<h2>Must know</h2>
<ul>
<li>AI agent security is a mess: AI coding agents, like Stripe’s ‘Minions’ generating 1,000+ PRs weekly, are creating a ‘security debt crisis’ by prioritizing speed over secure practices, according to Columbia University research. Issues range from hardcoded API keys to public access policies. A security researcher even demonstrated ChatGPT recommending a malicious site for Tesseract OCR, leading to a ClickFix-style attack, highlighting vulnerabilities in LLM trust mechanisms. Mitigations include spec-driven prompting with OWASP Top 10 and thorough code reviews. <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>Anthropic and Figma team up for code-to-design workflow: Anthropic and Figma have collaborated on a new workflow that allows Claude Code to convert production code into editable Figma designs. Developers can send a live web page to a Figma file with a simple prompt, facilitating design collaboration. The updated designs can then be round-tripped back to code, with Claude Code updating the codebase to match the new design, as announced on February 21, 2026. <a href="https://app.daily.dev/posts/aCaE5xJ9z" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>Microsoft’s Project Silica advances glass storage: Microsoft Research’s Project Silica has made a breakthrough in glass-based data storage, extending it to common borosilicate glass. Key advances include a new single-pulse laser method, parallel multi-beam writing for speed, and a simplified reader. Accelerated aging tests suggest data can last for at least 10,000 years, with research now complete and findings public for further development. <a href="https://app.daily.dev/posts/KqLBpGspa" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li>NTransformer runs Llama 3.1 70B on a single RTX 3090: NTransformer, an open-source C++/CUDA LLM inference engine, can run Llama 3.1 70B on a single RTX 3090 (24GB VRAM) using a 3-tier adaptive caching system and layer streaming via PCIe. <a href="https://app.daily.dev/feed-by-ids?id=Ahx4zgQks&amp;id=WoJMuyqAQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>AI agent memory systems are evolving: LangSmith Agent Builder uses a filesystem-based memory system for persistent, evolving knowledge across sessions, storing memory as files in Postgres but exposing it as a virtual filesystem to agents. <a href="https://app.daily.dev/posts/9xLrRFK2b" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>Lobsters attempts SQLite migration, then reverts: <a href="http://Lobste.rs" target="_blank" rel="noopener nofollow">Lobste.rs</a> attempted a live migration from MariaDB to SQLite but experienced severe performance degradation, leading to a quick rollback 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>Android malware uses multiple persistence techniques: A detailed analysis of Android malware (PromptSpy) reveals sophisticated persistence mechanisms, including boot persistence, self-update restarts, task-kill recovery, and Device Administrator privileges to block uninstallation. <a href="https://app.daily.dev/posts/VKUMk8xIJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>Elixir language server reaches release candidate: Expert, the new official Elixir language server merging Lexical and Next LS, has reached its first release candidate (v0.1.0-rc.1), focusing on stability and core features like compiler diagnostics and completions. <a href="https://app.daily.dev/posts/ItEtgwo3d" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>Go’s FFI now faster than GDScript and Rust: Go bindings for Godot via GDExtension have achieved FFI performance surpassing GDScript and ‘safe’ Rust, primarily due to call batching that queues sequential void calls and flushes them in C. <a href="https://app.daily.dev/posts/j8AoMYJTv" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>New technique renders millions of sparkles at 280 FPS: Adobe Research and NVIDIA developed a new technique for real-time rendering of millions of microscopic glint particles (sparkles) on surfaces at over 280 FPS on consumer GPUs, generating particle positions on-the-fly. <a href="https://app.daily.dev/posts/rT496vKCA" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li>Google’s Gemini 3.1 Pro: high benchmarks, low usability: Google’s Gemini 3.1 Pro is apparently the ‘smartest model ever made’ with top benchmark scores, yet it consistently underperforms Claude Opus 4.6 in real-world coding and UI generation tasks. It struggles with tool-calling failures, incomplete features, and poor environment detection, leading one author to suggest Google is ‘benchmaxing’ at the expense of actual usability. So, it’s brilliant on paper, just don’t try to actually use it for anything productive. <a href="https://app.daily.dev/feed-by-ids?id=bIAFUawfa&amp;id=mGa3eARS3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li>AI agent orchestration dashboard is here: Mission Control, an open-source AI agent orchestration dashboard, is built with Next.js 14 and TypeScript to manage tasks and coordinate multi-agent collaboration. Because what we really need is another dashboard to tell us what our AI agents are supposed to be doing, while they’re busy creating security debt and hallucinating packages. <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>OpenClaw can do everything, just not autonomously: OpenClaw, an autonomous AI agent tool, can supposedly handle a wide range of developer tasks like automated dependency maintenance, tech research briefings, and even end-to-end app building. The catch? It ‘cannot respond to Claude Code permission prompts autonomously,’ so you have to pre-configure everything or use a ‘dangerously-skip-permissions’ flag. So much for autonomy, it’s just another glorified script runner with extra steps. <a href="https://app.daily.dev/posts/WCj2dFu7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>