<h2>Must know</h2>
<ul>
<li><strong>OpenAI undercuts competitors with GPT-5.4 mini and nano</strong>: OpenAI has launched GPT-5.4 mini and nano, targeting high-volume subagent tasks with a 400k context window. The nano model is priced at a bottom-of-the-barrel $0.20 per million input tokens, specifically positioned to beat Google’s Gemini 3.1 Flash-Lite on cost. Integration is already rolling out to GitHub Copilot, emphasizing a multi-model architecture where these smaller models handle execution while larger ones plan. <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</strong>: Mistral announced Forge at NVIDIA GTC, a platform allowing companies to train custom models from scratch on proprietary data rather than just fine-tuning. Aimed at high-compliance sectors like finance and defense, it supports both dense and mixture-of-experts architectures. Early partners include the European Space Agency and ASML, signaling a push for sovereign AI that doesn’t rely on US-centric providers. <a href="https://app.daily.dev/posts/6pejPwPmX" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA CloudXR 6.0 brings RTX streaming to Vision Pro</strong>: NVIDIA and Apple have integrated CloudXR 6.0 with visionOS, allowing the Apple Vision Pro to stream high-fidelity, RTX-powered 3D applications. The system uses dynamic foveated streaming to maintain 4K quality at low latency while protecting user gaze privacy. Enterprise partners like BMW and Volvo are already using the Swift-based framework for automotive design reviews and digital twins. <a href="https://app.daily.dev/posts/Lajg4CAxX" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Mistral Leanstral verifies code using formal proofs</strong>: This new 120B agent uses the Lean language to ground AI-generated code in formal specifications, claiming massive cost savings over Claude Sonnet. <a href="https://app.daily.dev/posts/XdxpqgLtJ" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Oracle revives Project Detroit for Java and Python interoperability</strong>: The project embeds V8 and CPython runtimes directly into the JVM process using the Foreign Function and Memory API for high-performance scripting. <a href="https://app.daily.dev/posts/aQEMjwL5W" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>LangSmith Sandboxes isolate untrusted agent code</strong>: A new preview feature provides ephemeral microVM environments to safely execute code generated by AI agents with native tracing. <a href="https://app.daily.dev/posts/Uz5xAKtF3" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>PostgreSQL 18 begins the end of MD5 encryption</strong>: The upcoming release deprecates MD5 in favor of SCRAM-SHA-256, issuing warnings for any users still relying on the legacy hashing method. <a href="https://app.daily.dev/posts/nekgWU5A0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Laravel 13 mandates PHP 8.3 and adds native AI SDK</strong>: The update introduces first-class support for PHP Attributes and a first-party SDK for vector search and tool-calling agents. <a href="https://app.daily.dev/posts/Tq7AY8USq" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Vite 8 adopts Rolldown for massive build speedups</strong>: The major release promises 10-30x faster builds by switching to the Rust-based Rolldown bundler. <a href="https://app.daily.dev/posts/db4oA9qt6" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>AMD releases C++ compiler for Ryzen AI NPUs</strong>: The MLIR-AIE v1.3 update replaces Python-based tooling with a native C++ compiler to unlock NPU hardware for non-AI signal processing workloads. <a href="https://app.daily.dev/posts/gLtp8gAqE" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>The World ID digital power of attorney</strong>: World (the artist formerly known as Worldcoin) is launching AgentKit to let you ‘delegate’ your identity to AI shopping bots via iris scans. Because nothing says ‘secure commerce’ like giving a hallucinating LLM a cryptographic power of attorney over your bank account. It’s a desperate bid for relevance for a token that’s down 76% and a hardware orb nobody wants in their house. <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 slop</strong>: Big Tech is patting itself on the back for funding an initiative to help open-source maintainers deal with the ‘bug report’ spam their own AI tools created. We’ve reached the peak of the cycle where the arsonists are selling fire extinguishers and calling it ‘sustainability.’ Good luck to the maintainers triaging 10,000 AI-generated hallucinations with a few million dollars in Azure credits. <a href="https://app.daily.dev/posts/LS9KJQCC0" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>NVIDIA DLSS 5 and the over-processed future</strong>: NVIDIA is rebranding ‘making things look like plastic’ as a ‘major graphics breakthrough’ with DLSS 5. Early feedback suggests the neural rendering looks over-processed and uncanny, yet the marketing machine is full steam ahead. It’s the classic tech pivot: if the upscaling artifacts get too obvious, just call it ‘artistic AI inference’ and charge for the upgrade. <a href="https://app.daily.dev/posts/79zWldGPG" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
