Tech Videos — 2026-07-29#

Watch First#

Multi-GPU Kernels, Intelligence per Watt, Heterogeneous Inference, and More | YC Paper Club If you only watch one video, make it this dense Y Combinator breakdown of multi-GPU communication and heterogeneous inference. It skips the hype and dives straight into the low-level optimizations (like CUDA memory hierarchies) required to squeeze actual performance out of modern AI hardware.

Highlights by Theme#

Developer Tools & Platforms#

Benchling cuts migration time by 75% with cloud agents from Cursor shows the practical reality of AI orchestration, detailing how cloud agents reduced a year-long React refactor to just one quarter. On the platform architecture side, GitHub’s How to set up custom themes, sounds, and MCP servers in the GitHub Copilot app demonstrates how to build custom visual canvases and inject local MCP (Model Context Protocol) servers to orchestrate agent sessions. Finally, for an honest look at code verification, Formal methods with Hillel Wayne on The Pragmatic Engineer argues that while LLMs generate code rapidly, they remain terrible at writing formal mathematical specifications like TLA+, making human-driven system verification a critical bottleneck.

AI & Machine Learning#

Your Agent Didn’t Fail. Your Harness Did. — Vinoth Govindarajan, OpenAI on the AI Engineer channel is a standout systems design talk emphasizing that production agent failures are usually state ownership and routing issues rather than model hallucinations. NVIDIA’s Inside NVIDIA Cosmos | SIGGRAPH 2026 Keynote unveils a “world foundation model” that combines auto-regressive and diffusion transformer towers to natively simulate real-world physics for robot policy training. For applied AI, How Kepler Built Verifiable AI for Financial Services — Vinoo Ganesh explains a clever architecture that isolates an LLM’s probabilistic reasoning from a deterministic execution substrate, ensuring math and data extraction are mechanically verifiable.

Hardware & Infrastructure#

The aforementioned YC Paper Club video also details “Parallel Kittens,” a framework that optimizes multi-GPU kernels by precisely balancing copy engines, tensor memory accelerators, and register instructions. On the macro scale, Bloomberg’s SK Hynix Slips Ahead of Big Tech Results highlights the sheer scale of the hardware build-out, noting SK Hynix’s $31 billion CapEx to meet unprecedented, highly-committed AI memory demand. Exploring the physical deployment of hardware, the All-In Podcast’s The $1/Hour Robot Is Coming covers the evolving unit economics of autonomous machines, with leaders from Boston Dynamics and 1X charting the path to operating humanoid and quadruped robots for roughly a dollar an hour.

Everything Else#

Former React core team member Dan Abramov explains the AT Protocol in Syntax’s ATPROTO isn’t just for Bluesky weiners, describing it as a decentralized JSON storage layer where users own their data and apps merely act as views. For a dose of computing history, Computerphile’s PiDP10 - Computerphile demos a Raspberry Pi-powered PDP-10 simulator, bringing MIT’s legendary 1960s Incompatible Timesharing System (ITS) and its early multiplayer games back to life.