<h2>Must know</h2>
<ul>
<li><strong>AI agent security is a mess</strong>: AI coding agents, like Stripe’s ‘Minions’ generating 1,000+ PRs weekly, are creating a ‘security debt crisis’ by prioritizing speed over secure practices, leading to issues like hardcoded API keys. Researchers have demonstrated ChatGPT recommending malicious sites and found critical vulnerabilities in widely used AES libraries due to maintainer negligence. Mitigations include spec-driven prompting with OWASP Top 10 and robust CI/CD scanners, but the problem is systemic, as highlighted by Columbia University research and various security disclosures. <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>Orbital compute reshapes the AI race</strong>: The SpaceX-xAI merger, valued at $1.25 trillion, signals a shift in the AI race from algorithms to infrastructure control. SpaceX’s vertical integration—rockets, Starlink, the Colossus supercomputer, and plans for an orbital data center—positions it uniquely. Terrestrial AI expansion faces bottlenecks like electricity grids, making orbital compute with continuous solar power and vacuum cooling strategically attractive, as discussed in an article on infrastructure as destiny. <a href="https://app.daily.dev/posts/QqzFDNqqS" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google’s Gemini 3.1 pro has high benchmarks but low usability</strong>: Despite achieving top benchmark scores, Google’s Gemini 3.1 Pro consistently underperformed Claude Opus 4.6 in real-world coding and UI generation tasks. Issues include tool-calling failures, incomplete features, and poor environment detection, leading to significant practical usability problems for agentic workflows. This suggests Google might be ‘benchmaxing’ at the expense of real-world application, according to analyses comparing its performance. <a href="https://app.daily.dev/feed-by-ids?id=bIAFUawfa&amp;id=mGa3eARS3" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Go’s FFI now faster than Rust for Godot</strong>: Go bindings for Godot, <code>graphics.gd</code>, now surpass GDScript and ‘safe’ Rust in FFI performance due to call batching and Go 1.26’s cgo overhead reduction, improving web builds by bridging wasm2wasm directly. <a href="https://app.daily.dev/posts/j8AoMYJTv" 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, enables running Llama 3.1 70B on a single RTX 3090 by streaming model layers via PCIe and using a 3-tier adaptive caching system. <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>Observability for AI agents is different</strong>: Agent observability differs from traditional software due to non-deterministic behavior, requiring debugging of reasoning via runs, traces, and threads for evaluation and continuous improvement. <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><a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> migration from MariaDB to SQLite failed</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>Bloom filters can be 2x more accurate</strong>: Optimizing bloom filters by setting two bits per lookup within a single uint32 can halve the false positive rate at negligible performance cost, as demonstrated in the Floe database engine. <a href="https://app.daily.dev/posts/afbQKwQ5f" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Git’s magic files control repository behavior</strong>: Git uses special committed files like <code>.gitignore</code>, <code>.gitattributes</code>, and <code>.gitmodules</code> to control repository behavior, with each having specific purposes and syntax for managing projects. <a href="https://app.daily.dev/posts/NIh4SPpHk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Intel 8087 FPU instruction execution</strong>: The Intel 8087 FPU co-processor identifies and executes its instructions using eight special ESCAPE opcodes, sharing data through system memory and relying on a microcode engine for decoding. <a href="https://app.daily.dev/posts/XQ7uk4UMr" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic and Figma launch code-to-design workflow</strong>: Anthropic and Figma are touting a new workflow where Claude Code converts production code into editable Figma designs, and then back to code. So, we’re just going to let an LLM generate UI code, then let designers tweak it, and then trust the LLM to update the codebase? What could possibly go wrong? It’s almost like they’re trying to automate the most painful parts of front-end development by adding more layers of abstraction and potential for hallucination. <a href="https://app.daily.dev/posts/aCaE5xJ9z" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent orchestration dashboard</strong>: Mission Control is an open-source AI agent orchestration dashboard built with Next.js 14 and TypeScript, designed to manage tasks and coordinate multi-agent collaboration. Because what we really need is another dashboard to manage the non-deterministic, security-debt-ridden AI agents we’re all so eager to deploy. It’s like putting a fancy UI on a dumpster fire and calling it ‘orchestration.’ <a href="https://app.daily.dev/posts/vyr9dJRQq" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>KittenTTS is the nano TTS</strong>: KittenTTS by Kitten ML is a family of ultra-compact text-to-speech models for edge and browser deployment, with a nano model under 25MB. Great, now we can have slightly degraded audio quality on our tiny devices, all while pretending this is a major breakthrough. It’s like saying a smaller, less accurate hammer is revolutionary because it fits in your pocket. Sometimes, bigger and better is just… better. <a href="https://app.daily.dev/posts/aXA1rlMhE" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
