Tech Videos — 2026-04-03#

Watch First#

37,000 Lines of Slop A vital, pragmatic teardown of AI-generated code hype that demonstrates why blindly shipping 37,000 lines of LLM output a day results in catastrophic, unreviewed production payloads.

Highlights by Theme#

Developer Tools & Platforms#

In The Download: LiteLLM hacked, Pretext layout engine, OpenAI news & more from GitHub, a critical security warning is issued regarding a LiteLLM supply chain attack where malicious PTH files were loaded via a Cursor MCP server to harvest credentials and Kubernetes secrets. Meanwhile, the Syntax channel provides a much-needed reality check on AI productivity metrics in 37,000 Lines of Slop; an audit of a heavily-bragged-about AI-generated project revealed a horrific production bundle containing 300KB of test files, 2MB uncompressed PNGs, and a 520KB backend text editor shipped to the client. For those looking to automate security reviews, Run Security Analysis using Gemini CLI locally and on GitHub shows how to integrate local vulnerability scans and GitHub Actions CI/CD gates using the Gemini CLI extension.

AI & Machine Learning#

AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering… on No Priors is highly technical, featuring Liam Fedus explaining how they use language models as an orchestration layer to direct specialized, symmetry-aware neural networks for atomic systems. On the open-source front, the The State of Open Source AI | NVIDIA GTC panel highlights AI2’s Molmo and MolmoBot, which outperformed larger proprietary models by abandoning generic web text in favor of heavily customized spatial datasets and 3D simulation environments. If you are building conversational interfaces, Build real-time conversational voice agents with Gemini 3 offers a quick, practical demo of the new Live API handling real-time user interruptions and multimodal screen vision.

Hardware & Infrastructure#

In Why did Uber have 5,000 microservices?, The Pragmatic Engineer unpacks the architectural reality that Uber’s microservice explosion was purely a survival tactic to bypass a monolithic backend API and unblock developer velocity under extreme pressure. Looking at physical infrastructure scale, “NVIDIA’s cost per token is the lowest in the world.” explores the brutal economics of AI data centers, noting that amortizing a $40 billion gigawatt facility over 15 years demands the most efficient compute architectures to keep per-token costs competitive.

Everything Else#

The SpaceX IPO, Iran War Fallout, Quantum Bitcoin Hack, The Space Opportunity episode of the All-In Podcast contains an interesting segment on the threat of quantum computing against cryptography, specifically highlighting Oded Regev’s 2023 paper which drastically improved Shor’s algorithm by reducing the operations required to factor large integers from 28 million down to 500,000. For a quick history break, Why Florence’s Top Cop Was Always a Foreigner explains the clever incentive structure of Renaissance Florence, which hired temporary foreign noblemen as police chiefs to enforce laws neutrally and prevent local power grabs.