Tech Videos — 2026-06-01#

Watch First#

Build long-running agents with Google’s Agentic Stack | The Agent Factory is the most substantive watch today, cutting through AI hype to offer a pragmatic architecture for true multi-day agent workflows using durable state, event-driven dormancy, and separated evaluation.

Highlights by Theme#

Developer Tools & Platforms#

On the Syntax channel, Red Hat Hacked! ⟡ NVIDIA Enters the PC Race ⟡ GTA 6 Goes Live ⌁ Syntax Weekly ⌁ breaks down the mechanics of recent npm supply chain attacks that exploit pre-install scripts with heavily obfuscated payloads—bizarrely downloading the Bun runtime to execute their theft. Another Syntax episode, 54% AI-Generated and Climbing — State of AI, reviews developer survey data revealing a significant shift toward AI-assisted coding, highlighting Claude’s dominance in positive sentiment and the growing frustration with Copilot’s unprompted code generation.

AI & Machine Learning#

The AI Engineer talk 20 days of compute vs 7 hours: rethinking what state-of-the-art means — Bertrand Charpentier, Pruna is a refreshing look at the hidden compute costs of benchmarking models, arguing for smaller, efficient models on the Pareto front over brute-force foundation models. On the release side, NVIDIA announced Cosmos 3 in Meet Cosmos 3: Our Latest Frontier Model for Physical AI, boasting a novel “Mixture-of-Transformer” architecture combining autoregressive and diffusion towers. To skip the marketing and see how it actually works, watch How to Run NVIDIA Cosmos 3 Reasoner NIM for Video Reasoning for a credible, technical walkthrough of deploying the containerized model and querying it with <think> tags for physical logic reasoning. Finally, OpenAI teased new capabilities via a partner in Lovable on How GPT-5.5 Unlocks Better Planning for Complex Builds, claiming a 31% increase in intent understanding and 22% less context amnesia.

Hardware & Infrastructure#

NVIDIA is aggressively re-architecting hardware specifically for local AI execution, announcing a new ARM-based platform in NVIDIA RTX Spark Reinvents Windows PCs for the Age of Personal AI. It packs a 20-core Grace CPU, 6,144 CUDA cores, and 128GB of unified memory into a Windows PC optimized for running continuous, local agentic sandboxes. For the data center, NVIDIA Vera—The CPU for Agents introduces a custom CPU designed to eliminate bottlenecks in branch-heavy Python runtimes, featuring LPDDR5X memory and a monolithic mesh to keep GPUs properly fed.

Everything Else#

For a break from software, Is string theory dead? | Don Lincoln and Lex Fridman provides a candid discussion on the state of string theory in physics. Lincoln pragmatically notes that while the theory isn’t strictly “dead,” its lack of testable predictions and vast landscape of possible universes makes it an unappealing bet for scientists wanting to see tangible progress in their lifetimes.