections succeeded between 0.49% and 4.00% of the time. Their BrowseSafe-Adapt method revises injections using feedback from the target agent and raised the reported range to 30.66% to 95.41%. These experimental results have not been independently verified.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://huggingface.co/datasets/perplexity-ai/browsesafe-plus-plus&quot; rel=&quot;noopener noreferrer&quot;&gt;training split&lt;/a&gt; is offered under CC BY 4.0. Test inputs carry an evaluation-only license that prohibits training and redistribution. The SDK uses the MIT License, but its repository labels the code a research artifact rather than a Perplexity product and says support or maintenance is not planned. The repository also warns that a benchmark score is not a security guarantee and advises against opening benchmark pages in an agent connected to tools, credentials or real accounts.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>perplexity</category><category>browsesafe-plus-plus</category><category>prompt-injection</category><category>browser-agents</category><category>ai-security</category><category>benchmarks</category></item><item><title>AWS adds usage metering to SageMaker bidirectional streams</title><link>https://dataphoenix.info/news/aws-sagemaker-streaming-usage-pricing</link><guid isPermaLink="true">https://dataphoenix.info/news/aws-sagemaker-streaming-usage-pricing</guid><description>AWS Marketplace providers can now meter software use during SageMaker AI bidirectional streams over a companion metadata connection; endpoint infrastructure remains a separate charge.</description><pubDate>Sat, 14 Nov 2026 16:53:11 GMT</pubDate><content:encoded>&lt;p&gt;AWS has &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/introducing-usage-based-pricing-for-amazon-sagemaker-ai-bidirectional-streaming&quot; rel=&quot;noopener noreferrer&quot;&gt;introduced usage-based AWS Marketplace metering&lt;/a&gt; for Amazon SageMaker AI bidirectional streaming inference. The option lets model providers report software use as a live session runs. AWS named Deepgram as the launch partner.&lt;/p&gt;
&lt;p&gt;The metering applies to the provider’s AWS Marketplace software charge. Customers still pay SageMaker AI infrastructure costs for the &lt;a href=&quot;https://dataphoenix.info/news/aws-sagemaker-mig-gpu-partitioning&quot;&gt;instances running the endpoint&lt;/a&gt;, including while the endpoint remains provisioned. AWS’s &lt;a href=&quot;https://docs.aws.amazon.com/marketplace/latest/userguide/machine-learning-pricing.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Marketplace pricing documentation&lt;/a&gt; describes inference pricing as an additional software charge, not a replacement for those infrastructure costs.&lt;/p&gt;
&lt;p&gt;The mechanism adds a second WebSocket connection alongside the session’s existing data connection. A model server opts in by returning &lt;code&gt;X-Amzn-SageMaker-Metadata-Stream-Supported: true&lt;/code&gt; in the data WebSocket upgrade response. SageMaker AI then opens a companion metadata WebSocket, using &lt;code&gt;/invocations-bidirectional-stream-metadata&lt;/code&gt; by default. Both connections carry the same request ID, allowing the server to associate metering records with the correct session. The underlying &lt;a href=&quot;https://docs.aws.amazon.com/en_en/sagemaker/latest/dg/your-algorithms-inference-code.html&quot; rel=&quot;noopener noreferrer&quot;&gt;SageMaker container contract&lt;/a&gt; uses &lt;code&gt;/invocations-bidirectional-stream&lt;/code&gt; as the default data path.&lt;/p&gt;
&lt;p&gt;Providers &lt;a href=&quot;https://docs.aws.amazon.com/marketplace/latest/userguide/ml-inference-metering.html&quot; rel=&quot;noopener noreferrer&quot;&gt;report consumption during the session&lt;/a&gt; by sending one JSON record in each metadata WebSocket text frame. Every record uses the fixed &lt;code&gt;inference.count&lt;/code&gt; dimension and a nonnegative &lt;code&gt;ConsumedUnits&lt;/code&gt; value for usage since the previous record, rather than a running total. The provider defines one unit; AWS’s examples include a minute of speech-to-text audio or a block of text-to-speech characters. An optional client token makes duplicate records idempotent on the same connection.&lt;/p&gt;
&lt;p&gt;AWS recommends reporting units incrementally, followed by a final record for any unreported usage when the data session ends. AWS limits each metadata frame to one record and no more than one record per second per connection. Bidirectional sessions can last up to 30 minutes. The companion path addresses a response-header constraint: once a streaming response begins, the count in an HTTP response header cannot be updated as more usage accrues.&lt;/p&gt;
&lt;p&gt;If the metadata connection cannot open or drops, inference continues by default, but subsequent usage is not reported. Providers can instead set the metadata-stream &lt;code&gt;Required&lt;/code&gt; header to &lt;code&gt;true&lt;/code&gt;, making a failure stop the request or session. AWS named Deepgram as the launch partner and said its model server calculates metering automatically for usage-based speech-to-text and text-to-speech billing through AWS Marketplace.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>aws</category><category>amazon-sagemaker-ai</category><category>aws-marketplace</category><category>bidirectional-streaming</category><category>usage-based-pricing</category><category>deepgram</category></item><item><title>NVIDIA releases Dynamo deployment recipes for four Korean open models</title><link>https://dataphoenix.info/news/nvidia-dynamo-korea-open-model-recipes</link><guid isPermaLink="true">https://dataphoenix.info/news/nvidia-dynamo-korea-open-model-recipes</guid><description>NVIDIA released Dynamo recipes for A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3, outlining multi-GPU configurations for B200 systems.</description><pubDate>Sat, 14 Nov 2026 16:47:09 GMT</pubDate><content:encoded>&lt;p&gt;NVIDIA &lt;a href=&quot;https://blogs.nvidia.com/blog/nvidia-dynamo-korea-open-models&quot; rel=&quot;noopener noreferrer&quot;&gt;released Dynamo-optimized deployment recipes&lt;/a&gt; for A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from &lt;a href=&quot;https://dataphoenix.info/news/nvidia-korean-startup-investments-point2&quot;&gt;Motif Technologies&lt;/a&gt;. The configurations provide model-specific starting points for serving the four open models across multiple NVIDIA B200 GPUs.&lt;/p&gt;
&lt;p&gt;The recipes are optimized for B200 GPUs and can deliver “up to 2x performance out of the box,” according to NVIDIA, which also said they reduce deployment work. The performance figure is a company claim, not an independently verified benchmark result. NVIDIA’s release does not disclose a workload, baseline, per-model result or test environment.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.nvidia.com/dynamo/v-0-8-0/design-docs/overall-architecture&quot; rel=&quot;noopener noreferrer&quot;&gt;Dynamo is NVIDIA’s open-source distributed inference-serving framework&lt;/a&gt;. In an aggregated setup, one worker pool handles both prompt processing and token generation. A disaggregated setup divides those jobs between separate prefill and decode pools; &lt;a href=&quot;https://docs.nvidia.com/dynamo/knowledge-base/modular-components/router/disaggregated-serving&quot; rel=&quot;noopener noreferrer&quot;&gt;the frontend sends a request to a prefill worker first and then to a decode worker&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Dynamo’s &lt;a href=&quot;https://docs.nvidia.com/dynamo/v1.3.0/user-guides/kv-cache-aware-routing&quot; rel=&quot;noopener noreferrer&quot;&gt;KV-aware router&lt;/a&gt; weighs the relevant key-value cache state held by each worker against its decode load, then chooses the lowest-cost worker. A KV cache stores attention state from tokens the model has already processed, so reusing it can avoid repeating part of the computation. Workers can report cache changes to the router over ZMQ or NATS Core. Without those events, the router can estimate cache state from its own routing decisions.&lt;/p&gt;
&lt;p&gt;In disaggregated deployments, prefill workers must pass that cache state to the decode workers that continue generation. Dynamo uses the &lt;a href=&quot;https://github.com/ai-dynamo/nixl/blob/main/docs/nixl.md&quot; rel=&quot;noopener noreferrer&quot;&gt;NVIDIA Inference Xfer Library, or NIXL&lt;/a&gt;, for these point-to-point transfers. NIXL provides a common interface for moving data across GPU and CPU memory and storage backends, with plugins for transports including UCX.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://github.com/ai-dynamo/dynamo/blob/main/recipes/README.md&quot; rel=&quot;noopener noreferrer&quot;&gt;live recipe index&lt;/a&gt; lists aggregated and disaggregated B200 deployments for each model. A.X K2 uses eight GPUs in aggregated mode or 12 in a two-prefill, one-decode topology. K-EXAONE 2.0 uses four or eight GPUs, respectively; its disaggregated recipe uses NIXL over UCX on InfiniBand RDMA. Solar Open 2 uses four GPUs in aggregated mode or eight with one prefill and two decode workers. Motif-3 uses two GPUs in aggregated mode or eight in a three-prefill, one-decode layout.&lt;/p&gt;
&lt;p&gt;The repository currently lists Motif-3 in its experimental section, which it describes as functional but under active development and not yet fully validated for production use.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>nvidia</category><category>nvidia-dynamo</category><category>open-models</category><category>inference</category><category>multi-gpu</category><category>b200</category></item><item><title>Google makes Agent Substrate 1.0 generally available on GKE</title><link>https://dataphoenix.info/news/google-agent-substrate-1-0-gke-ga-cncf</link><guid isPermaLink="true">https://dataphoenix.info/news/google-agent-substrate-1-0-gke-ga-cncf</guid><description>Google released Agent Substrate 1.0 and made the runtime generally available on GKE, separating stateful agent sessions from reusable sandbox workers as CNCF onboarding continues.</description><pubDate>Sat, 14 Nov 2026 16:39:11 GMT</pubDate><content:encoded>&lt;p&gt;Google &lt;a href=&quot;https://cloud.google.com/blog/products/containers-kubernetes/open-infrastructure-for-autonomous-agents-comes-to-cncf&quot; rel=&quot;noopener noreferrer&quot;&gt;released Agent Substrate 1.0 and made it generally available on Google Kubernetes Engine&lt;/a&gt; on November 5, giving eligible Google Cloud customers a managed deployment path for the open-source agent runtime. The project has also been accepted into the Cloud Native Computing Foundation’s Sandbox, with onboarding still in progress.&lt;/p&gt;
&lt;p&gt;Agent Substrate separates an agent’s durable state from the compute that runs it. An Actor holds the agent’s memory, files and identity; a Worker is a sandbox that can be started in advance and reused rather than remaining permanently assigned to one Actor.&lt;/p&gt;
&lt;p&gt;When an application suspends an idle Actor, the runtime snapshots its memory and local files to storage, then releases the Worker. A later request restores that snapshot onto an available warm Worker. In this way, idle sessions can give up compute without discarding their state.&lt;/p&gt;
&lt;p&gt;Google says Agent Substrate can resume suspended Actors in under 500 milliseconds, reach 10 times the density of standard container runtimes and handle more than 500 suspend-and-resume activations per second. The opened evidence did not independently verify those performance figures.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://docs.cloud.google.com/kubernetes-engine/ai-ml/about-agent-substrate&quot; rel=&quot;noopener noreferrer&quot;&gt;Google’s GKE documentation&lt;/a&gt; says Agent Substrate supports gVisor and Cloud Hypervisor for sandbox isolation. Gateways can enforce inbound and outbound policy and inject credentials according to Actor identity, keeping secrets outside the agent sandbox.&lt;/p&gt;
&lt;p&gt;The GKE deployment requires a Standard cluster running version 1.37 or later and Workload Identity Federation for GKE. Autopilot clusters are not supported. Open network connections do not survive suspension, and GPU device passthrough into Actor containers is not supported.&lt;/p&gt;
&lt;p&gt;Google says the core project can run on Kubernetes clusters outside Google Cloud, while the GKE deployment tooling is limited to eligible Google Cloud customers. The documentation does not define eligibility or provide pricing.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://github.com/cncf/sandbox/issues/523&quot; rel=&quot;noopener noreferrer&quot;&gt;The CNCF Sandbox application record&lt;/a&gt; shows that the Technical Oversight Committee vote passed, Google and CNCF completed a contribution agreement, and an onboarding issue was created on October 20. Google described the donation as a move toward vendor-neutral ownership. The available record, however, does not establish that every onboarding task was complete by the November 5 announcement.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>infrastructure-and-compute</category><category>agent-substrate</category><category>google-cloud</category><category>gke</category><category>cncf</category><category>kubernetes</category><category>ai-infrastructure</category></item><item><title>Databricks adds Unity Catalog metric views to Power BI in private preview</title><link>https://dataphoenix.info/news/databricks-unity-catalog-metric-views-power-bi</link><guid isPermaLink="true">https://dataphoenix.info/news/databricks-unity-catalog-metric-views-power-bi</guid><description>Databricks has opened private-preview support for publishing Unity Catalog metric views to Power BI semantic models, with Unity Catalog controls retained through DirectQuery and scheduled republishing available through Power BI tasks.</description><pubDate>Sat, 14 Nov 2026 15:40:52 GMT</pubDate><content:encoded>&lt;p&gt;&lt;a href=&quot;https://www.databricks.com/blog/unity-catalog-metric-views-now-power-bi-one-unified-governed-semantic-layer-all-your-reporting&quot; rel=&quot;noopener noreferrer&quot;&gt;Databricks added private-preview support&lt;/a&gt; for publishing Unity Catalog metric views into Microsoft Power BI semantic models. Teams can keep measures and dimensions defined in Unity Catalog and make them available in Power BI without rebuilding those definitions there.&lt;/p&gt;
&lt;p&gt;Publishing turns a metric view into a table in the Power BI semantic model. Dimensions become columns, and eligible measures become explicit DAX calculated measures. Power BI queries the metric view rather than its underlying tables. Dimensions from joined tables appear in the same Power BI table without Power BI recreating those relationships.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/azure/databricks/partners/bi/power-bi/publish-metric-views&quot; rel=&quot;noopener noreferrer&quot;&gt;DirectQuery is the default storage mode&lt;/a&gt;. Visual queries run through Databricks SQL to the metric view, so Unity Catalog access controls and security policies apply. Import mode is also supported, but it copies data into Power BI. That copy is governed by Power BI permissions rather than Unity Catalog controls.&lt;/p&gt;
&lt;p&gt;Users can publish once from Catalog Explorer or add a &lt;a href=&quot;https://learn.microsoft.com/en-us/azure/databricks/jobs/tasks/powerbi&quot; rel=&quot;noopener noreferrer&quot;&gt;Power BI task to an Azure Databricks job&lt;/a&gt; to republish on a schedule. Each job run publishes or refreshes the semantic model in the selected Power BI workspace. Republishing updates existing DAX measures, adds newly declared measures and dimensions, and retains existing fields. The broader Power BI task feature is in public preview, separate from the private preview for metric-view publishing.&lt;/p&gt;
&lt;p&gt;Publication runs in one direction. Changes to a published measure in Power BI are not written back to Unity Catalog. Supported changes to the metric view can flow into Power BI when the model is republished.&lt;/p&gt;
&lt;p&gt;The preview also comes with model constraints. A publishable metric view must have at least one dimension and no more than one fact source. Measures can be dropped when their source columns are not declared as named dimensions. The documentation also lists unsupported constructs—including leading window measures, percentiles, parameters, wildcard measures and complex-type dimensions—that are dropped during publication. Publishing multiple metric views together can fail if they define the same measure name differently.&lt;/p&gt;
&lt;p&gt;&lt;a href=&quot;https://learn.microsoft.com/en-us/azure/databricks/partners/bi/power-bi/publish-metric-views&quot; rel=&quot;noopener noreferrer&quot;&gt;Access requires an Azure Databricks workspace enrolled in the private preview&lt;/a&gt;, followed by an administrator enabling the Publish Metric Views To Power BI preview. The available documentation does not specify pricing, a general-availability date, named preview customers or equivalent support for deployments on other clouds.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>databricks</category><category>unity-catalog</category><category>metric-views</category><category>power-bi</category><category>directquery</category><category>semantic-models</category></item><item><title>Fed report says market contacts cited AI as top potential shock</title><link>https://dataphoenix.info/news/fed-market-contacts-ai-financial-stability-risk</link><guid isPermaLink="true">https://dataphoenix.info/news/fed-market-contacts-ai-financial-stability-risk</guid><description>AI topped the list of potential shocks cited by 23 market contacts surveyed by New York Fed staff, but the result does not represent the Federal Reserve&apos;s views.</description><pubDate>Sat, 14 Nov 2026 15:36:22 GMT</pubDate><content:encoded>&lt;p&gt;Artificial intelligence was the most frequently cited potential shock among 23 market contacts surveyed by Federal Reserve Bank of New York staff, according to the Federal Reserve’s November Financial Stability Report. That puts AI-linked market, cyber and financing concerns at the top of this small group’s risk list, but it is not a Federal Reserve declaration that AI is the financial system’s largest threat.&lt;/p&gt;
&lt;p&gt;In the &lt;a href=&quot;https://www.federalreserve.gov/publications/files/financial-stability-report-20261113.pdf&quot; rel=&quot;noopener noreferrer&quot;&gt;official report&lt;/a&gt;, 87% of respondents cited AI as a risk over the next 12 to 18 months, up from 50% in the spring survey of 20 contacts. New York Fed staff collected the fall responses in September and October from professionals at broker-dealers, banks, investment funds and advisory firms.&lt;/p&gt;
&lt;p&gt;The report is explicit about the survey’s limits: The box summarizes respondents’ views and “should not be interpreted as representing the views of the Federal Reserve Board or the Federal Reserve Bank of New York.”&lt;/p&gt;
&lt;p&gt;Respondents’ concerns fell into three main channels. They pointed to elevated equity valuations that could reverse if AI-related earnings disappoint; &lt;a href=&quot;https://dataphoenix.info/news/uk-cyber-agency-ai-attacks-routine-2028&quot;&gt;AI-related cyber risk&lt;/a&gt; from sophisticated attackers or frontier AI models; and complex financing for AI investment. Some said stretched valuations, circular financing and debt-heavy funding for AI infrastructure could magnify losses from an equity correction and spill into business investment and household consumption.&lt;/p&gt;
&lt;p&gt;Most frequently cited did not mean most severe. A risk-asset correction was cited by 52% of contacts, higher long-term interest rates by 48%, and cyberattacks and fiscal debt sustainability by 43% each. Respondents nevertheless viewed cyberattacks and fiscal debt sustainability as carrying the most severe potential consequences.&lt;/p&gt;
&lt;p&gt;The Board’s formal monitoring framework tracks vulnerabilities across asset valuations, borrowing by businesses and households, financial-sector leverage and funding risks, while separately considering plausible near-term shocks raised through market outreach. In its own analysis, the report says a cyber incident could reduce market liquidity, erode investor and depositor confidence, force asset sales and spread across institutions through a disrupted critical third-party provider.&lt;/p&gt;
&lt;p&gt;The Board published the report on Nov. 13 and said it generally reflects information available through Oct. 28.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>policy-safety-and-incidents</category><category>federal-reserve</category><category>artificial-intelligence</category><category>financial-stability</category><category>market-risk</category><category>cybersecurity</category></item><item><title>AWS details how HyperPod data capture feeds model workflows</title><link>https://dataphoenix.info/news/aws-hyperpod-data-capture</link><guid isPermaLink="true">https://dataphoenix.info/news/aws-hyperpod-data-capture</guid><description>AWS explains how SageMaker HyperPod routes captured inference traffic into monitoring, debugging and model-improvement workflows.</description><pubDate>Sat, 14 Nov 2026 13:39:19 GMT</pubDate><content:encoded>&lt;p&gt;AWS has detailed how SageMaker HyperPod customers can turn captured inference traffic into data for monitoring, debugging and model improvement. Its &lt;a href=&quot;https://aws.amazon.com/blogs/machine-learning/build-a-model-improvement-flywheel-with-amazon-sagemaker-hyperpod-data-capture&quot; rel=&quot;noopener noreferrer&quot;&gt;technical walkthrough&lt;/a&gt; connects live inputs and outputs stored in Amazon S3 with Athena analysis, labeling, fine-tuning and redeployment on a HyperPod cluster.&lt;/p&gt;
&lt;p&gt;This is not a new launch. The post expands on a capability AWS introduced earlier in 2026. The company’s &lt;a href=&quot;https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-hyperpod-inference-release-notes.html&quot; rel=&quot;noopener noreferrer&quot;&gt;HyperPod Inference release notes&lt;/a&gt; date its introduction to May 6, while AWS &lt;a href=&quot;https://aws.amazon.com/about-aws/whats-new/2026/05/amazon-sagemaker-hyperpod-data-capture/&quot; rel=&quot;noopener noreferrer&quot;&gt;announced broader availability&lt;/a&gt; on May 20.&lt;/p&gt;
&lt;p&gt;Teams can enable &lt;a href=&quot;https://docs.aws.amazon.com/sagemaker/latest/dg/sagemaker-hyperpod-model-deployment-data-capture.html&quot; rel=&quot;noopener noreferrer&quot;&gt;HyperPod Data Capture&lt;/a&gt; independently at three points in the request path. Tier 1 records input and output payloads at the SageMaker endpoint and works with SageMaker Model Monitor. Tier 2 saves Application Load Balancer access logs, including request paths, client IP addresses and latency, but not request or response bodies. Tier 3 records payloads at the model pod, with controls for sampling, buffering and payload-size limits. All three tiers write to S3. A separate AWS infrastructure update added &lt;a href=&quot;https://dataphoenix.info/news/aws-sagemaker-mig-gpu-partitioning&quot;&gt;GPU partitioning for SageMaker real-time inference&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;Configuration requires at least one tier in the &lt;code&gt;dataCapture&lt;/code&gt; section of an &lt;code&gt;InferenceEndpointConfig&lt;/code&gt; or &lt;code&gt;JumpStartModel&lt;/code&gt; custom resource. Tier 1 and Tier 3 can use customer-managed AWS KMS keys. According to the documentation, Tier 2 relies on S3’s default encryption because Application Load Balancer access logs do not support KMS encryption.&lt;/p&gt;
&lt;p&gt;AWS says capture delivery to S3 is asynchronous and does not block inference. In a synthetic test of 400 routing requests at concurrency eight, the company measured p99 latency of 249.4 milliseconds without capture and 249.8 milliseconds through the Tier 3 proxy. Both paths handled 39.1 requests per second. These are AWS’s measurements and were not independently reproduced.&lt;/p&gt;
&lt;p&gt;For the walkthrough, AWS used synthetic support-ticket traffic to demonstrate a model-improvement loop. It queried captured records with Athena, applied a SQL labeling policy, fine-tuned the model and deployed a second endpoint. AWS reported that accuracy on a synthetic 280-item held-out set rose from 83.9% to 100%. On a separate 24-item hand-written probe, correct responses increased from 16 to 20. The exercise does not establish results for an external production deployment.&lt;/p&gt;
&lt;p&gt;Capturing full payloads also creates data-governance obligations. AWS warns that records may contain personal information and recommends sampling, lifecycle policies, KMS encryption for Tiers 1 and 3, and redaction before the data is reused for training. Enabling both payload tiers stores the same traffic twice. AWS estimates that Tier 3’s three sidecars request a combined 250 millicpu and 640 MiB per pod, in addition to S3 storage and request charges.&lt;/p&gt;
&lt;p&gt;AWS says the feature is available for EKS-orchestrated HyperPod clusters in every AWS Region where HyperPod is supported. AWS identifies S3 storage and requests, Athena scans and Tier 3 sidecar resources as potential cost categories.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>infrastructure-and-compute</category><category>aws</category><category>sagemaker-hyperpod</category><category>data-capture</category><category>inference-monitoring</category><category>amazon-s3</category></item><item><title>Databricks makes Apache Spark 4.3 available in Runtime 19</title><link>https://dataphoenix.info/news/databricks-spark-43-runtime-19</link><guid isPermaLink="true">https://dataphoenix.info/news/databricks-spark-43-runtime-19</guid><description>Databricks made Apache Spark 4.3 available in Runtime 19, bringing stateful processing to Real-Time Mode alongside history-preserving Auto CDC.</description><pubDate>Sat, 14 Nov 2026 13:18:49 GMT</pubDate><content:encoded>&lt;p&gt;Databricks &lt;a href=&quot;https://www.databricks.com/blog/introducing-apache-spark-43-now-available-databricks-runtime-19&quot; rel=&quot;noopener noreferrer&quot;&gt;made Apache Spark 4.3 available in Databricks Runtime 19&lt;/a&gt;, bringing stateful workloads to &lt;a href=&quot;https://dataphoenix.info/news/databricks-real-time-mode-spark-pipelines-ga&quot;&gt;Spark’s Real-Time Mode&lt;/a&gt;. Developers can now run streaming aggregations, duplicate removal and the JVM Scala/Java &lt;code&gt;transformWithState&lt;/code&gt; operator through the existing Structured Streaming APIs.&lt;/p&gt;
&lt;p&gt;Databricks said its tests of those three workloads produced P99 end-to-end latency of 136 milliseconds for &lt;code&gt;dropDuplicates&lt;/code&gt;, 139 milliseconds for window aggregation and 141 milliseconds for &lt;code&gt;transformWithState&lt;/code&gt;. The company limited those results to the tested workloads and said they were not a latency guarantee for every pipeline.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://spark.apache.org/docs/latest/streaming/real-time-mode.html&quot; rel=&quot;noopener noreferrer&quot;&gt;stateful Real-Time Mode implementation&lt;/a&gt; uses concurrent stage scheduling and pipelined shuffles. Producer and consumer stages can run at the same time, allowing consumers to begin reading shuffle output before producers finish.&lt;/p&gt;
&lt;p&gt;Spark 4.3 also adds Slowly Changing Dimension Type 2 support to Auto CDC through Python, SQL and Spark Connect. It preserves changes as versioned rows bounded by &lt;code&gt;__START_AT&lt;/code&gt; and &lt;code&gt;__END_AT&lt;/code&gt; metadata columns. Because the feature applies micro-batches with &lt;code&gt;MERGE&lt;/code&gt;, its target catalog connector must support Data Source V2 row-level operations. Databricks cited Apache Iceberg as a common choice and said a built-in file format such as Parquet does not qualify on its own.&lt;/p&gt;
&lt;p&gt;Project Feather adds two opt-in paths intended to cut fixed overhead on smaller workloads: shuffle-free single-task execution for eligible small, single-partition query fragments and Arrow-native Dataset caching. Other &lt;a href=&quot;https://spark.apache.org/releases/spark-release-4-3-0.html&quot; rel=&quot;noopener noreferrer&quot;&gt;Spark 4.3 release changes&lt;/a&gt; include REST endpoints for inspecting Spark Connect sessions and operations, operation IDs for request attribution, expanded JDBC and API coverage, Python &lt;code&gt;DataFrame.zip&lt;/code&gt;, SQL &lt;code&gt;JSON_TABLE&lt;/code&gt;, VARIANT transformation functions and &lt;code&gt;ASOF JOIN&lt;/code&gt;, plus preview nanosecond-precision timestamp types.&lt;/p&gt;
&lt;p&gt;The November 5 announcement was not Runtime 19’s first release. Databricks’ &lt;a href=&quot;https://docs.databricks.com/aws/en/release-notes/runtime/19&quot; rel=&quot;noopener noreferrer&quot;&gt;Runtime 19 release notes&lt;/a&gt; list a Spark 4.3.0-powered beta on June 15 and general availability on July 23, while the Apache Spark project &lt;a href=&quot;https://spark.apache.org/news/spark-4-3-0-released.html&quot; rel=&quot;noopener noreferrer&quot;&gt;announced the upstream Spark 4.3.0 release&lt;/a&gt; on October 30. The opened sources do not explain that pre-release versioning relationship in detail. Apache said the upstream release addressed more than 1,100 Jira tickets with contributions from more than 210 people.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>databricks</category><category>apache-spark</category><category>real-time-mode</category><category>structured-streaming</category><category>auto-cdc</category><category>project-feather</category></item><item><title>KORA announces AI-focused Linux desktop for 2027</title><link>https://dataphoenix.info/news/kora-ai-linux-desktop-announcement</link><guid isPermaLink="true">https://dataphoenix.info/news/kora-ai-linux-desktop-announcement</guid><description>KORA Systems announced a Linux desktop built around local AI models and agents for 2027, but its central capability and compatibility claims remain unverified.</description><pubDate>Sat, 14 Nov 2026 13:01:33 GMT</pubDate><content:encoded>&lt;p&gt;KORA Systems &lt;a href=&quot;https://kora.fun/?lang=en&quot; rel=&quot;noopener noreferrer&quot;&gt;announced KORA&lt;/a&gt;, a Linux-based desktop operating system for developers and creators, with PC availability planned for 2027. KORA is not yet shipping: the company is taking early-access signups and has not released a public build, exact launch date, price or final system requirements.&lt;/p&gt;
&lt;p&gt;The company describes KORA as a full desktop environment designed to move work between local and cloud-hosted models while keeping context intact across applications and agents. In its &lt;a href=&quot;https://press.kora.fun/&quot; rel=&quot;noopener noreferrer&quot;&gt;announcement materials&lt;/a&gt;, KORA says its proprietary Cortex engine schedules local-model inference across CPU, GPU and NPU resources and can extend that work to other computers on a local network. Neither capability has been independently demonstrated in a public build or reproducible benchmark.&lt;/p&gt;
&lt;p&gt;KORA also says its memory system stores context on the user’s computer, while its neural file system retrieves files by meaning inside permission-bound workspaces. According to the company, applications and agents will run in sandboxes with AppArmor enforcement and granular permissions. No independent security or privacy audit was identified for those mechanisms, and Cortex is not open source.&lt;/p&gt;
&lt;p&gt;KORA says it is working with Qualcomm Technologies to support the Snapdragon X Series. Qualcomm’s separate &lt;a href=&quot;https://www.qualcomm.com/developer/blog/2026/09/announcing-linux-on-snapdragon-x2-series-early-developer-preview&quot; rel=&quot;noopener noreferrer&quot;&gt;Snapdragon X2 Linux developer preview&lt;/a&gt; documents broader Linux enablement work for X2 laptops, but describes the support as an early developer preview rolling out in stages and continuing to mature toward production readiness. The preview does not mention KORA.&lt;/p&gt;
&lt;p&gt;The company further claims its compatibility technology can run selected Windows x86-64 applications inside Linux on Snapdragon hardware. It has not published an application compatibility matrix, performance results or a supported-hardware list for x86-64 PCs. KORA says native email, calendar, contacts, tasks, chat and media-editing applications are also in development, along with an SDK whose documentation and launch details have yet to be released.&lt;/p&gt;
&lt;p&gt;KORA grew out of Playtron and its Linux-based GameOS, and the company dates work on that foundation to November 2025. KORA says it will disclose pricing and its business model closer to its targeted 2027 availability.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>kora</category><category>linux</category><category>desktop-os</category><category>local-ai</category><category>ai-agents</category><category>snapdragon</category><category>developer-tools</category></item><item><title>IBM says sovereignty confidence is running ahead of proof</title><link>https://dataphoenix.info/news/ibm-digital-sovereignty-index-2026</link><guid isPermaLink="true">https://dataphoenix.info/news/ibm-digital-sovereignty-index-2026</guid><description>IBM says its first Digital Sovereignty Index found enterprise confidence in technology control running ahead of routine testing, while key survey methods remain undisclosed.</description><pubDate>Sat, 14 Nov 2026 12:58:03 GMT</pubDate><content:encoded>&lt;p&gt;IBM released its inaugural Digital Sovereignty Index on November 11, saying enterprises were more confident in their control of technology than their routine testing could substantiate.&lt;/p&gt;
&lt;p&gt;The company said 94% of surveyed enterprises had faced at least one sovereignty challenge in the previous three years, while 33% of respondents regularly or continuously tested their ability to demonstrate control. The percentages are IBM-reported survey results, not independently audited estimates of the wider enterprise population.&lt;/p&gt;
&lt;p&gt;Digital sovereignty, as IBM uses the term, is an organization’s ability to retain and demonstrate control over its AI, technology, operations and data, including when providers fail or legal and operational conditions change. &lt;a href=&quot;https://newsroom.ibm.com/2026-11-11-ibm-digital-sovereignty-index-in-the-ai-era,-control-is-no-longer-enough-enterprises-must-prove-it&quot; rel=&quot;noopener noreferrer&quot;&gt;IBM said the index assesses reported evidence across six dimensions&lt;/a&gt;: responsibility, trace, constraint, intervention, legitimacy and exit.&lt;/p&gt;
&lt;p&gt;Among surveyed enterprises that had encountered a sovereignty challenge, IBM said 81% had been required to provide formal evidence of control to regulators, customers, auditors or internal stakeholders. AI-related issues were the most frequently cited individual challenge, according to IBM, selected by 44% of respondents. The company’s examples included incorrect AI outputs, explainability problems, model or provider failures and governance concerns. The AI-related result sits alongside &lt;a href=&quot;https://dataphoenix.info/news/ibm-agent-identity-control-launch&quot;&gt;IBM’s rollout of identity and access controls for enterprise AI agents&lt;/a&gt;.&lt;/p&gt;
&lt;p&gt;IBM also reported that an average of 82% of respondents were very or extremely confident they could provide evidence of sovereign control across six evidence dimensions. By contrast, 26% of surveyed enterprises qualified for IBM’s Leading Digital Sovereignty tier, a separate maturity classification intended to reflect more comprehensive and integrated evidence. Confidence and tier placement are different measures, so the figures are not a like-for-like comparison.&lt;/p&gt;
&lt;p&gt;The &lt;a href=&quot;https://www.ibm.com/downloads/documents/us-en/187b1233b0643da5&quot; rel=&quot;noopener noreferrer&quot;&gt;IBM report&lt;/a&gt; describes a 20-minute online survey of 2,003 senior decision-makers responsible for areas including data, AI, technology, operations, governance, risk and compliance. It lists samples from the United States, United Kingdom, Germany, India, Japan, France and Saudi Arabia. IBM said the research also included 25 interviews with policymakers, regulators, public-sector officials and policy experts across those seven markets, plus perspectives from 17 senior customer leaders in five countries.&lt;/p&gt;
&lt;p&gt;The published materials do not disclose the survey’s field dates, recruitment and sampling design, weighting, response rate, question wording, fieldwork provider or uncertainty estimates. Nor do they clarify whether IBM alone sponsored, commissioned, designed, analyzed or fielded each stage of the work. IBM says participating organizations were not independently audited, and it has not published the questionnaire, raw data, cross-tabs or a reproducible formula for the index tiers. Those gaps prevent independent recalculation of the headline findings, and IBM’s materials do not establish that the combined results are globally representative.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>industry-and-adoption</category><category>ibm</category><category>digital-sovereignty</category><category>ai-governance</category><category>enterprise-technology</category><category>technology-risk</category></item><item><title>Apple study finds stronger gating can slow in-context learning in Mamba</title><link>https://dataphoenix.info/news/apple-gating-state-space-model-learning</link><guid isPermaLink="true">https://dataphoenix.info/news/apple-gating-state-space-model-learning</guid><description>Stronger gating can lead Mamba models to memorize fixed mappings before learning to retrieve prompt-specific answers, researchers found.</description><pubDate>Sat, 14 Nov 2026 12:53:03 GMT</pubDate><content:encoded>&lt;p&gt;Apple and Harvard researchers found that stronger gating can lead &lt;a href=&quot;https://dataphoenix.info/news/apple-attention-to-mamba-distillation-research&quot;&gt;Mamba state-space models&lt;/a&gt; to learn a parameter-stored shortcut before retrieving the correct answer from the current prompt. The &lt;a href=&quot;https://machinelearning.apple.com/research/gating-memorization-in-context&quot; rel=&quot;noopener noreferrer&quot;&gt;study of gating in state-space models&lt;/a&gt;, published by Apple in November, also found that the same mechanism often helped models handle sequences longer than those used in training.&lt;/p&gt;
&lt;p&gt;The finding describes an optimization tradeoff in the paper’s theoretical model and tested Mamba configurations, not a universal limit for every state-space model. Some retrieval failures can arise from training dynamics even when a model can represent the correct retrieval solution.&lt;/p&gt;
&lt;p&gt;The researchers separated two behaviors using tasks in which an input had a fixed underlying label but the label shown in each prompt could be randomly changed. In-weights memorization meant storing the usual input-to-label mapping in model parameters. In-context learning meant finding the matching input in the current prompt and returning the label attached there. Memorization could therefore score above chance while still ignoring the prompt-specific answer; perfect performance required retrieval.&lt;/p&gt;
&lt;p&gt;In the &lt;a href=&quot;https://arxiv.org/html/2609.16540v1&quot; rel=&quot;noopener noreferrer&quot;&gt;paper’s simplified Mamba-like model&lt;/a&gt;, gating exponentially reduced the influence of earlier tokens. Stronger gating preserved less distant context, weakening the training signal for retrieval while leaving the model able to learn the fixed mapping in its weights.&lt;/p&gt;
&lt;p&gt;The authors described the result as a memorization plateau, not necessarily a permanent outcome. Retrieval remained reachable in the reduced theoretical dynamics, but the estimated escape time grew sharply as the initial decay strengthened. Weak gating made retrieval easier to learn, while also allowing noise from more irrelevant tokens to accumulate in long contexts. Stronger decay reduced that noise but imposed locality, so useful information still had to remain within the model’s effective context.&lt;/p&gt;
&lt;p&gt;The multi-token retrieval experiments used eight-token keys and values, changed each value token with 10% probability, and trained on contexts containing up to 512 key-value pairs. Memorizing the fixed dictionary produced about 43% sequence accuracy. Across five random seeds, the default Mamba initialization did not leave that baseline during 100,000 training steps, while weaker-gated variants reached retrieval sooner. When evaluation contexts grew to as many as 2,048 pairs, stronger-gated models generalized better.&lt;/p&gt;
&lt;p&gt;A more naturalistic test fine-tuned 1.5-billion-parameter Mamba models on a variable-pool version of the Berkeley Function Calling Leaderboard after pretraining on 300 billion tokens. Within the tested range, weaker gating produced moderately better accuracy and fewer syntactically valid calls to tools absent from the prompt. Gating strength did not measurably change length generalization in that experiment, although all tested Mamba variants generalized better than the 2-billion-parameter Transformer baseline as tool pools grew.&lt;/p&gt;
&lt;p&gt;The authors said the study does not immediately establish a remedy. Simply weakening, reinitializing or removing gates may sacrifice long-context performance, and the results do not establish the same tradeoff across all architectures, model sizes, datasets or production deployments.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>foundation-models-and-research</category><category>apple</category><category>mamba</category><category>state-space-models</category><category>in-context-learning</category><category>machine-learning-research</category></item><item><title>Google opens Tabletop collaboration experiment via U.S. waitlist</title><link>https://dataphoenix.info/news/google-tabletop-collaboration-experiment</link><guid isPermaLink="true">https://dataphoenix.info/news/google-tabletop-collaboration-experiment</guid><description>Google Labs has introduced Tabletop, an experimental shared workspace that generates interactive group tools, through a waitlist limited to U.S. users aged 18 and older.</description><pubDate>Sat, 14 Nov 2026 12:34:49 GMT</pubDate><content:encoded>&lt;p&gt;Google Labs has &lt;a href=&quot;https://blog.google/innovation-and-ai/models-and-research/google-labs/tabletop/&quot; rel=&quot;noopener noreferrer&quot;&gt;opened a U.S. waitlist for Tabletop&lt;/a&gt;, an experimental real-time shared workspace, for people aged 18 and older. The limited rollout is not general availability.&lt;/p&gt;
&lt;p&gt;Groups can add links, files and ideas to a Tabletop space, then prompt &lt;code&gt;@tabletop&lt;/code&gt; or choose a starter template to generate an interactive tool. Google gives trip itineraries, dinner polls, expense splitters and tournament brackets as examples.&lt;/p&gt;
&lt;p&gt;When a space is ready, its creator can share an invite link, allowing group members to enter and use it at the same time. Google has not identified the AI model or models behind Tabletop.&lt;/p&gt;
&lt;p&gt;Other details also remain undisclosed. Google provided no waitlist or admission totals, usage or adoption figures, pricing, general-availability date, or plans to expand beyond the United States.&lt;/p&gt;
</content:encoded><dc:creator>DP News Team</dc:creator><category>news</category><category>products-and-features</category><category>google</category><category>google-labs</category><category>tabletop</category><category>collaboration</category><category>generative-ai</category></item></channel></rss>