<h2>Must know</h2>
<ul>
<li><strong>AI coding assistants: more planning, more bugs, more cost</strong>: AI coding assistants, while perceived as productivity boosters, actually demand significantly more upfront architectural planning and specification work. A 2025 study revealed that teams using AI assistants took 19% longer to complete tasks, with code review times increasing by 91% and 45% of AI-generated code containing security vulnerabilities. This suggests that treating AI as a junior developer needing clear direction, rather than an autonomous agent, is crucial to avoid the ‘zig-zag problem’ and the hidden ‘architecture tax’ that can negate perceived gains. <a href="https://app.daily.dev/posts/pYjmmMIkn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>TSMC’s flat CapEx creates AI chip bottleneck for hyperscalers</strong>: Despite the AI boom ignited by ChatGPT, TSMC maintained flat capital expenditures from 2021 to 2024, leading to a significant supply bottleneck for AI chips. This disparity is costing hyperscalers billions in lost AI revenue, and the shortage is projected to persist through at least 2028. The analysis suggests hyperscalers should diversify their supply chains, as relying solely on TSMC is riskier than exploring alternatives like Samsung or Intel to meet surging demand. <a href="https://app.daily.dev/posts/pPxEIulIE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>OpenAI upgrades ChatGPT code execution with Bash and package installs</strong>: OpenAI has quietly enhanced ChatGPT’s code execution feature, now allowing users to run Bash commands and install packages via pip and npm within its sandboxed environment. This upgrade, which also includes support for multiple programming languages and a container.download tool, significantly expands ChatGPT’s utility for developers. However, these new capabilities remain undocumented in official release notes, raising questions about security and OpenAI’s long-term strategy for managing potential misuse. <a href="https://app.daily.dev/posts/7V1v7MpYn" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Docker Sandboxes isolates AI coding agents with microVMs</strong>: Docker Sandboxes now uses microVM-based isolation for AI coding agents like Claude Code and Gemini CLI, allowing them to install packages and modify files without affecting the host system, currently available on macOS with Windows and Linux support planned. <a href="https://app.daily.dev/posts/6YdOupFK7" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Microsoft launches Maia 200 AI inference accelerator</strong>: Microsoft has unveiled the Maia 200, a custom AI inference accelerator built on TSMC’s 3nm process, featuring FP8/FP4 tensor cores and 216GB HBM3e memory, claiming 30% better performance per dollar for tasks like GPT-5.2 models and Microsoft 365 Copilot. <a href="https://app.daily.dev/posts/HBgJElvHn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google launches FunctionGemma for mobile/edge AI function calling</strong>: Google released FunctionGemma, a 270M parameter model optimized for on-device function calling on mobile and edge hardware, translating natural language into API calls with 85% accuracy and integrating with Hugging Face, Keras, and vLLM. <a href="https://app.daily.dev/posts/DFyxBQQFb" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>KDE Plasma 6.6 shifts to systemd, ends FreeBSD support</strong>: KDE Plasma 6.6 beta introduces a systemd-logind dependent login manager, effectively ending FreeBSD support and signaling an increasing Linux-specific direction for the desktop environment, mirroring GNOME’s growing reliance on Linux-only infrastructure. <a href="https://app.daily.dev/posts/satawafBY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Node.js 25.5.0 released with security and feature updates</strong>: Node.js 25.5.0 includes an upgrade to NSS 3.119 root certificates, adds LIEF for Single Executable Application (SEA) generation, and introduces an ‘ignore’ option for fs.watch, alongside SQLite defensive mode and test_runner improvements. <a href="https://app.daily.dev/posts/BrQRGrBKE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Uncloud gains traction as Kubernetes alternative for small teams</strong>: Uncloud, an open-source deployment platform built on Docker, Caddy, and WireGuard, is gaining momentum among small teams as a simpler Kubernetes alternative, offering single-command deployments and nearing a stability milestone with plans for a managed control center. <a href="https://app.daily.dev/posts/FrS1fa85P" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 6.19 scheduler performance changes and NEXT_BUDDY rollback</strong>: The Linux 6.19 kernel’s initial test versions experienced performance regressions with the NEXT_BUDDY feature for the EEVDF scheduler, leading to its disabling in the rc7 release candidate after community feedback and benchmarks showed drops in MySQL and other workloads. <a href="https://app.daily.dev/feed-by-ids?id=85ivk605B&amp;id=mNRQJiXzS" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Cursor AI agents build a browser ‘from scratch’ (spoiler: they didn’t)</strong>: Cursor’s CEO proudly announced their AI agents built a fully functional web browser ‘from scratch’ in a week, generating over 3 million lines of code. Turns out, independent developers found the browser barely compiles, doesn’t work, and heavily relies on existing projects like Servo and QuickJS. So, ‘from scratch’ apparently means ‘cobbled together with existing parts and still broken,’ while costing millions in tokens. Impressive marketing, less impressive reality. <a href="https://app.daily.dev/posts/DovPZUpOP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI will make your code unmaintainable, but it’s faster!</strong>: After two years of using AI coding agents, one developer discovered that while the AI generates seemingly high-quality code that passes reviews, it leads to ‘cumulative architectural degradation’ and an unmaintainable codebase. The conclusion? Manual coding is ultimately faster and more productive when you consider the full development lifecycle. So, if you’re wondering why your AI-generated project can’t ship, it’s not the AI’s fault, it’s just ‘slop.’ <a href="https://app.daily.dev/posts/MVxk3C8r1" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-generated tests: the illusion of quality</strong>: Apparently, using large language models to generate software tests <em>after</em> the code is written is just a ‘ceremony’ that provides a false sense of security. The real experts say effective tests should fail first to prove they actually cover functionality. So, if your AI is churning out tests that never fail, you’re not getting quality assurance, you’re just getting a pat on the back from a very confident, very wrong algorithm. Who knew testing required actual thought? <a href="https://app.daily.dev/posts/wLM9iDorf" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
