<h2>Must know</h2>
<ul>
<li><strong>Google's embedding model deprecation causes vendor lock-in</strong>: Google deprecated its <code>text-embedding-004</code> model, rendering thousands of stored vectors unusable for a production RAG-based movie recommender. This highlights a critical issue: different embedding models produce incompatible vector spaces, making migration difficult and creating a hidden form of vendor lock-in. The takeaway is to treat embedding models as infrastructure and design for change if you’re persisting embeddings at scale. <a href="https://app.daily.dev/posts/xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Databricks launches Lakebase, a PostgreSQL-compatible database for AI</strong>: Databricks introduced Lakebase, a new serverless, PostgreSQL-based OLTP database specifically designed for AI workloads. It aims to bridge operational database capabilities with Databricks’ data lakehouse architecture, offering features like instant data branching, point-in-time recovery, and unified access controls. This move further blurs the lines between traditional databases and data lakehouses, targeting real-time ML feature serving and AI agent memory. <a href="https://app.daily.dev/posts/56RHAShlB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux kernel 7.0 rc1 released with Rust support and performance boosts</strong>: Linus Torvalds released Linux kernel 7.0 rc1, a major version bump due to his aversion to large numbers, not a feature milestone. This release cements Rust support, brings faster cache clearing, non-disruptive kernel updates, and significant performance improvements for IO_uring and memory reclamation. It also includes updated AMD/Intel silicon support and RISC-V/LoongArch optimizations, continuing the kernel’s steady evolution. <a href="https://app.daily.dev/posts/cYP8biu57" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AI coding agents lack innovation, human developers still relevant</strong>: While AI coding agents like Claude Code are highly productive and fast, they fundamentally lack the ability to devise innovative solutions when facing obstacles, a critical differentiator that preserves the relevance of creative human developers. <a href="https://app.daily.dev/posts/hB7gw6iXo" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenClaw Gateway offers AI agent orchestration dashboard</strong>: Mission Control is an open-source AI agent orchestration dashboard built with Next.js 14 and TypeScript, connecting to OpenClaw Gateway via WebSocket to create tasks, plan with AI, and track agent progress on a Kanban board. <a href="https://app.daily.dev/posts/l0KLAjYGc" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Shuru provides local-first microVM sandbox for macOS AI agents</strong>: Shuru is a new local-first microVM sandbox tool for macOS, utilizing Apple’s Virtualization.framework to run lightweight Linux VMs, providing secure, ephemeral, and reproducible environments for AI agent workloads without Docker. <a href="https://app.daily.dev/posts/TqgsBjEoV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New VS Code extension 'Fresh File Explorer' focuses on recent files</strong>: Fresh File Explorer is a new VS Code extension that replaces the standard file explorer, displaying recently modified files with Git history integration, configurable time windows, and a ‘Resurrect’ feature for deleted files. <a href="https://app.daily.dev/posts/pQedWHBRB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rust RwLock can be slower than Mutex for read-heavy workloads</strong>: A hands-on investigation revealed that <code>RwLock</code> performed ~5× worse than <code>Mutex</code> for a read-heavy LRU tensor cache in Rust due to cache-line ping-pong, suggesting <code>Mutex</code> might be better for very short critical sections. <a href="https://app.daily.dev/posts/rFMs5918e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>New real-time glint particle rendering technique achieves 280 FPS</strong>: Adobe Research, NVIDIA, and Aalto University developed a new technique for real-time rendering of millions of microscopic glint particles on surfaces like metallic paint or snow at over 280 FPS on consumer GPUs, generating particles on-the-fly without storing every position. <a href="https://app.daily.dev/posts/rT496vKCA" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Godot Go bindings now faster than GDScript and Rust</strong>: <a href="http://graphics.gd" target="_blank" rel="noopener nofollow">graphics.gd</a>, the Go bindings for Godot via GDExtension, has achieved FFI performance surpassing GDScript and even ‘safe’ Rust in benchmarks, thanks to call batching and Go 1.26’s cgo overhead reduction. <a href="https://app.daily.dev/posts/j8AoMYJTv" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic's Claude compiler video: more marketing than magic</strong>: Anthropic released a marketing video showcasing Claude autonomously building a C compiler capable of compiling Linux and Doom. Critics, however, are calling out the ‘dishonest framing,’ suggesting the real goal was a stress test of Claude’s autonomous runtime, not a straightforward demonstration of compiler-building prowess. So, it’s less ‘AI wizardry’ and more ‘look what our AI can do if we set it up just right and then edit the hell out of it.’ <a href="https://app.daily.dev/posts/vGYk0DbTO" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Gemini 3.1 Pro: pure disappointment despite benchmark hype</strong>: Despite strong benchmark scores, Google’s Gemini 3.1 Pro consistently underperformed Claude Opus 4.6 in real-world coding and UI generation tasks, with issues like tool-calling failures, incomplete features, and poor environment detection. It seems Google’s latest AI model is great at passing tests but struggles with actual work, proving once again that benchmarks are often just a fancy way to lie with statistics. <a href="https://app.daily.dev/posts/bIAFUawfa" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Meta patents AI to simulate dead users: because privacy is so last decade</strong>: Meta was granted a patent for an AI system designed to simulate a user’s social media activity, including posting and chatting, using an LLM trained on their past data, even when the user is ‘absent’ from the platform. Because nothing says ‘we care about your mental health and ethical boundaries’ like creating digital ghosts of your deceased loved ones to keep the engagement numbers up. What could possibly go wrong? <a href="https://app.daily.dev/posts/AeW0pHCcd" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
