<h2>Must know</h2>
<ul>
<li><strong>AI coding assistants increase planning demands and security risks</strong>: AI coding assistants, while boosting code writing speed, demand significantly more upfront architectural planning and detailed specifications. A 2025 study revealed that teams using AI assistants took 19% longer to complete tasks overall, 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 requiring clear direction, rather than an autonomous agent, is crucial to avoid the ‘zig-zag problem’ and ‘context pollution’ that degrade code quality and increase debugging efforts. <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. This accelerator, claiming 30% better performance per dollar, is designed for GPT-5.2 models, Microsoft 365 Copilot, and synthetic data generation. Initial deployment is in Microsoft’s US Central datacenter, with a Maia SDK preview planned, signaling Microsoft’s aggressive push into AI hardware. <a href="https://app.daily.dev/posts/HBgJElvHn" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>TSMC CapEx strategy creates AI chip supply bottleneck</strong>: TSMC’s flat capital expenditures from 2021-2024, despite a surge in hyperscaler AI investments, have created a significant AI chip supply bottleneck, costing hyperscalers billions in lost revenue. With TSMC’s planned 2026 CapEx still lagging, shortages are projected to persist until at least 2028. Analysts suggest hyperscalers diversify their supply chains to mitigate risks, as competition from Samsung or Intel could incentivize TSMC to expand production more aggressively. <a href="https://app.daily.dev/posts/pPxEIulIE" 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 microVMs for secure isolation of AI coding agents like Claude Code and Gemini CLI, allowing them to install packages and modify files without host system impact, 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 installs</strong>: OpenAI’s ChatGPT code execution feature now supports Bash commands and package installation via pip and npm within its sandboxed environment, expanding its utility for developers across multiple programming languages. <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 model optimized for on-device function calling on mobile and edge hardware, improving natural language-to-mobile-action accuracy from 58% to 85% 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>AMDGPU Linux driver adds HDMI VRR and ALLM gaming support</strong>: AMDGPU Linux kernel driver patches now enable HDMI Variable Rate Refresh (VRR) and Auto Low Latency Mode (ALLM) for gaming, aiming for feature parity with DisplayPort and Windows platforms despite HDMI Forum restrictions. <a href="https://app.daily.dev/posts/bsF1QAXcQ" 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 from Isovalent, into its networking hardware and software to enable device- and application-level security, live patching, and enhanced protection for Kubernetes and AI workloads. <a href="https://app.daily.dev/posts/VWEt8pf7l" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Uncloud Kubernetes alternative gains momentum for small teams</strong>: Uncloud, an open-source deployment platform using 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>KDE Plasma runs on Apple M3 with Fedora Asahi Remix</strong>: KDE Plasma desktop is now functional on Apple M3 chips via Fedora Asahi Remix, marking a significant milestone for Linux compatibility on Apple Silicon devices. <a href="https://app.daily.dev/posts/cxEiHEruW" 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’ in a week, ‘from scratch,’ 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 much for ‘from scratch’—it seems the AI was better at generating marketing hype than actual code, and at a cost of millions in tokens, according to external analysis. <a href="https://app.daily.dev/posts/DovPZUpOP" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI-generated software tests are ‘ineffective’ without human babysitting</strong>: Apparently, using LLMs to generate software tests after the code is written is just a ‘ceremony’ that gives a ‘false sense of security.’ Who knew that blindly trusting an AI to validate your code might lead to problems? The experts now suggest you should actually make the tests fail first to prove they work, or better yet, write the tests yourself and let the AI write the production code. Revolutionary. <a href="https://app.daily.dev/posts/wLM9iDorf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI coding assistants are privatizing knowledge, because sharing is for suckers</strong>: LLM-powered coding assistants are increasingly trained on proprietary user data, leading to a ‘knowledge hoarding’ scenario where critical programming insights are moving from public forums like Stack Overflow into private, paywalled AI ecosystems. The author predicts market consolidation and regionalized pricing for AI subscriptions. Because why have freely accessible knowledge when you can have a few dominant providers selling you back what you collectively created? Open source, meet your new overlords. <a href="https://app.daily.dev/posts/p61M36Xco" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
