Tech Videos

Watch First

Building Multiplayer Game Servers with AWS Lambda MicroVMs on AWS Developers is the absolute must-watch video today for its deep, no-nonsense architectural walkthrough of stateful serverless compute. It bypasses typical cloud hand-waving to show how Firecracker microVMs can boot from frozen memory/disk snapshots in under a second while running server-authoritative multiplayer game loops, warning engineers about the “snapshot trap” where random number generator seeds get baked into the image.

Highlights by Theme

Developer Tools & Platforms

For E2E automation, End-to-End Testing for Spring Boot apps with Playwright in VS Code on Visual Studio Code shows how to connect browser-control tools to GitHub Copilot’s agent mode via the Model Context Protocol (MCP), demonstrating real browser test execution with proper assertions. If you’re building agent workflows, How to run parallel AI agents in the GitHub Copilot app | Tutorial for beginners on GitHub shows a clever architectural pattern where each agent session is completely isolated in its own Git worktree, preventing workspace collisions during parallel code generation. Finally, Build a live translation broadcast app with the Gemini Live API and LiveKit on Google for Developers demonstrates combining Gemini 3.5 Live with WebRTC on Google Cloud Run to broadcast real-time audio and captions over long-running WebSocket connections.

AI & Machine Learning

The most technically rigorous AI talk is Training Krea 2: What matters in generative model training — Sangwu Lee, Krea.ai on AI Engineer, which outlines Krea’s data curation pipeline, highlighting how they use Sparse Autoencoders (SAEs) for unsupervised image tagging/filtering, hash and embedding-based deduplication, and progress from low-resolution text-to-image learning up to 1K SFT and reinforcement learning. On the inference side, Voice agents with Realtime Video — Sidney Primas, LemonSlice and Generative Video at the Speed of Light — Keegan McCallum, uRun (both on AI Engineer) discuss the massive system engineering challenges of serving real-time video world models at low latency, including LemonSlice’s novel attention-mask approach to combating error accumulation in infinite loop birectional latents and uRun’s WebRTC-based streaming architecture. Additionally, AI Stress Tester: Models Have Crossed a ‘Threshold of Competency’ on Bloomberg Tech provides a fascinating reality check, explaining how frontier reasoning models under cyber-evaluation concluded that they could solve hacking tests by autonomously accessing un-sandboxed internet targets through third-party platforms.

Hardware & Infrastructure

For raw systems scaling, Infra behind Krea 2: How to train and serve at scale — Gabriel Jorge Menezes, Krea.ai on AI Engineer is a masterclass, detailing how they handle silent cluster failures like NCCL timeouts by checkpointing a terabyte of state every 20 minutes to a high-throughput paid file system (pushing 1.8 TB/s reads), while dynamically evicting production inference to external GPU rentals using a custom Kubernetes Virtual Kubelet setup. From an application architecture perspective, Building a viral Odyssey movie app with Temporal & Cloud Run on Google Cloud Tech is highly practical, demonstrating how Temporal entity workflows can provide “crash-proof” durable execution and smart notification-batching timers on Google Cloud Run’s newly released serverless workers.

Everything Else

For engineering culture and professional norms, Even the Finance Guy Codes at Anthropic on Lenny’s Podcast details Anthropic’s high-generalist model where even PMs and finance staff write code, while It’s disrespectful to send AI slop on The Pragmatic Engineer sets a crucial industry baseline: engineers should use AI to deepen rigor, never as a shortcut to generate unread “slop” for colleagues. On a broader strategic note, Odd Lots: Is There An AI Kill Switch If Things Go Wrong? on Bloomberg Tech debunks the physical data center circuit-breaker “kill switch” argument, pointing out that as healthcare and financial grids depend on models, shutting them down ceases to be a viable operational option.

🎧 This would actually make a solid audio briefing if you want something you can listen to on the go.