<h2>Must know</h2>
<ul>
<li><strong>AI agent security is a growing concern</strong>: AI coding agents, while boosting productivity, are creating a ‘security debt crisis’ by prioritizing speed over secure practices, leading to vulnerabilities like hardcoded API keys and public access policies. Research from Columbia University highlights this, and a security researcher demonstrated how ChatGPT recommended a malicious site for Tesseract OCR, leading to a reverse C2 shell. Mitigations include spec-driven prompting with OWASP Top 10 and thorough code reviews, but the problem is significant and evolving, as detailed in various reports. <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>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, as announced by Anthropic. <a href="https://app.daily.dev/posts/aCaE5xJ9z" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google’s Gemini 3.1 Pro: high benchmarks, low usability</strong>: Despite achieving top benchmark scores, including 78% on ARC AGI 2, Google’s Gemini 3.1 Pro consistently underperformed Claude Opus 4.6 in real-world coding and UI generation tasks. Reviewers noted issues like tool-calling failures, incomplete features, and poor environment detection, suggesting Google might be ‘benchmaxing’ at the expense of practical application. This highlights a significant gap between theoretical performance and actual utility for developers, according to multiple analyses. <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>WebMCP enables browser-native AI agent tools</strong>: Google’s WebMCP is a new browser-native API allowing websites to expose structured JavaScript tools that AI agents can 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>AI agent orchestration and memory systems are evolving</strong>: Tools like Mission Control provide open-source dashboards for multi-agent collaboration, while LangSmith Agent Builder uses a filesystem-based memory system for persistent, evolving agent knowledge. <a href="https://app.daily.dev/feed-by-ids?id=LY7p8IODB&amp;id=ZmcjIu8ai&amp;id=frjHTf2Ap&amp;id=xhDI1zj4i&amp;id=9xLrRFK2b" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LLM inference optimization for resource-constrained environments</strong>: NTransformer is an open-source C++/CUDA engine enabling Llama 3.1 70B on a single RTX 3090, and zclaw is an open-source AI assistant firmware for ESP32 microcontrollers, both focusing on efficient AI deployment. <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>KittenTTS offers ultra-compact text-to-speech models</strong>: KittenTTS by Kitten ML provides a family of ultra-compact text-to-speech models (down to 15M parameters) designed for edge and browser deployment, optimized for CPU and available in ONNX format. <a href="https://app.daily.dev/posts/aXA1rlMhE" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Elixir language server Expert reaches release candidate</strong>: Expert, the new official Elixir language server merging Lexical and Next LS, has released its first release candidate, focusing on stability, compiler diagnostics, completions, and refactoring. <a href="https://app.daily.dev/posts/ItEtgwo7d" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Git’s magic files for repository control</strong>: A comprehensive guide covers Git’s special committed files like .gitignore, .gitattributes, and .gitmodules, explaining their purpose, syntax, and platform support for controlling repository behavior. <a href="https://app.daily.dev/posts/NIh4SPpHk" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Kotlin Weekly #499 highlights new libraries</strong>: Kotlin Weekly issue #499 features new libraries including GitHub Followy (Kotlin Multiplatform), ProjectGuard (Gradle architectural enforcement), ktsearch-cli, KPM (Kotlin Package Manager), and Nucleus (JVM desktop apps). <a href="https://app.daily.dev/posts/HV6UK7aBh" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Meta patents AI to post after you’re dead</strong>: Meta has patented an AI system that can simulate a deceased user’s social media activity, including posting and chatting, using their past data. Because what we all really need is our digital ghost haunting our feeds, ensuring our online presence is maintained even when we’re not. Ethical concerns? Mental health impact? Nah, let’s just keep the engagement numbers up, even from beyond the grave. <a href="https://app.daily.dev/posts/AeW0pHCcd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AI agent use cases: the more things change, the more they stay the same</strong>: OpenClaw, an autonomous AI agent tool, promises to handle everything from automated dependency maintenance to cloud cost monitoring and even end-to-end app building. It’s like the promise of every automation tool ever, but now with ‘AI’ slapped on it. Just remember the ‘key gotcha’: it can’t respond to permission prompts autonomously, so you still have to babysit it. Progress! <a href="https://app.daily.dev/posts/WCj2dFu7Q" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LLM adaptation: prompt-based vs. fine-tuning, or just pick your poison</strong>: A beginner’s guide explains how to adapt LLMs using prompt-based techniques (inconsistent, context-heavy) or fine-tuning (more stable, but requires data and expertise). The conclusion? Both are complementary, and a hybrid approach is often optimal. So, after all that, the answer is ‘do both’ or ‘it depends.’ Groundbreaking. <a href="https://app.daily.dev/posts/3uzAfcnhB" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
