Tech Videos — 2026-04-07#

Watch First#

Agentic Engineering: Working With AI, Not Just Using It — Brendan O’Leary A highly pragmatic talk on moving from “AI as autocomplete” to “AI as collaborator,” outlining a concrete “Research, Plan, Implement” workflow that prevents coding agents from hallucinating or mutating your architecture blindly.

Highlights by Theme#

Developer Tools & Platforms#

How Pretext is fixing browser layout struggles on the GitHub channel showcases a high-performance layout engine that predicts text height and routes lines without relying on expensive DOM reads. On the AI tooling front, Agentic Engineering: Working With AI, Not Just Using It — Brendan O’Leary details how overusing Model Context Protocol (MCP) servers bloats the context window and actively degrades code quality, emphasizing the need for isolated context management. Conversely, How to use MCP servers with GitHub Copilot CLI | Tutorial for beginners offers a practical demo on binding local Playwright and remote Svelte MCP servers into Copilot CLI to run end-to-end tests autonomously. For a faster, fully-integrated IDE experience, 3 updates to know about Google AI Studio highlights a new anti-gravity coding agent that maintains context across multi-file edits and integrates directly with the npm registry.

AI & Machine Learning#

Build, Optimize, Run: The Developer’s Guide to Local Gen AI on NVIDIA RTX AI PCs provides an excellent technical deep dive into running Small Language Models (SLMs) locally, evaluating quantization trade-offs (Post-Training vs. Quantization-Aware Training) for formats like GGUF and NVFP4. It backs this up with a live demo of Ollama chaining bash and file-read tools with a local Qwen model on an RTX 5090. On the security front, An initiative to secure the world’s software | Project Glasswing reveals Anthropic’s “Claude Mythos Preview” chaining multiple vulnerabilities to discover a 27-year-old zero-day in OpenBSD. Contrasting this offensive capability, Simon Willison in The AI Challenger disaster prediction warns that the industry’s normalization of deviance regarding prompt injection makes a massive, headline-grabbing security disaster inevitable. Shifting to macro policy, Sam Altman on Building the Future of AI discusses a “superintelligence blueprint,” arguing that future cybersecurity will require a robust ecosystem of AI defenders rather than simply gating the frontier models.

Hardware & Infrastructure#

While part of a broader science and philosophy discussion, Michael Nielsen – How science actually progresses includes a notable infrastructure highlight regarding Jane Street’s GPU cluster optimizations. Their engineers used CUDA graphs and streams to diagnose rank bottlenecks, shaving 25 milliseconds off training steps to effectively free up thousands of B200 accelerators.

Everything Else#

In a grounding discussion, Martin Fowler & Kent Beck: Frameworks for reinventing software, again and again compare the current AI boom to the dawn of microprocessors, noting that Test-Driven Development is now vital for reigning in coding “genies”. Cutting through the hype, AlphaFold isn’t about AI - Michael Nielsen reminds us that protein prediction relied primarily on decades of experimental X-ray diffraction data from the Protein Data Bank, not just algorithms. Finally, for a break from AI, 4211 - The Party Pooper Prime - Numberphile details an amusing mathematical sequence anomaly involving prime coordinate lines, and Palantir Exec Fires Back at ‘Surveillance State’ Claims defends their platform as simple data infrastructure for lawful authorities rather than a proprietary dragnet.