☕️ Microsoft releases VS Code 1.129

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

💻 Microsoft releases VS Code 1.129

🐧 Torvalds tells AI critics to fork Linux

🧠 Murati's startup launches first AI model

🔧 Atlassian rebuilds Jira for AI agents

⚡ Apache Spark 4.2 is out

Plus: 🎁 11 other news you might like, 🧰 4 tools, and 📚 5 papers.

💻 Microsoft releases VS Code 1.129 LINK

Microsoft shipped Visual Studio Code 1.129, a weekly release notable for previewing a modernized workbench UI and adding an experimental editor panel to the Agents window.

Enabling the "Workbench > Experimental: Modern UI" setting renders sidebars and the bottom panel as floating cards with rounded corners and gaps, applying refreshed styles that match the new Agents window.

The new editor panel merges the editor and detail area into one docked pane with a shared tab bar, and chat messages prefixed with an exclamation mark now run as terminal commands in agent host sessions.

🐧 Torvalds tells AI critics to fork Linux LINK

Linus Torvalds declared that Linux is not an "anti-AI" project and told developers who object to AI-assisted contributions that they can fork the kernel or simply walk away.

Responding to anti-LLM sentiment on lore.kernel.org, Torvalds said he'll "put my foot down" as top-level maintainer, calling AI a useful tool while conceding it strains maintainer workloads and "keeps finding embarrassing bugs."

The stance marks a shift from his 2024 view that AI was "90% marketing," coming after he began using AI on a personal project in January and criticized AI-generated fixes for introducing kernel bloat.

🧠 Murati's startup launches first AI model LINK

Thinking Machines Lab, the startup founded by ex-OpenAI CTO Mira Murati, released Inkling, its first open-weights foundation model, a 975-billion-parameter Mixture-of-Experts system built for multimodal reasoning, coding, tool use, and long-context work.

An adjustable effort parameter between 0.2 and 0.99 lets developers trade reasoning tokens for cost, matching Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly a third as many tokens, alongside strong audio and safety scores.

Full weights ship through Hugging Face, including an NVFP4 checkpoint for Nvidia Blackwell, with fine-tuning via Tinker, deployment across Together AI, Fireworks, Modal, Databricks, and Baseten, plus SGLang, vLLM, and Llama.cpp support.

🔧 Atlassian rebuilds Jira for AI agents LINK

Atlassian rebuilt Jira around AI coding agents, adding tools to turn rough ideas into specs, convert work items into agent requests, and assign tasks automatically as tickets move through workflows.

The new Agentic Engineering Template sets up boards where columns carry state, so moving an issue between columns automatically assigns an agent, targeting advanced teams wanting Jira workflows running alongside agents.

Jira Coding Agent runs bounded work in the cloud without a coding editor, and any work item can be delegated to Claude, Codex, Cursor, or GitHub Copilot while staying grounded in Jira's project history.

⚡ Apache Spark 4.2 is out LINK

Apache Spark 4.2 has shipped, pulling more of the modern data and AI stack into the engine with governed metric views, native vector search primitives, first-class change data capture, and an Arrow-first Python path.

Metric views add a semantic layer to Spark SQL for defining non-additive measures like ratios and distinct counts once, while new vector distance, NEAREST BY top-K joins, sketches, and geospatial types enable retrieval and location-aware analytics directly in SQL.

Arrow-optimized Python UDF execution is now default so existing UDFs get the columnar path without rewrites, Pandas 3 is supported, and Real-Time Mode extends to stateless PySpark streaming queries for millisecond latency.

Other news you might like

• Running a self-hosted LLM in Kubernetes with vLLMLINK

• How to build a cloud software factory - self-improving code reviewLINK

• Lessons Learned Rewriting a Sticky DetectorLINK

• Context engineering with Dex HorthyLINK

• I use a local LLM to scrub my bank statements before they ever reach ChatGPTLINK

• Zig creator calls Bun's Claude Rust rewrite 'unreviewed slop'LINK

• Ex-Sega engineer creates 'super realistic' CRT monitor emulator, incredible retro offering even includes TV screen tapping to fix pictureLINK

• What building Shippy taught us about building agentsLINK

• Sheetz moves 838 stores off VMware: Broadcom created “too much uncertainty"LINK

• Is a Pod the right deployment unit for an AI agent?LINK

🧰 Trending tools

Humalike: adds social skills and proactiveness to AI agents through behavioral APIs, models, and benchmarks, making them feel more humanlike in conversation.LINK

Tiptap AI Toolkit: a set of building blocks for creating AI agents that edit documents, offering real-time streaming, track-changes review, and context-aware selections.LINK

Firefox in WebAssembly: runs the entire Firefox browser (Gecko, UI components, and SpiderMonkey JS engine) compiled to WebAssembly and rendered to a canvas element in-browser.LINK

vscodium-rust: an AI-native code editor offering agentic workflows, iPhone emulation on Windows/Linux, a PyTorch ML studio, and ROCm-optimized local AI tools.LINK

📚 Trending papers & reports

Solving huge scheduling puzzles gets faster and more general purpose when a flexible planning method replaces custom-built solvers, beating existing automated tools across four problem types.LINK

Self-driving trajectory planning gets a built-in safety check during route generation, fixing collision risks at every planning step without distorting the AI's learned driving path.LINK

Old Manchu handwriting gets read almost as accurately by a smart routing system that picks the right specialist for each writing style as by hand-picking the ideal expert yourself, hitting 99.3 percent page-matching accuracy and error rates as low as 0.30 percent.LINK

AI judges debating moral dilemmas flip their verdicts far more depending on how they take turns, with revision rates swinging from just 0.6-3.1% to 28-41% across different chatbot models.LINK

Cleaning up noisy audio before transcribing it can backfire, making speech recognition errors worse even though the sound quality itself improves, with Whisper base's word error rate jumping from 10.53% to 21.66% on English speech.LINK

See you tomorrow for a new dose of ☕️ Devshot!