<h2>Must know</h2>
<ul>
<li><strong>Databricks launches Lakebase for AI workloads</strong>: Databricks just dropped Lakebase, a new PostgreSQL-compatible database specifically engineered for AI workloads. It aims to merge operational database capabilities with their existing data lakehouse architecture, offering a serverless OLTP solution that separates compute from storage. Currently GA on AWS, with Azure in public preview and GCP coming later this year, it supports up to 8TB per instance and includes Postgres 17 with pgvector, targeting real-time ML feature serving and AI agent memory. <a href="https://app.daily.dev/posts/56RHAShlB" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Linux kernel 7.0 rc1 released, succession plan hinted</strong>: The Linux kernel 7.0 rc1 is out, and Linus Torvalds says the major version bump is just because he gets confused by big numbers, not a huge feature. Still, it brings cemented Rust support, faster cache clearing, non-disruptive kernel updates, and updated AMD/Intel silicon support. More importantly, Torvalds mused about a future maintainer, referencing the community’s recently developed succession plan, which is a big deal for long-term project health. <a href="https://app.daily.dev/feed-by-ids?id=cYP8biu57&amp;id=8ssuMcdSU&amp;id=JXfUDc5F4&amp;id=QdVE1yqfx&amp;id=TcabAvodN&amp;id=cQun6gD6y&amp;id=zaDpYuaxU" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Google embedding model deprecation causes vendor lock-in</strong>: Google deprecated its <code>text-embedding-004</code> model, instantly bricking a production RAG-based movie recommender because different embedding models create incompatible vector spaces. This highlights a nasty form of vendor lock-in: your persisted embeddings are tied to a specific model’s geometry, not just an API. The fix involved re-embedding the entire dataset and a blue/green deployment, underscoring the need to treat embedding models as critical infrastructure and design for change. <a href="https://app.daily.dev/posts/xhDI1zj4i" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Good to know</h2>
<ul>
<li><strong>Kubernetes 1.35 enhances AI/ML support</strong>: Kubernetes v1.35, codenamed ‘Timbernetes’, is strengthening its role as the OS for AI/ML workloads with features like workload-aware scheduling and stable in-place Pod resource resizing. <a href="https://app.daily.dev/posts/mAyDHBOpw" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Shuru: local microVM sandbox for AI agents on macOS</strong>: Shuru is a new local-first microVM sandbox for macOS, designed for AI agent workloads, providing secure, ephemeral, and reproducible environments without Docker. <a href="https://app.daily.dev/posts/TqgsBjEoV" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Uber and OpenAI reveal infrastructure traffic control platforms</strong>: Uber’s “Global Rate Limiter” and OpenAI’s “Access Engine” are new infrastructure-level traffic control platforms, handling 80 million requests per second and ensuring fair access, respectively. <a href="https://app.daily.dev/posts/OVoWP2oNj" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Developer productivity lost to non-coding tasks</strong>: A developer’s 10-second code fix for <code>flash-attention</code> took over 10 hours due to non-coding tasks like fighting sandboxed test environments, <code>CUDA</code> conflicts, and build system issues, highlighting that <code>LLM tools</code> don’t help with these bottlenecks. <a href="https://app.daily.dev/posts/OBIqfWI7x" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Rust RwLock underperforms Mutex in read-heavy workloads</strong>: A hands-on investigation found that <code>RwLock</code> performed 5x worse than <code>Mutex</code> for a read-heavy LRU tensor cache in Rust due to cache-line ping-pong, suggesting <code>Mutex</code> might be better for tiny critical sections. <a href="https://app.daily.dev/posts/rFMs5918e" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>JDK 24 brings major security enhancements</strong>: JDK 24 introduces significant security enhancements, including post-quantum cryptography (ML-KEM, ML-DSA), SHA-3 performance gains, and the permanent removal of the Security Manager. <a href="https://app.daily.dev/posts/TLN02dr7b" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>ChatDetector uses LLMs to find RM-API misuse</strong>: ChatDetector, an LLM-powered system, detects resource management API misuse in software, identifying 115 security bugs across six popular libraries by using a ReAct-inspired framework and two-dimensional prompting. <a href="https://app.daily.dev/posts/UGo4Ea1t8" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
<h2>Bullshit police</h2>
<ul>
<li><strong>Anthropic’s Claude compiler video: marketing fluff or real progress?</strong>: Anthropic released a marketing video showing Claude autonomously building a C compiler capable of compiling Linux and Doom. Critics, however, are calling it “dishonest framing,” suggesting the real goal was just a stress test of Claude’s ability to run for an extended period, not a straightforward compiler-building feat. So, another day, another AI marketing video stretching the truth. <a href="https://app.daily.dev/posts/vGYk0DbTO" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Meta patents AI for deceased user simulation</strong>: Meta was granted a patent for an AI system that simulates deceased users’ social media activity, including posting and chatting, using an LLM trained on their past data. Because what we all need is our dead relatives’ digital ghosts haunting our feeds, courtesy of Meta. Ethical and mental health concerns? Nah, just more engagement! <a href="https://app.daily.dev/posts/AeW0pHCcd" target="_blank" rel="noopener nofollow">Read more</a></li>
<li><strong>Neiry tests bio-drones: pigeons with neural implants</strong>: Neiry, a deeptech company, is testing ‘bio-drones’—pigeons with neural implants, solar panels, and cameras—to guide them along preset routes for infrastructure inspection and search-and-rescue. They claim “privacy compliance” through on-device processing. Because nothing says “privacy” like a pigeon with a brain implant and a camera flying over your house. What could possibly go wrong? <a href="https://app.daily.dev/posts/m7v8MA0Dh" target="_blank" rel="noopener nofollow">Read more</a></li>
</ul>
