Tech Videos — 2026-04-17#

Watch First#

Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI is the single most valuable watch for engineering leaders looking to operationalize AI. Lopopolo outlines a pragmatic approach to treating code generation as effectively free, explaining how to shift your team’s focus toward system design, automated linting, and continuous context injection rather than synchronous code review.

Highlights by Theme#

Developer Tools & Platforms#

The Visual Studio Code channel offers a crisp, gimmick-free demo of installing a Formula 1 Model Context Protocol (MCP) server directly into the GitHub Copilot CLI via standard IO. Meanwhile, GitHub’s The Download outlines their new Copilot SDK, which allows developers to embed Copilot’s agent runtime directly into custom applications with open telemetry and BYOK support. For a look at the bleeding edge of UI generation, the Syntax channel hosts a live coding battle using open code and Composer 2, demonstrating how models often get bogged down in deep thinking modes instead of executing practical rendering tasks.

AI & Machine Learning#

In a highly substantive update on the AI Engineer channel, OpenClaw creator Peter Steinberger reviews the massive security challenges of maintaining a popular agent, noting how advanced models can easily chain exploits to break out of sandboxes. A related walkthrough from NVIDIA Developer provides a solid tutorial on deploying a 120-billion parameter Neotron 3 Super model locally in an OpenShell sandbox to create a Telegram assistant with zero cloud dependencies. Additionally, the GitHub update highlights Anthropic’s unreleased Claude Mythos model, which autonomously chained Linux kernel vulnerabilities and discovered a 27-year-old OpenBSD bug, triggering a $100 million defensive security initiative.

Hardware & Infrastructure#

The All-In Podcast features an extensive discussion on how compute constraints, power limitations, and rising local opposition to data center builds could soon cap the exponential growth of frontier AI models. Providing a geopolitical angle on hardware ecosystems, Jensen Huang argues on Dwarkesh Patel’s channel that continuing to sell chips to China is a pragmatic necessity to ensure global developers remain locked into the American software ecosystem rather than optimizing for competitors like Huawei.

Everything Else#

For a fascinating dive into computing history, Computerphile details the reverse-engineering of a B language compiler to execute the original “Hello World” program on a PDP-11 emulator. On the enterprise management front, No Priors interviews ServiceNow CEO Bill McDermott, who pushes back on the “SaaS apocalypse” narrative by calculating that trying to replace mature workflow platforms with raw LLM code generation would cost ten times more and introduce unacceptable unreliability into corporate systems of record.