Le meilleur du dev et de l'IA, scoré chaque jour par un agent. Filtré, résumé, classé. Aucune couleur, aucun bruit — juste la matière.
The article argues that LLMs should not be treated as junior engineers. They are powerful tools that require careful integration, supervision, and governance within real development workflows. It emphasizes that while LLMs can assist coding tasks, they do not replace skilled engineering judgment and design.
Linux kernel vulnerabilities sometimes reach distributions too late, leaving deploys unpatched. The discussion highlights gaps between kernel disclosures and downstream advisories, and calls for improved visibility for maintainers. For a senior dev, this suggests reworking patching workflows and adding kernel alert feeds to CI/CD.
Techniques for better software testing surveys a range of approaches to improve test quality and reliability. It discusses strategies for designing tests, selecting appropriate testing methodologies, and integrating testing into the development workflow. The article serves as a practical primer for shaping more robust testing practices in modern software projects.
Researchers uncovered a malicious dependency in the PyTorch Lightning AI training library, tied to a Shai-Hulud themed malware. The post explains how the compromised package could be exploited within AI training workflows and the potential security risks.
The GCC 16 Release Series outlines changes, new features, and fixes across its compiler toolchain. The notes cover improvements to optimization, language standards support, and bug fixes, with version-by-version details. Developers should review the changes to understand potential impacts on builds and flags.
Google is rolling out the Gemini AI assistant to cars with Google built-in, replacing Google Assistant with a more natural conversational experience and better access to vehicle information and settings. The upgrade will reach both new cars and existing ones via a software update, continuing Google's promise that your car will get better over time. This marks a notable shift in in-car AI capabilities and OTA updates for automotive systems.
OpenAI is addressing its 'goblin problem' after reports that its coding model was instructed to never mention goblins or similar creatures. The company says these references are a 'strange habit' from training, appearing with GPT-5.1 and the Nerdy personality, and they're taking steps to curb them.
CSS Zen Garden demonstrates how a single HTML page can be dramatically re-styled with just CSS, underscoring the power of CSS as a design tool. It serves as a reference for frontend theming and CSS techniques, showing how typography, layout, and visuals can change without altering markup. A senior dev can draw patterns for design systems and inspiration for CSS architecture.
Amber-Lang 0.6.0 adds multi-shell support (Bash, Zsh, Ksh, and Bash 3.2), along with recursion, union types, and public variables, plus a built-in testing suite with assert. It also delivers performance gains via native Bash arithmetic, new builtins (fetch, touch, rm, ls), LGPL licensing, and breaking changes like mandatory parentheses for builtins and stricter out-of-bounds array checks.
Servo makes progress with keyboard navigation improvements, better debugging tooling, and FreeBSD support, plus additional updates. The post highlights enhancements aimed at developers, improving ergonomics and tooling. This marks ongoing open-source collaboration and maintenance for the engine.
Streaming UIs look simple at first glance but are technically hard. Designers must handle layout shifts, motion preferences, correct markup, and multiple UI states, while also planning for interruptions and ensuring keyboard navigation works as the UI shifts. ARIA attributes and accessibility considerations must be addressed.
Grok 4.3 is now available on Vercel AI Gateway, with a December 2025 knowledge cutoff and a 1M token context window. The model improves accuracy, tool calling, and instruction following. Use xai/grok-4.3 via the AI Gateway/SDK, which provides a unified API, usage/cost tracking, automatic retries, and features like custom reporting, observability, and Bring Your Own Key.
Futhark breaks a decade of C API compatibility, introducing breaking changes to its C bindings. The post explains the motivation, outlines the migration path, and discusses the impact on existing projects relying on the old API.
Day 4 focuses on turning subjective agent vibes into objective, data-driven evaluation using LLM Evals. It introduces Evalite, an open-source eval framework built on Vitest to run tests without cloud providers, and guides you to define success criteria and write a first scorer to aim for a 100 score. By end, you'll have a foundational evaluation setup for the DeepSearch agent.
This workshop examines how much control the LLM should have over the app's control flow, pitting agents against workflows. It notes that more agentic systems offer flexibility, while deterministic workflows improve predictability. In upcoming exercises, the dial will be turned back toward determinism to test concrete patterns.
Using Vercel AI SDK's useChat, you can prototype a chat quickly, but turning it into a durable, persistent product requires backend persistence, real-time updates, and solid routing. The workshop guides you through building a POST /api/chat endpoint that streams OpenAI completions, using Drizzle to create and return a chatId, saving messages as they stream, rehydrating the UI with past conversations, and keeping the sidebar in sync via route revalidation.
Observability is essential when building AI-driven products, enabling you to see, debug, and improve agent behavior as conversations and calls grow. Day 3 guides you through setting up Langfuse, exploring traces with custom metadata, and building a time-aware web scraper tool to turn the DeepSearch agent into a well-understood system.
Showing Sources In The Frontend outlines an evaluator-optimizer loop for LLM workflows and recommends displaying sources in the UI as card-like entries (favicon, title and snippet). It reframes the agent as a decision loop—answer when you have enough information, otherwise search—and adapts frontend patterns to present sources clearly.
Observability is essential when building AI-driven products. Day 3 teaches you to see and debug agent behavior using Langfuse, explore traces, and add custom trace metadata, then to build a time-aware web scraper tool to improve your DeepSearch agent.
That workshop reframes the agent as an evaluator-optimizer loop: one LLM call generates a response while another provides evaluation and feedback. It also covers showing sources in the frontend with a card-like UI (favicon, title, snippet) to surface provenance and improve traceability.