<h2>Must know</h2>
<ul>
<li><strong>AI agents are a security nightmare</strong>: AI coding agents, while boosting productivity (Stripe’s Minions generate 1,000+ PRs weekly), are creating a “security debt crisis” by prioritizing speed over secure practices. Research from Columbia University highlights issues like hardcoded API keys and public access policies. A security researcher even demonstrated ChatGPT recommending a malicious site for Tesseract OCR, leading to a reverse C2 shell, showing how LLMs can be tricked into trusting bad actors. Mitigations include spec-driven prompting with OWASP Top 10 and robust CI/CD scanners. <a href="https://app.daily.dev/feed-by-ids?id=DD6LZnw0Q&amp;id=Et2d8uHu3&amp;id=LUO8dPKde&amp;id=RH1TYGIET&amp;id=UGo4Ea1t8&amp;id=bDJrJKMJX&amp;id=em7afQ5sm&amp;id=l0KLAjYGc&amp;id=ohhVPexbM&amp;id=vXBWwcwBk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google’s Gemini 3.1 pro: great benchmarks, terrible usability</strong>: Despite Google touting Gemini 3.1 Pro as the “smartest model ever made” with top benchmark scores (78% on ARC AGI 2), real-world coding and UI generation tasks show it consistently underperforms Claude Opus 4.6. Reviewers found Gemini 3.1 Pro struggles with tool-calling failures, incomplete features, and poor environment detection, often requiring multiple prompts where Claude Opus 4.6 succeeds in one. This suggests Google might be “benchmaxing” at the expense of practical application, as noted by multiple sources. <a href="https://app.daily.dev/feed-by-ids?id=bIAFUawfa&amp;id=mGa3eARS3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Anthropic and Figma team up for code-to-design workflow</strong>: Anthropic and Figma have collaborated to enable a new workflow where Claude Code can convert production code into editable Figma designs. This allows developers to send a live web page to a Figma file with a simple prompt, facilitating design collaboration. Updated designs can then be round-tripped back to code, with Claude Code updating the codebase to match the new design, streamlining the iteration process. <a href="https://app.daily.dev/posts/aCaE5xJ9z" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>WebMCP: browser-native tools for AI agents</strong>: Google’s WebMCP is a new browser-native API allowing websites to expose schema-defined JavaScript tools for AI agents to invoke directly within the browser, improving determinism and reducing token overhead. <a href="https://app.daily.dev/posts/61GmU0Tcy" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NTransformer runs Llama 3.1 70B on a single RTX 3090</strong>: NTransformer, an open-source C++/CUDA LLM inference engine, claims to run Llama 3.1 70B on a single 24GB RTX 3090 using adaptive caching and layer streaming, achieving 0.5 tok/s with layer skipping. <a href="https://app.daily.dev/feed-by-ids?id=Ahx4zgQks&amp;id=WoJMuyqAQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent observability is different</strong>: Agent observability requires debugging reasoning rather than code, focusing on runs, traces, and threads to evaluate single-step, full-turn, and multi-turn agent interactions. <a href="https://app.daily.dev/feed-by-ids?id=LY7p8IODB&amp;id=ZmcjIu8ai&amp;id=frjHTf2Ap&amp;id=xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Embedding vendor lock-in is real</strong>: The deprecation of Google’s <code>text-embedding-004</code> model rendered thousands of stored vectors unusable, highlighting how persisted embeddings create vendor lock-in and necessitate designing for change. <a href="https://app.daily.dev/feed-by-ids?id=LY7p8IODB&amp;id=ZmcjIu8ai&amp;id=frjHTf2Ap&amp;id=xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Go’s FFI now faster than GDScript and Rust</strong>: Go bindings for Godot (<a href="http://graphics.gd" target="_blank" rel="noopener nofollow">graphics.gd</a>) have achieved FFI performance surpassing GDScript and ‘safe’ Rust, 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>
<li><strong><a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> SQLite migration fails</strong>: <a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> attempted a live migration from MariaDB to SQLite but quickly reverted due to severe performance degradation, with plans to retry after addressing the issues. <a href="https://app.daily.dev/posts/RLQeRSeS0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Database design starts with consequences</strong>: Effective database design begins by understanding business domain consequences, emphasizing rigorous schema changes, deliberate denormalization, and using constraints to encode business rules. <a href="https://app.daily.dev/posts/fI73pTtiz" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>OpenClaw: the AI agent for everything</strong>: Apparently, OpenClaw, an “autonomous AI agent tool,” can handle everything from automated dependency maintenance to cloud cost monitoring and even building entire apps from a PRD using Claude Code and Gemini. Because who needs actual human developers when you have an AI that can do literally everything, except, you know, respond to permission prompts autonomously? Just pre-configure those permissions, and let the magic happen! <a href="https://app.daily.dev/posts/WCj2dFu7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>KittenTTS: nano models for edge AI</strong>: KittenTTS offers “ultra-compact” text-to-speech models, with the nano variant under 25MB, designed for edge and browser deployment. Sure, the audio quality has “noticeable but modest degradation” as the models shrink, and the smallest quantized version has “some artifacts.” But hey, it’s small! Who needs clear speech when you can have tiny models? <a href="https://app.daily.dev/posts/aXA1rlMhE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>SwiftUI’s GeometryReader: the gift that keeps on breaking</strong>: A common SwiftUI pitfall involves using <code>GeometryReader</code> in every list row, causing CPU spikes and dropped frames during scrolling, and even worse, layout shifts when new data is inserted. The fix? Stop using <code>GeometryReader</code> as a screen size API. Because apparently, a tool designed to read geometry isn’t actually for reading geometry in a way that makes sense for dynamic lists. Who would’ve thought? <a href="https://app.daily.dev/feed-by-ids?id=dGEGPWHMv&amp;id=pxYDPz0nO" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
