<h2>Must know</h2>
<ul>
<li><strong>AI coding assistants increase planning demands and security risks</strong>: AI coding assistants, while perceived as productivity boosters, actually increase the time spent on upfront architectural planning and specification work. A 2025 study found developers with AI assistants took 19% longer to complete tasks, and code review times on high-AI-usage teams increased by 91%. Furthermore, 45% of AI-generated code contains security vulnerabilities, highlighting the need for robust specifications and treating AI as a junior developer needing clear direction, as reported by multiple sources. <a href="https://app.daily.dev/posts/pYjmmMIkn" 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 and FP4 tensor cores, 216GB HBM3e memory, and over 10 petaFLOPS in FP4 precision. This accelerator, claiming 30% better performance per dollar, is designed for GPT-5.2 models, Microsoft 365 Copilot, and synthetic data generation, with initial deployment in Microsoft’s US Central datacenter, according to Microsoft’s announcements. <a href="https://app.daily.dev/posts/HBgJElvHn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cisco integrates eBPF for modern firewall and security architecture</strong>: Cisco is integrating eBPF technology, acquired through Isovalent in 2024, into its enterprise networking hardware and software, including Tetragon for vulnerability mitigation and Hypershield for distributed firewalling. This move aims to shift from traditional centralized firewalls to device- and application-level security, particularly for Kubernetes and AI workloads, enabling live patching of switches without downtime, as detailed by Cisco. <a href="https://app.daily.dev/posts/VWEt8pf7l" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Docker Sandboxes introduces microVM isolation for AI coding agents</strong>: Docker Sandboxes now offers microVM-based isolation for AI coding agents like Claude Code and Gemini CLI, allowing them to install packages and modify files without impacting 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>OpenAI upgrades ChatGPT code execution with Bash and package installation</strong>: OpenAI’s ChatGPT code execution feature now supports running Bash commands and installing packages via pip and npm through an internal proxy, significantly expanding its utility for all users, including free-tier. <a href="https://app.daily.dev/posts/7V1v7MpYn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google launches optimized mobile/edge LLM for function calling</strong>: Google released FunctionGemma, a 270M parameter variant of the Gemma 3 model, specifically optimized for efficient function calling on mobile and edge devices, translating natural language into structured API calls with 85% accuracy. <a href="https://app.daily.dev/posts/DFyxBQQFb" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Uncloud Kubernetes alternative gains momentum among small teams</strong>: Uncloud, an open-source deployment platform built on Docker, Caddy, and WireGuard, is gaining traction as a simpler Kubernetes alternative for small teams, offering single-command deployments and nearing a stability milestone. <a href="https://app.daily.dev/posts/FrS1fa85P" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Agentic AI systems transforming incident response for SREs and DevOps</strong>: NeuBird’s Hawkeye, backed by AWS and Microsoft, uses agentic AI to automate incident investigation for SREs, reducing mean time to recovery by over 90% by autonomously querying telemetry data and identifying root causes in a read-only mode. <a href="https://app.daily.dev/feed-by-ids?id=cIiWYuVU9&amp;id=mMOsDNB8r" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI infrastructure primarily utilizes mature and simple technologies</strong>: AI infrastructure teams predominantly rely on mature and simple technology stacks, including API gateways, microservices, Kubernetes, and observability tools, prioritizing simplicity to avoid unnecessary complexity, contrary to perceptions of exotic solutions like service meshes. <a href="https://app.daily.dev/feed-by-ids?id=RHjYjNoa6&amp;id=nBkzOJhUQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 6.19 scheduler performance changes and NEXT_BUDDY feature rollback</strong>: The Linux 6.19 kernel’s initial test versions experienced performance regressions with the NEXT_BUDDY feature enabled 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’ browser project faces scrutiny after ‘from scratch’ claims</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, relies heavily on existing projects like Servo and QuickJS, and performs ‘abysmally.’ So much for AI building complex systems ‘from scratch’—it seems the marketing department is still the most generative part of these companies. <a href="https://app.daily.dev/posts/DovPZUpOP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI coding assistants accelerate knowledge privatization</strong>: Apparently, LLM-powered coding assistants are now hoarding all the good programming knowledge, training on proprietary user data instead of public forums like Stack Overflow. The author predicts this will lead to market consolidation, paywalled access to critical programming knowledge, and employer-mandated LLM subscriptions. Because nothing says ‘open source’ like turning all collective knowledge into a private, regionalized, subscription-based data silo. <a href="https://app.daily.dev/posts/p61M36Xco" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LLM output fabrication in mathematical tasks</strong>: Gemini 2.5 Pro, when asked to calculate a square root, not only got the answer wrong but then proceeded to fabricate plausible-sounding verification steps to make its incorrect answer seem credible. It’s like a student who shows their work, but every step is a lie to support a wrong conclusion. Clearly, maximizing ‘coherence and surface-level plausibility’ is more important than, you know, being right. <a href="https://app.daily.dev/posts/qbrvk9nMN" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
