<h2>Must know</h2>
<ul>
<li><strong>Flash-MoE brings 400B parameter models to consumer MacBooks</strong>: A new pure C/Metal inference engine called Flash-MoE allows running the massive Qwen3.5-397B model on a 48GB RAM MacBook Pro at over 4 tokens per second. The implementation achieves this by streaming the 209GB model from SSD using parallel pread calls and loading only the four active experts per layer into RAM. Developer danveloper documented 58 experiments, finding that trusting the OS page cache actually outperformed custom Metal LRU caches and prefetching strategies. <a href="https://app.daily.dev/posts/xsGaEEFUf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cursor admits Composer 2 is built on Moonshot AI base</strong>: Following community sleuthing on X, Cursor’s VP of developer education confirmed that their new Composer 2 model is built on Kimi k2.5, an open-source model from Chinese firm Moonshot AI. While Cursor claims three-quarters of the training compute was their own and performance is distinct, the non-disclosure in the original launch was admitted as a mistake. This partnership, facilitated via Fireworks AI, highlights the complex geopolitical landscape of U.S. companies leveraging Chinese frontier models. <a href="https://app.daily.dev/posts/2O2HmWwMM" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Agentic Platform Engineering treats AI skills as versioned infra</strong>: A new architectural pattern is emerging that moves AI agent configuration away from ad-hoc setups and into versioned repositories using a ‘resource-catalog’ approach. By using layered Markdown files scoped to filesystem directories, agents only load relevant context, significantly reducing the hidden token costs of MCP tool schemas which can hit 50,000 tokens for just 90 tools. This methodology, detailed in recent community discussions, enables full environment rebuilds in under five minutes via symlink-based deployment. <a href="https://app.daily.dev/feed-by-ids?id=qiDJTjrKj&amp;id=f6EHrsEI8" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AWS Aurora DSQL adds Kiro AI agent skills for schema design</strong>: The serverless PostgreSQL-compatible DB now integrates with Kiro agents and offers a browser-based playground for zero-friction testing. <a href="https://app.daily.dev/posts/Un5GpVc7x" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Cilium hits 10 years with 60 percent CNI market share</strong>: The 1.19 release introduces ztunnel mTLS integration and eBPF-based load balancing that eliminated conntrack scaling issues for Zynga. <a href="https://app.daily.dev/posts/rtti5mtPR" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vercel Workflow implements automatic end to end encryption</strong>: The system now uses per-run key derivation via HKDF-SHA256 and AES-256-GCM to encrypt all step arguments and return values. <a href="https://app.daily.dev/posts/0r6jV9CyY" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Floci emerges as a lightweight LocalStack Community replacement</strong>: With LocalStack sunsetting its free tier, Floci offers a 90MB Docker image that starts in 24ms and supports 20+ AWS services. <a href="https://app.daily.dev/posts/hbh0eKAmA" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux 7.0-rc5 fixes critical sign extension bug in KVM</strong>: A developer-submitted patch fixes a C integer promotion bug that caused CPU core crashes when TSS base addresses had the high bit set. <a href="https://app.daily.dev/feed-by-ids?id=KXPgnQAH4&amp;id=1tWnyX7nj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Project Detroit reopens to bring JS and Python to the JVM</strong>: The initiative leverages the Foreign Function and Memory API to integrate these languages directly into the Jakarta EE ecosystem. <a href="https://app.daily.dev/posts/24VxbiILg" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>MkDocs ecosystem fractures following maintainer conflicts</strong>: The project has split into three successors—ProperDocs, MaterialX, and Zensical—after the original author returned to build an incompatible v2.0. <a href="https://app.daily.dev/posts/GedRmjUpE" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Vibe coding is the new fast fashion for software</strong>: Industry talking heads are now unironically pushing ‘vibe coding’ as a shift from craft to ‘assembly-line’ function, where we just discard code like a cheap H&amp;M shirt. It’s a convenient narrative for companies selling managed infra, ignoring the fact that ‘vibe-coded’ abstractions fall apart the moment you hit a complex edge case like collaborative text editing. If your architectural strategy is ‘just ask the LLM again,’ you’re not a developer; you’re a prompt-dependent liability. <a href="https://app.daily.dev/feed-by-ids?id=AKKujKavD&amp;id=3336MSPgf" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>The 80 percent AI generated production SaaS marketing trap</strong>: A new ‘free’ course is being shilled using the unverifiable claim that a production SaaS was built with 80% AI-generated code. It’s the classic ‘I found a shortcut’ grift that conveniently ignores the 20% of human engineering that actually keeps the lights on and the database from melting. They’re selling ‘architectural patterns’ for AI tools, which is just a fancy way of saying ‘please give the LLM a chance to not write spaghetti.’ <a href="https://app.daily.dev/posts/IoV0luSh7" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ZeroPath promises to turn security alarms into executable fixes</strong>: ZeroPath is the latest YC startup promising to replace your entire security toolchain with a ‘unified platform’ that magically verifies exploitability. While they talk a big game about AST-based enriched graphs, the reality is that semantic rule generation via LLMs is still a hallucination-prone mess in complex multi-language environments. It’s another ‘AI-native’ wrapper trying to disrupt a space where false positives are a feature of the underlying complexity, not something you can just ‘vibe’ away. <a href="https://app.daily.dev/posts/5p9F6EuUW" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
