# v2 Published articles for v2. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Asahi Linux Drops MIT and Goes Full GPL After AI Tools Were Used to Steal Unreleased Code DevFeed: [Asahi Linux Drops MIT and Goes Full GPL After AI Tools Were Used to Steal Unreleased Code]() Original publisher: [Read original article]() Author: Abhishek Prakash Published: 2026-11-08T12:56:22Z Content type: article Language: en Sources: [It's FOSS]() Topics: [generative-ai-tools](), [AI tools for developers]() Tags: [ai](), [gpl](), [linux](), [llms](), [mit](), [news](), [repository](), [v2]() ### AI overview Asahi Linux says it is moving all repositories from permissive licensing to GPL v2 or later after code from developers' unreleased personal forks was reportedly copied, sometimes rewritten with LLMs to obscure its origin. The article explains that MIT permits reuse in closed commercial products if its copyright notice is retained, while GPL requires distributed modified versions to provide source under the same license, reducing the scope for closed commercial forks and making reuse easier to detect. The change may hinder code laundering but cannot stop determined actors from copying or rewriting code. The excerpt says Asahi did not name the actors or explain how past contributions will be relicensed or whether all prior contributors agreed; the move has drawn criticism over possible contributor consent and may complicate BSD reuse. ### Source excerpt Every Asahi Linux repository is moving to GPL v2 or later, after bad actors lifted unreleased code and ran it through LLMs to hide where it came from. ## Kubeflow Trainer v2.4: Bringing Workload-Aware Scheduling to TrainJob DevFeed: [Kubeflow Trainer v2.4: Bringing Workload-Aware Scheduling to TrainJob]() Original publisher: [Read original article]() Author: Shristi, Andrey Velichkevich Published: 2026-11-06T06:00:00Z Content type: article Language: en Sources: [Kubeflow]() Topics: [scheduling](), [distributed-training](), [AI Infrastructure](), [Deadlock](), [Multi-GPU](), [Distributed Systems & Parallel Computing](), [CUDA](), [Boilerplate]() Tags: [ai](), [ai-ml](), [distributed](), [distributed-training](), [gpus](), [kubeflow](), [release](), [scheduling](), [trainer](), [v2]() ### AI overview Kubeflow Trainer v2.4 adds Workload-Aware Scheduling (WAS) for distributed TrainJobs. The controller creates Kubernetes Workload and PodGroup resources so workers can be admitted as an all-or-nothing gang, avoiding GPUs sitting idle while a job waits for capacity. Runtime configuration can give initializers a basic scheduling policy while GPU workers wait for their own gang, addressing dependency deadlocks. Admins choose one scheduling level per ReplicatedJob: TrainJob level for zone placement, ReplicatedJob level for network-block placement, or Job level for rack placement. In v2.4 these are independent PodGroups, not a nested hierarchy; Job-level groups support at most one replica and require Kubernetes v1.38. CompositePodGroup support is roadmap work, with Kubernetes 1.37 identified as alpha, and would allow nested placement and coordinated groups. Other listed additions include DRA support, per-TrainJob scheduler overrides, multi-slice TPU replicas, secured metrics, and OptimizationJob alpha integrations. The release also changes MPI image behavior and several configuration and upgrade requirements. ### Source excerpt Imagine launching an AI/ML distributed training job across 64 GPUs. Two frequent problems happen: ## Hardening Production Multi-Container Topologies with Docker Compose v2 DevFeed: [Hardening Production Multi-Container Topologies with Docker Compose v2]() Original publisher: [Read original article]() Author: Siva Sai Krishna Suryadevara Published: 2026-11-05T15:31:57Z Content type: article Language: en Sources: [HackerNoon]() Topics: [Docker Compose](), [container-networking](), [container escape](), [CI/CD]() Tags: [agent](), [ai](), [ai-agent](), [architecture](), [backend](), [blueprint](), [compose](), [container](), [container-hardening](), [container-security](), [docker](), [docker-compose](), [docker-networking](), [docker-secrets](), [docker-security](), [hackernoon-top-story](), [hardening](), [health](), [least-privilege](), [network](), [production](), [race-conditions](), [startup](), [v2](), [yaml-configuration]() ### AI overview The excerpt's production Compose blueprint makes startup readiness, network access, secrets, and resource limits explicit. Long-form `depends_on` gates the API on a TCP health check for Postgres, then on a one-shot migration's successful exit; this avoids startup races and concurrent migration runs when scaling the app. Separate `dmz` and internal backend networks leave only Nginx's port published and isolate the data stores. File-sourced Compose secrets stay out of environment variables and `docker inspect`, though outside Swarm their host file permissions control container access. The blueprint also applies CPU, memory, and process limits, non-root users, read-only filesystems, and dropped capabilities. It notes named-volume compatibility with image ownership, bind-mount ownership pitfalls, Redis cache headroom, and rootless Docker's restriction on publishing ports 80/443. It recommends validating the rendered configuration, service health, migration exit status, network isolation, and applied limits. ### Source excerpt Eliminate silent startup failures, race conditions, and network exposure in multi-container stacks using Docker Compose v2 health gates and least privilege. ## Coarse Semantic Blocks for Efficient Training-Free Sparse Attention DevFeed: [Coarse Semantic Blocks for Efficient Training-Free Sparse Attention]() Original publisher: [Read original article]() Author: Sosuke Kobayashi, Shoichiro Yamaguchi, Seiya Tokui Published: 2026-11-04T05:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [LLM optimization](), [Large Language Model](), [long-context]() Tags: [accelerate](), [attention](), [blocks](), [cs-ai](), [cs-cl](), [cs-lg](), [decoding](), [hybrid](), [llms](), [long-context](), [nvidia](), [nvidia-h200](), [prefill](), [qwen3-5-4b](), [ruler](), [semantic](), [training](), [v2]() ### AI overview The method speeds up prefill and decoding by clustering prompt keys into large semantic blocks, including tokens from noncontiguous positions. Each query scores block centroids, attends to original keys and values in selected blocks, and is grouped with queries selecting the same blocks for efficient prefill. On one NVIDIA H200, including index construction and use, it achieves 2.3-2.4x prefill and 1.7-2.0x decoding speedups over dense attention at 512K tokens, and 2.9-3.0x prefill speedups at 1M. Qwen3.5-35B-A3B stays close to dense attention on five RULER tasks through 1M, LongBench-v2 overall, and MRCR through 256K. Qwen3.5-4B loses accuracy mainly on multi-key retrieval, though it beats dense Qwen3.5-2B on all three benchmarks at similar 512K latency. Larger blocks with a larger token budget are reported as both faster and more accurate than smaller-block, smaller-budget settings. ### Source excerpt arXiv:2611.04313v1 Announce Type: new Abstract: We introduce a training-free sparse attention method that builds a prompt-specific key index to accelerate both prefill and decoding. The method clusters the keys into large semantic blocks that group tokens from noncontiguous positions. Each query position selects blocks by scoring their centroids and attends to the original keys and values within them. Queries that select the same block are processed together, so prefill stays fast even though each query chooses its own blocks. Large blocks make centroid scoring and this grouped processing cheap, so each query can attend to more tokens in the same time. Larger blocks with a larger token budget can thus be both faster and more accurate on long-context tasks than smaller blocks with a smaller budget. We evaluate two hybrid LLMs, Qwen3.5-4B and 35B-A3B, on one NVIDIA H200. Including the cost of building and using the index, whole-model wall-clock speedups over dense attention are 2.3-2.4x for prefill and 1.7-2.0x for decoding at 512K tokens, and 2.9-3.0x for prefill at 1M. The 35B-A3B model stays close to dense attention on five RULER tasks through 1M, on LongBench-v2 overall, and on MRCR through 256K. The 4B model loses accuracy mainly on multi-key retrieval, yet outperforms dense Qwen3.5-2B, whose latency is similar at 512K, on all three benchmarks. ## NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt DevFeed: [NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt]() Original publisher: [Read original article]() Author: Maryam Motamedi; Piotr Żelasko; Kunal Dhawan; Nikolay Karpov; Nune Tadevosyan; Meline Mkrtchyan; Nithin Rao Koluguri; Francesco Published: 2026-10-28T14:58:17Z Content type: release Language: en Sources: [Hugging Face]() Topics: [Nemotron](), [voice ai](), [Fine-tuning](), [Prompt Engineering](), [Enhancing AI model performance](), [Internationalization (i18n)](), [Open Source]() Tags: [applications](), [audio](), [data](), [data-pipeline](), [domain](), [fine-tuning](), [format](), [granary-v2](), [leaderboard](), [model](), [nemotron](), [nemotron-3-5](), [nemotron-3-5-transcribe](), [nvidia](), [nvidia-nemotron](), [open](), [pipeline](), [prompt](), [transcription](), [v2]() ### AI overview NVIDIA Nemotron 3.5 Transcribe combines multilingual transcription, speaker attribution, translation, and audio Q&A in one promptable open-weight checkpoint. It supports transcription in 40 languages, translation between English and 31 languages, up to eight speakers, and recordings up to an hour; language support varies by task, and audio Q&A is trained in English. It is designed for recorded audio, with vLLM serving available and quantized local deployment described as forthcoming. Granary v2 provides multilingual manifests and a reusable curation pipeline, while NeMo Speech tools support fine-tuning. In an Indian medical ASR case study, fine-tuning reduced held-out WER from 22.44% to 12.17% versus the best-prompt baseline, with a slight LibriSpeech test-clean WER increase from 1.54% to 1.69%; a medical specialist still led on exact brand-name matches. The source recommends evaluating domain entities and general-speech retention. ### Source excerpt # NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt ## One open model for multilingual transcription, speaker attribution, translation, and audio Q&A, with top leaderboard results, plus the open Granary v2 data pipeline and fine-tuning recipes to adapt it. A transcript can capture every spoken word and still leave users or applications wondering who said what, how to translate it, or how to format domain-specific jargon. ## dbt without the warehouse (or the bill): DuckDB end to end DevFeed: [dbt without the warehouse (or the bill): DuckDB end to end]() Original publisher: [Read original article]() Author: Daniel Poppy Published: 2026-10-27T17:36:00Z Content type: news Language: en Sources: [dbt Labs]() Topics: [DuckDB]() Tags: [2026](), [about](), [adapter](), [agent](), [aggregation](), [all](), [analysis](), [analytics](), [analytics-engineering](), [bill](), [built-in](), [change](), [dbt](), [duckdb](), [learn](), [offline](), [profile](), [small](), [v2](), [warehouse](), [with]() ### AI overview dbt v2 includes a built-in DuckDB adapter, and connecting it requires setting the profile output to duckdb and specifying a file path; omitting the path creates a temporary in-memory database. In a live demonstration, Hannes Mühleisen ran dbt fully offline over about 10 GB of compressed Wikipedia page-view CSVs stored locally. A build with window-function calculations across millions of article titles and an aggregation over roughly 300 million rows finished in about three seconds. DuckDB can query CSV, Parquet, and Iceberg files directly without importing them, and includes SQL features such as GROUP BY ALL, UNION BY NAME, and PIVOT. The excerpt presents DuckDB as a single-machine option for many datasets, while previewing DuckDB 2.0 features expected around October 2026. ### Source excerpt dbt v2 ships with a built-in DuckDB adapter, so connecting dbt to a local, offline warehouse takes only a small profile change. ## dbt v2 lands: faster parsing, one shared adapter layer, and a docs experience that skips the giant manifest.json DevFeed: [dbt v2 lands: faster parsing, one shared adapter layer, and a docs experience that skips the giant manifest.json]() Original publisher: [Read original article]() Author: Daniel Poppy Published: 2026-10-26T21:58:00Z Content type: news Language: en Sources: [dbt Labs]() Topics: [adapter pattern](), [Open Source](), [JSON](), [Tree-sitter](), [User Experience](), [GeoJSON]() Tags: [adapter](), [adapters](), [ai](), [analytics-engineering](), [autofix](), [bigquery](), [blast-radius](), [claude](), [column](), [compatibility](), [core](), [dbt](), [docs](), [faster](), [files](), [json](), [layer](), [learn](), [manifest](), [parsing](), [v2]() ### AI overview dbt v2 aims to bring v1 and v2 toward one engine and binary, while offering practical upgrade checks in v1.12. The excerpt says v1.12 includes a v2 parser for compatibility checks and an autofix deprecations command; a live demo resolved 22 flagged issues in about five minutes. In v2, adapters share a layer built on the ADBC driver standard and ship with the engine, avoiding the dependency conflicts that could arise when installing separate v1 adapters. On a synthetic 1,800-model project, parsing took about 3 seconds in v2 versus 10-13 seconds in v1. SQL typo detection took about 1.5 seconds on the OSS distribution, which round-trips to the warehouse, versus 0.06 seconds on the full distribution, which understands SQL directly. Docs v2 replaces a large manifest.json with queryable Parquet files; demos used Claude to query lineage and downstream impact. The excerpt notes that migration will take years and that v2 was generally available on most adapters at the time of the talk. ### Source excerpt dbt v2 bundles adapters, speeds up parsing, and replaces manifest.json with smaller, queryable files for people and AI. ## Depth Anything: Models, Comparisons, and How to Use It DevFeed: [Depth Anything: Models, Comparisons, and How to Use It]() Original publisher: [Read original article]() Author: Dikshant Shah Published: 2026-10-26T19:18:06Z Content type: tutorial Language: en Sources: [Roboflow]() Topics: [Machine learning](), [3D](), [Tool](), [Documentation]() Tags: [comparisons](), [computer-vision](), [deploy](), [how-to](), [images](), [models](), [run](), [streams](), [v1](), [v2](), [v3](), [video](), [workflows]() ### AI overview Depth Anything V1 and V2 estimate relative depth from a single image; V2 uses a teacher trained on synthetic images with precise depth labels, then pseudo-labels real-world images, improving fine detail and boundaries. Metric-depth variants exist, but accuracy may vary outside their training domains. V3 extends the task to spatially consistent 3D geometry using depth, per-pixel rays, and camera outputs, combining multiple views or video and supporting camera-pose estimation and reconstruction. Roboflow Workflows offers V2 and V3 blocks that can be assembled manually or generated with Roboflow Agent, then tested and published. Deployment options include a credit-billed Serverless Cloud API, which avoids managing a dedicated GPU, and running on users' own hardware; generated code supports images, video, webcam, and RTSP inputs. ### Source excerpt Compare Depth Anything V1, V2, and V3, then run Depth Anything V2 or V3 in Roboflow Workflows and deploy it to images, video, or RTSP streams. ## Actions Runner v2.338.2: Node 24 internal operations and broker connectivity probes DevFeed: [Actions Runner v2.338.2: Node 24 internal operations and broker connectivity probes]() Original publisher: [Read original article]() Author: github-actions\[bot\] Published: 2026-10-26T14:45:48Z Content type: release Language: en Sources: [GitHub Action Runners]() Topics: [self-hosted runner](), [GitHub](), [Dependabot](), [CI/CD]() Tags: [2026](), [actions](), [buildx](), [codeql](), [cryptography](), [dependabot](), [docker](), [dotnet](), [encoding](), [eslint](), [github](), [github-actions](), [node](), [node-js](), [repository](), [sdk](), [update](), [v2](), [version](), [yaml]() ### AI overview Actions Runner v2.338.2 updates dependencies and build tooling, adds broker-session and WebSocket probe changes, and moves internal operations to Node 24. The listed functional changes include accepting `$/` self-repository references in action manifests, including the owned session ID when deleting broker sessions, and implementing broker-listener WebSocket connectivity probes. The release also updates Docker, Buildx, the .NET SDK, and several JavaScript and .NET packages. Installation instructions provide packages for Windows x64 and arm64, macOS x64 and arm64, and Linux x64, arm64, and arm; Windows guidance recommends placing the runner folder at the drive root to help avoid permission and long-path issues. The notes state that releases roll out progressively, so availability may vary by enterprise, organization, or repository. ### Source excerpt What's Changed Bump @typescript-eslint/eslint-plugin from 8.67.0 to 8.68.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4661 chore: update Node versions by @github-actions[bot] in #4665 Bump browserslist from 4.28.5 to 4.28.9 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4677 Update Docker to v29.8.0 and Buildx to v0.37.0 by @github-actions[bot] in #4676 Update Node.js 20 removal date to September 23rd, 2026 by @salmanmkc in #4685 Bump @typescript-eslint/parser from 8.68.0 to 8.69.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4673 Bump Microsoft.DevTunnels.Connections from 1.3.51 to 1.3.56 by @dependabot[bot] in #4667 Bump github/codeql-action from 4.37.8 to 4.37.9 by @dependabot[bot] in #4664 Bump lint-staged from 17.3.0 to 17.5.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4663 Bump System.Formats.Asn1 and System.Security.Cryptography.Pkcs by @dependabot[bot] in #4688 Bump @typescript-eslint/parser from 8.69.0 to 8.70.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4705 Bump js-yaml from 4.3.1 to 4.3.2 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4696 Bump lint-staged from 17.5.0 to 17.5.1 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4715 chore: update Node versions by @github-actions[bot] in #4716 Bump github/codeql-action from 4.37.9 to 4.38.0 by @dependabot[bot] in #4713 Bump @typescript-eslint/eslint-plugin from 8.68.0 to 8.70.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4714 Update Buildx to v0.37.1 by @github-actions[bot] in #4712 Bump System.Text.Encoding.CodePages from 10.0.11 to 10.0.12 by @dependabot[bot] in #4701 Update dotnet sdk to latest version @8.0.425 by @github-actions[bot] in #4717 Bump github/codeql-action from 4.38.0 to 4.38.1 by @dependabot[bot] in #4731 Update Docker to v29.8.1 by @github-actions[bot] in #4730 Accept $/ self-repository references in action manifests by @nodeselector in #4669 Bump Microsoft ## How Virgin Media O2 cut compute costs 30% by giving dbt a memory DevFeed: [How Virgin Media O2 cut compute costs 30% by giving dbt a memory]() Original publisher: [Read original article]() Author: Daniel Poppy Published: 2026-10-26T14:04:00Z Content type: news Language: en Sources: [dbt Labs]() Topics: [cost-optimization](), [resource orchestration](), [Tool]() Tags: [2026](), [adopting](), [ai](), [ai-workflows](), [analytics](), [analytics-engineering](), [bigquery](), [ci-cd](), [ci-cd-pipeline](), [compute](), [compute-costs](), [costs](), [dbt](), [dependency](), [enterprise](), [learn](), [model](), [scale](), [state](), [times](), [v2]() ### AI overview Virgin Media O2 reports that migrating its BigQuery dbt platform to dbt v2 with dbt State cut compute costs by about 30% and model run times by about 35% across jobs. Its platform serves 300+ users across 100+ projects and 5,500+ models; previously, even small changes could trigger redundant full rebuilds. In a 1,424-model test, dbt State reused 32 unchanged models, reducing runtime from about 1 hour 53 minutes to 1 hour 13 minutes and BigQuery consumption from roughly 4,500 to 3,400 slots per hour. The team says migration required addressing accumulated technical debt and outdated packages. Full production rollout still depends on aligning dbt v2 with custom FinOps configurations that route work between on-demand and reservation pricing. ### Source excerpt Virgin Media O2 cut dbt compute costs 30% and model run times 35% by adopting dbt v2 and dbt State at scale. ## How DocuSign cut analytics engineering time in half by pairing Claude Code with dbt v2, dbt MCP, and agent skills DevFeed: [How DocuSign cut analytics engineering time in half by pairing Claude Code with dbt v2, dbt MCP, and agent skills]() Original publisher: [Read original article]() Author: Daniel Poppy Published: 2026-10-26T14:04:00Z Content type: news Language: en Sources: [dbt Labs]() Topics: [Claude Code](), [Agent Skills](), [developer-productivity](), [DataOps](), [Pull Request]() Tags: [2026](), [agent](), [agent-skills](), [aggregation](), [ai](), [analytics](), [analytics-engineering](), [approach](), [asset](), [claude-code](), [dbt](), [docusign](), [half](), [learn](), [mcp](), [pairing](), [productivity](), [skills](), [time](), [v2](), [workflow]() ### AI overview DocuSign's workflow uses Claude Code to coordinate analytics engineering across three phases: discovery, dbt development, and orchestration. dbt Agent Skills provide project conventions, dbt MCP supplies metadata and lineage, and dbt v2 compiles and validates SQL before a pull request. Phase-specific retries can address failures such as compile errors or skipped unit tests without restarting the full workflow. The demo produced a quarterly aggregation model, YAML, and orchestration job in an isolated branch, with human review required before merge. Gupta reported a 50% productivity gain from requirements to validated, review-ready assets in a pilot, plus more consistent naming, testing, and semantic patterns. Guardrails included capping test runs at 1,000 rows; governance for teams' custom skills was described as early and still evolving. ### Source excerpt DocuSign wired Claude Code, dbt Agent Skills, dbt MCP, and dbt v2 into one workflow and measured a 50% productivity gain. ## GitHub Actions Runner v2.338.1 release DevFeed: [GitHub Actions Runner v2.338.1 release]() Original publisher: [Read original article]() Author: github-actions\[bot\] Published: 2026-10-22T14:15:24Z Content type: release Language: en Sources: [GitHub Action Runners]() Topics: [self-hosted runner](), [CI/CD](), [Dependabot](), [GitHub](), [Dockerfile]() Tags: [2026](), [actions](), [buildx](), [codeql](), [dependabot](), [docker](), [dotnet](), [encoding](), [eslint](), [github](), [github-actions](), [js](), [lint](), [node](), [node-js](), [repository](), [sdk](), [security](), [update](), [v2]() ### AI overview Actions Runner 2.338.1 updates dependencies and build tooling, moves internal runner operations to Node 24, and adds broker-listener WebSocket connectivity probes. Other listed changes include support for `$/` self-repository references in action manifests, retaining the owned session ID when deleting broker sessions, and updating the Node.js 20 removal date to September 23, 2026. The excerpt provides download and extraction instructions for Windows x64/arm64, macOS x64/arm64, and Linux x64/arm64/arm; Windows guidance recommends installing under the drive root to reduce service identity permission and long-path issues. Progressive release means availability may vary by enterprise, organization, or repository. ### Source excerpt What's Changed Bump @typescript-eslint/eslint-plugin from 8.67.0 to 8.68.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4661 chore: update Node versions by @github-actions[bot] in #4665 Bump browserslist from 4.28.5 to 4.28.9 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4677 Update Docker to v29.8.0 and Buildx to v0.37.0 by @github-actions[bot] in #4676 Update Node.js 20 removal date to September 23rd, 2026 by @salmanmkc in #4685 Bump @typescript-eslint/parser from 8.68.0 to 8.69.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4673 Bump Microsoft.DevTunnels.Connections from 1.3.51 to 1.3.56 by @dependabot[bot] in #4667 Bump github/codeql-action from 4.37.8 to 4.37.9 by @dependabot[bot] in #4664 Bump lint-staged from 17.3.0 to 17.5.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4663 Bump System.Formats.Asn1 and System.Security.Cryptography.Pkcs by @dependabot[bot] in #4688 Bump @typescript-eslint/parser from 8.69.0 to 8.70.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4705 Bump js-yaml from 4.3.1 to 4.3.2 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4696 Bump lint-staged from 17.5.0 to 17.5.1 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4715 chore: update Node versions by @github-actions[bot] in #4716 Bump github/codeql-action from 4.37.9 to 4.38.0 by @dependabot[bot] in #4713 Bump @typescript-eslint/eslint-plugin from 8.68.0 to 8.70.0 in /src/Misc/expressionFunc/hashFiles by @dependabot[bot] in #4714 Update Buildx to v0.37.1 by @github-actions[bot] in #4712 Bump System.Text.Encoding.CodePages from 10.0.11 to 10.0.12 by @dependabot[bot] in #4701 Update dotnet sdk to latest version @8.0.425 by @github-actions[bot] in #4717 Bump github/codeql-action from 4.38.0 to 4.38.1 by @dependabot[bot] in #4731 Update Docker to v29.8.1 by @github-actions[bot] in #4730 Accept $/ self-repository references in action manifests by @nodeselector in #4669 Bump Microsoft ## Qiskit SDK v2.6 release summary DevFeed: [Qiskit SDK v2.6 release summary]() Original publisher: [Read original article]() Author: Qiskit Team Published: 2026-10-22T04:00:00Z Content type: release Language: en Sources: [IBM Research]() Topics: [SDK](), [Release notes](), [version]() Tags: [api](), [features](), [new-features](), [qiskit](), [quantum-software-updates](), [release](), [release-summary](), [sdk](), [summary](), [technical](), [top](), [updates](), [v2]() ### AI overview Qiskit SDK v2.6 expands the C API with a customizable pass manager, QPY serialization, and user-defined operations, giving compiled-language programs more control over circuit compilation and exchange without a Python round-trip. Passes can declare input and output intermediate representations so pipelines can be type-checked across multiple stages. C API QPY reads and writes native circuits compatible with Python's qiskit.qpy; format v18 is now the default and may reduce files with large or numerous registers, but older Qiskit releases may not load v18 files. The release also advances work toward standalone C API use without Python installed, though that support is not yet complete. Python 3.10 remains supported in v2.6 and is dropped in 2.7; source builds now require Rust 1.89. The C API is not yet declared stable, but this release's changes are additive. ### Source excerpt Technical release summary for Qiskit SDK v2.6, including updates on top new features. ## EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver DevFeed: [EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver]() Original publisher: [Read original article]() Author: Dastin Huang, Baran Nama, Salaheddin Alzubi Published: 2026-10-21T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence](), [arXiv: Machine Learning]() Topics: [Multi-step AI processes](), [Agent Skills](), [Benchmark](), [Machine learning]() Tags: [agent](), [agentic](), [benchmark](), [benchmarks](), [claude](), [coding](), [coding-agent](), [cost](), [cs-ai](), [cs-cl](), [cs-lg](), [evolution](), [harness](), [instructions](), [model](), [pass](), [plugin](), [run](), [sequence](), [skills](), [stateful](), [tasks](), [v2]() ### AI overview EvoSkill v2 improves an agent by evolving its instructions while leaving its model and harness unchanged. A benchmark adaptor converts a user's benchmark description and grader into a plugin; a solver generates training rollouts, and one stateful coding agent with bash rewrites the skills each epoch. The authors distinguish this from SkillOpt's sequence of single model calls and SkillBoost's fresh agent each round. Their approach measures rollout variability, gives the evolver multiple attempts per training task, and accepts every epoch's skills rather than choosing based on noisy scores. With fine-grained ground truth, held-out pass rates exceeded the stronger baseline by 7.7 percentage points on SpreadsheetBench v2 and 5.4 on τ³-bench banking. The evolver also cost less per percentage point gained than either baseline optimizer in both benchmarks and feedback settings. When Claude Haiku 5.5 powered both solver and evolver, the banking experiment reported a gain of up to 21.5 percentage points. ### Source excerpt arXiv:2610.25931v1 Announce Type: cross Abstract: EvoSkill v2 is a three-agent system that edits the instructions (skills) an agent is given, leaving its model and harness intact. The user starts a run by describing a benchmark and supplying a grader, which a benchmark adaptor turns into a benchmark-specific plugin. A solver attempts the tasks, and an evolver reads the solver's training rollouts and rewrites the skills at each epoch. We implement the evolver as one stateful coding agent with a bash tool for the whole run, unlike the two most common designs among the skill optimizers we surveyed: a sequence of single model calls, as in SkillOpt, and an agent that starts afresh each round, as in SkillBoost. We also measure how much the solver's rollouts vary from one attempt to the next, give the evolver several attempts per training task, and accept every epoch's skills instead of selecting among them on noisy scores. With fine-grained ground truth, EvoSkill v2's skills raised the held-out pass rate above the stronger of SkillOpt and SkillBoost by 7.7 percentage points on SpreadsheetBench v2 and by 5.4 on $\tau^3$-bench banking. In both benchmarks and both feedback settings, EvoSkill v2's evolver cost less per percentage point of held-out pass rate gained than either baseline's optimizer. EvoSkill v2 also enables a less capable model to improve its own skills, demonstrated by a gain of up to 21.5 percentage points in the $\tau^3$-bench banking experiment where Claude Haiku 5.5 powers both solver and evolver. ## Claude Desktop v2.0: You Don't Need VS Code Anymore! DevFeed: [Claude Desktop v2.0: You Don't Need VS Code Anymore!]() Original publisher: [Read original article]() Author: Paweł Huryn Published: 2026-10-20T15:12:41Z Content type: tutorial Language: en Sources: [The Product Compass]() Topics: [Claude Code](), [Claude](), [AI-assisted coding](), [Visual Studio Code](), [generative-ai-tools](), [Agent Browser](), [Multi Agent Systems](), [anthropic](), [Pull Request](), [codex]() Tags: [agentic-systems](), [agents](), [agents-md](), [ai](), [analyze](), [apps](), [artifacts](), [automate](), [claude](), [code](), [collections](), [courses](), [desktop](), [features](), [mcp](), [mcp-servers](), [servers](), [skills](), [v2](), [vs-code]() ### AI overview The article recommends Claude Desktop's Code tab as a lower-friction way for Product Managers to prototype and build with Claude Code without VS Code. It describes local project folders with AGENTS.md instructions, connector and permission setup, reusable skills, and a built-in browser for running apps and reviewing changes. The author says Cowork lacks app previews, project-specific skills and connections, guardrails, and AGENTS.md support, and reports that its 10 GB virtual machine limit became restrictive in their tests. Other described features include artifacts, scheduled tasks, remote control, cloud sessions, and session compaction. Code access requires a paid Pro, Max, Team, or Enterprise subscription; the article's model and usage recommendations reflect the author's experience. ### Source excerpt How to start, the features that matter for PMs, and why I now recommend the Code tab. Plus my collections of skills, MCP servers, free courses, and AGENTS.md snippets. ## Benchmarking DuckDB v2.0-dev on NVIDIA Vera DevFeed: [Benchmarking DuckDB v2.0-dev on NVIDIA Vera]() Original publisher: [Read original article]() Author: Gábor Szárnyas Published: 2026-10-16T00:00:00Z Content type: news Language: en Sources: [DuckDB]() Topics: [NVIDIA Vera](), [DuckDB](), [Nvidia]() Tags: [benchmark](), [benchmarking](), [dev](), [duckdb](), [nvidia](), [nvidia-vera](), [v2](), [vera]() ### AI overview DuckDB preview build 1.6.0.dev379 scored 46% higher than an Intel Xeon 6 baseline on the full TPC-H SF1,000 benchmark when run on NVIDIA Vera. The comparison used Ubuntu 26.04, 768 GB RAM, one NVMe SSD, and three measured runs after a discarded warmup; DuckDB was limited to 96 threads on the 96-core Xeon and 88 threads on Vera, with Vera restricted to one socket. The authors attribute Vera's advantage to memory bandwidth and latency that speed scans and keep cores occupied, plus large Olympus-core caches that help with join and aggregation hash tables. The result is from a preview build and single-socket Vera run; the authors say they plan to repeat with DuckDB 2.0.0 stable and the full dual-socket system. ### Source excerpt We put NVIDIA's new Vera CPU through its paces using the full TPC-H workload on scale factor 1,000. ## October 2026 LangChain Newsletter: Everything we shipped at Interrupt NYC & London DevFeed: [October 2026 LangChain Newsletter: Everything we shipped at Interrupt NYC & London]() Original publisher: [Read original article]() Author: LangChain Accounts Published: 2026-10-15T16:50:48Z Content type: article Language: en Sources: [LangChain Blog]() Topics: [LangChain](), [AI Context Management](), [Multi Agent Systems](), [Artificial Intelligence](), [Enhancing AI model performance]() Tags: [agent-evals](), [agent-memory](), [agents](), [ai-search](), [custom-apps](), [fine-tuning](), [langchain](), [langsmith](), [v2]() ### AI overview The October 2026 LangChain newsletter highlights product releases from Interrupt NYC and London, including LangSmith AI Search, Managed Deep Agents v0.9, LangSmith Engine v2, trajectories, custom apps, and fine-tuning. It also covers LangChain v1.4.0, agent building with LangGraph, workshops, events, and a customer example involving Zip's chatbot and agent. ### Source excerpt What we shipped at Interrupt NYC and London: LangSmith AI Search, Engine v2, Trajectories, Custom Apps, Fine-Tuning, and Managed Deep Agents v0.9. ## How to send Webflow form leads to Pipedrive as contacts with Webflow Cloud DevFeed: [How to send Webflow form leads to Pipedrive as contacts with Webflow Cloud]() Original publisher: [Read original article]() Author: Ismail Ajagbe Published: 2026-10-09T00:00:00Z Content type: tutorial Language: en Sources: [Webflow Blog]() Topics: [webflow](), [hosting](), [Learning]() Tags: [2](), [2026](), [22](), [3](), [account](), [acme](), [adapter](), [added](), [api](), [build](), [cloud](), [form](), [guides](), [handler](), [how-to](), [route](), [send](), [submission](), [v2](), [webflow](), [with]() ### AI overview A Next.js Route Handler on Webflow Cloud can receive a Webflow form webhook and create a Pipedrive person through the v2 API, keeping the Pipedrive token in server-side environment variables. The guide maps case-sensitive form fields to the person's name, email and optional phone, and uses a shared secret in the webhook URL to authenticate submissions. It covers setup, deployment and curl verification; the mount path must be included in the live URL. Responses distinguish bad keys (401), missing Pipedrive configuration (500), missing email (400) and API or network failures (502). The handler creates a new person on every call, so repeated submissions can create duplicates; Webflow retains submissions for recovery if Pipedrive is unavailable. The URL secret is exposed in the query string and may appear in logs, and debug payload logging should be removed after checking the incoming field structure. ### Source excerpt Learn how to build a Webflow Cloud Route Handler that turns each Webflow form submission into a Pipedrive person through the v2 API. ## Agent Mode in the DuckDB CLI DevFeed: [Agent Mode in the DuckDB CLI]() Original publisher: [Read original article]() Author: The DuckDB team Published: 2026-10-09T00:00:00Z Content type: article Language: en Sources: [DuckDB]() Topics: [DuckDB](), [console](), [Claude Code](), [GitHub Copilot]() Tags: [agent-mode](), [agents](), [ai](), [ai-coding](), [ai-coding-agents](), [claude-code](), [cli](), [codex](), [copilot](), [cost](), [csv](), [duckdb](), [errors](), [json](), [markdown](), [queries](), [tables](), [tokens](), [using-duckdb](), [v2]() ### AI overview DuckDB v2.0 agent mode adapts CLI output for coding agents: it uses compact Markdown tables, marks omitted rows and cell text, emits JSON errors, and can stop oversized results early while reporting counts and hashes. It activates when an agent environment variable is set, output is piped, and no format flag is specified; `-agent`, `-no-agent`, and `DUCKDB_AGENT_MODE` can control it. Explicit CLI settings and `.duckdbrc` still take precedence. In a Claude Code test of 22 TPC-H questions, all 132 runs were correct and agent mode reduced output tokens from 123.6k to 50.8k (59%), but total input savings were about 0.5%, runtime did not improve, and agent-mode runs used slightly more turns. The feature is an initial version, and agent detection conventions remain in progress. ### Source excerpt The DuckDB v2.0 CLI has an agent mode that gets AI coding agents to a correct result faster and with fewer tokens. When the CLI detects an agent, it prints compact Markdown tables instead of padded boxes, says clearly when a result was cut, stops runaway queries early, and reports errors as JSON. Long queries announce their expected cost before they start, so the agent can decide whether to wait. ## New Spectre v2 Variant Exposes Intel, AMD, Arm CPUs to Data Leaks DevFeed: [New Spectre v2 Variant Exposes Intel, AMD, Arm CPUs to Data Leaks]() Original publisher: [Read original article]() Author: Eduard Kovacs Published: 2026-09-29T17:00:00Z Content type: news Language: en Sources: [SecurityWeek]() Topics: [Security](), [Kernel](), [Linux Kernel](), [B4X]() Tags: [amd](), [arm](), [branch](), [browsers](), [compilers](), [cpu](), [cpu-attack](), [cpu-vulnerability](), [data-leak](), [data-leaks](), [endpoint-security](), [firefox](), [intel](), [javascript](), [jit](), [kernel](), [linux](), [linux-kernel](), [spectre](), [v2](), [vulnerabilities](), [web-browsers]() ### AI overview Researchers disclosed Branch Target Reuse (BTR), a Spectre v2 variant affecting Intel, AMD, and Arm CPUs. It targets just-in-time compilers used by kernels, browsers, and runtimes, and can enable attackers who run code on a machine to leak sensitive memory. The researchers demonstrated Linux kernel exploits that could leak a root password hash; a complete browser exploit has not yet been built. ### Source excerpt Branch Target Reuse (BTR) is a new Spectre v2 attack targeting JIT compilers in web browsers, language runtimes, and the operating system kernel The post New Spectre v2 Variant Exposes Intel, AMD, Arm CPUs to Data Leaks appeared first on SecurityWeek. ## Git v2.56.0 released DevFeed: [Git v2.56.0 released]() Original publisher: [Read original article]() Author: jake Published: 2026-09-28T17:33:58Z Content type: release Language: en Sources: [LWN.net]() Topics: [Git](), [version-control]() Tags: [git](), [github](), [merge](), [v2](), [version-control]() ### AI overview Git 2.56 has been released with 748 non-merge commits from 104 developers, including 39 first-time contributors. Changes include a safer conflict-resolution workflow, smaller path-walk repacks, and a new git history drop subcommand. ### Source excerpt Version 2.56 of the Git distributed version-control system has been released. It has 748 non-merge commits since Git 2.55 was released back in June; those commits came from 104 developers, 39 of whom are first-time contributors. New features include a safer workflow for conflict resolution, smaller path-walk repacks, a new git history drop sub-command, and much more. LWN looked at Git 2.56 recently and the GitHub blog has a lengthy look at 2.56 as well. ## Apache Iceberg Geospatial Types: Geometry and Geography in the Lakehouse DevFeed: [Apache Iceberg Geospatial Types: Geometry and Geography in the Lakehouse]() Original publisher: [Read original article]() Author: Alex Merced Published: 2026-09-25T12:43:09Z Content type: article Language: en Sources: [Dremio]() Topics: [Apache Iceberg](), [geospatial](), [parquet](), [data](), [systems]() Tags: [apache-iceberg](), [compatibility](), [data-lakehouse](), [dremio-blog-open-data-insights](), [file-formats](), [geospatial](), [geospatial-data](), [lakehouse](), [migrations](), [parquet](), [v2](), [v3]() ### AI overview This article explains how Apache Iceberg v3 represents geometry and geography as primitive types, including Well Known Binary storage in Avro and Parquet, optional coordinate reference system metadata, and spatial bounds for pruning. It also covers engine compatibility, CRS and axis-order testing, and migration risks from v2 to v3. ### Source excerpt Learn how Iceberg v3 stores geometry and geography, handles CRS metadata, writes Parquet values, and preserves spatial statistics across engines. The post Apache Iceberg Geospatial Types: Geometry and Geography in the Lakehouse appeared first on Dremio. ## AWS CRT for JavaScript and AWS IoT Device SDK for JavaScript v2 align with the Node.js release schedule DevFeed: [AWS CRT for JavaScript and AWS IoT Device SDK for JavaScript v2 align with the Node.js release schedule]() Original publisher: [Read original article]() Author: Rakshilkumar Modi Published: 2026-09-23T17:13:23Z Content type: release Language: en Sources: [AWS Developer Tools Blog]() Topics: [Node.js](), [JavaScript](), [Internet of things](), [releases](), [SDK]() Tags: [announcements](), [aws](), [aws-iot](), [aws-sdk](), [deprecation](), [developer-tools](), [end-of-life](), [iot](), [javascript](), [lts](), [node-js](), [open-source](), [release](), [release-notes](), [release-schedule](), [releases](), [sdk](), [upgrade](), [v2](), [version]() ### AI overview AWS CRT for JavaScript and AWS IoT Device SDK for JavaScript v2 will align their minimum supported Node.js versions with the Node.js release schedule. Beginning in January 2027, end-of-life Node.js versions will be dropped after an eight-month maintenance window, with changes documented in release notes. ### Source excerpt The AWS CRT for JavaScript is adopting a policy that aligns with the official Node.js release schedule, providing a predictable timeline for when support for older Node.js versions is dropped. The AWS IoT Device SDK for JavaScript v2 is built on top of AWS CRT for JavaScript, so it shares the same minimum Node.js version. [...] ## DuckDB Now Ships inside dbt v2 DevFeed: [DuckDB Now Ships inside dbt v2]() Original publisher: [Read original article]() Author: Geertjan Wielenga Published: 2026-09-22T00:00:00Z Content type: release Language: en Sources: [DuckDB]() Topics: [DuckDB](), [parquet]() Tags: [adapter](), [apache-2-0](), [autocomplete](), [catalog](), [ci](), [cli](), [columnar](), [dbt](), [duckdb](), [ducklake](), [fusion](), [iceberg](), [metadata](), [parquet](), [rust](), [using-duckdb](), [v2]() ### AI overview dbt v2 includes a built-in DuckDB adapter, so a local project needs no separate adapter package; dbt downloads and caches its DuckDB driver on first use. The excerpt describes new catalog support for DuckLake and Iceberg REST catalogs through `catalogs.yml`, which requires the v2 engine and `use_catalogs_v2`; the legacy Python adapter lacks this feature. dbt v2 can also write project metadata as Parquet, queryable with DuckDB, and produce column lineage through strict static analysis. It ships a pinned DuckDB build that supports read-write Iceberg REST catalogs. For migration, the excerpt recommends trying the opt-in v2 parser in dbt v1.12 with `dbt parse --use-v2-parser`, then following the upgrade guide; `dbt-autofix` handles many required changes. Python versions of dbt Core remain available for users who prefer not to move. ### Source excerpt dbt v2, which runs on the new Rust-based Fusion engine, is the first dbt release that ships with a built-in DuckDB adapter. This post covers setup, DuckLake and Iceberg catalogs, querying dbt's Parquet metadata with DuckDB, plus other v2 features that matter to DuckDB users, including migrating to dbt v2. [Next page]()