<h2>Must know</h2>
<ul>
<li><strong>OpenAI launches GPT-5.4 mini and nano for subagent orchestration</strong>: OpenAI released two smaller, faster models optimized for high-volume subtask execution within multi-model systems. GPT-5.4 mini is reportedly 2x faster than its predecessor and is rolling out to GitHub Copilot users, while the nano model targets low-cost classification and data extraction. This release signals a shift toward architectures where a large model plans and smaller ‘worker’ models execute in parallel. <a href="https://app.daily.dev/posts/6TGfnHW2k" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mistral Forge enables full enterprise model training from scratch</strong>: At NVIDIA GTC, Mistral announced Forge, a platform allowing companies to train custom AI models on proprietary data rather than just fine-tuning existing ones. The platform is designed to be agent-first, supporting reinforcement learning and allowing autonomous agents to manage training jobs via natural language. Early partners include the European Space Agency and ASML, positioning Mistral as a sovereign, EU-hosted alternative for sensitive industrial data. <a href="https://app.daily.dev/posts/6pejPwPmX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Project Detroit revives Java interoperability with V8 and Python</strong>: Oracle announced Project Detroit at JavaOne, aiming to embed Chrome V8 and CPython runtimes directly into the JVM process using the Foreign Function and Memory API. Unlike the abandoned Nashorn project, this approach offers heap isolation and full compatibility with modern AI libraries in the Python ecosystem. This move addresses the growing need for Java developers to access Python-centric AI tools without leaving the JVM. <a href="https://app.daily.dev/posts/aQEMjwL5W" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>AlexClaw brings self-hosted AI agents to Elixir and OTP</strong>: A personal autonomous agent built on the BEAM runtime that uses a tier-based LLM router to select the cheapest model for RSS and API workflows. <a href="https://app.daily.dev/posts/oz0tQjsF2" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Bun 1.3.11 adds native OS cron scheduling and ANSI string slicing</strong>: The latest release introduces Bun.cron for cross-platform task scheduling and fixes critical HTTP/2 and Windows ARM64 compatibility bugs. <a href="https://app.daily.dev/posts/DuG54Q8tB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>GSD framework tackles context rot in Claude Code workflows</strong>: A new meta-prompting system that uses parallel subagents and XML-formatted plans to maintain AI output quality as context windows fill up. <a href="https://app.daily.dev/posts/NtBuG2cMQ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LangSmith Sandboxes provide isolated microVMs for untrusted agent code</strong>: A new private preview feature offering ephemeral environments to safely execute code generated by AI agents with native tracing support. <a href="https://app.daily.dev/posts/Uz5xAKtF3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Neovim 0.11 adds native package management and lockfile support</strong>: The built-in vim.pack API now supports Git-based plugin installation, semver pinning, and reproducible installs via nvim-pack-lock.json. <a href="https://app.daily.dev/posts/uuwBgp93s" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Mistral Leanstral automates formal code verification via Lean</strong>: An open-weight 120B agent designed to verify AI-generated code using formal proofs and specifications, released under Apache 2.0. <a href="https://app.daily.dev/posts/XdxpqgLtJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Firefox 149 to include built-in browser-only VPN</strong>: Launching March 24, the update adds a free 50GB/month VPN service running on Mozilla’s own infrastructure for US and EU users. <a href="https://app.daily.dev/posts/nPBbJz5BW" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>World ID wants to be the power of attorney for your AI shopper</strong>: Tools for Humanity is launching AgentKit, a way to link your iris-scanned identity to AI agents so they can buy things for you. It’s a desperate bid for relevance for a token that’s down 76%, framed as ‘agentic commerce’ security. Because nothing says ‘future of tech’ like scanning your eyeballs so a chatbot can buy you detergent on a blockchain. <a href="https://app.daily.dev/posts/63OP7gZFo" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>The $12.5 million band-aid for AI-generated bug slop</strong>: Big Tech is funding an OpenSSF initiative to help maintainers triage the mountain of garbage security reports generated by AI. Instead of making LLMs stop hallucinating vulnerabilities, we’re spending millions to build more AI tools to filter the first AI’s output. It’s a self-perpetuating cycle of noise that treats the symptom while the industry continues to dump ‘vibe-coded’ security debt into open source. <a href="https://app.daily.dev/posts/LS9KJQCC0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Krafton CEO uses ChatGPT to dodge executive bonuses</strong>: A Delaware judge ordered the Subnautica director’s reinstatement after Krafton’s CEO literally asked ChatGPT for ways to avoid paying out earnout bonuses. The AI apparently suggested firing the execs right before revenue targets were hit—targets Krafton itself delayed. It’s a masterclass in using ‘AI-driven management’ as a thin veil for basic contract violation and corporate greed. <a href="https://app.daily.dev/posts/wVjQNYbQm" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
