Tech Videos — 2026-05-10#

Watch First#

Two Roads to Durable Agents: Replay vs. Snapshot — Eric Allam, Trigger.dev Why: A highly practical look at how the shift to long-running LLM agents breaks traditional stateless backend architectures, requiring a return to stateful compute via microVM memory snapshots.

Highlights by Theme#

Developer Tools & Platforms#

In Two Roads to Durable Agents: Replay vs. Snapshot, Eric Allam argues that standard replay-based durable execution falls apart for long-running AI sessions due to context bloat, advocating instead for using Firecracker microVMs to compress and snapshot machine state in milliseconds. On the product engineering side, Mehedi Hassan details in You can’t just one shot it — Mehedi Hassan, Granola how his team built internal LLM tool-call tracing to debug search trails and refactored their Electron app’s frontend into a web shell for faster CI preview testing. Finally, Sally-Ann Delucia’s Hierarchical Memory: Context Management in Agents outlines the transition from prompt engineering to context engineering, showing how naive context truncation breaks reasoning and how offloading heavy data tasks to sub-agents is necessary to avoid hitting context limits.

AI & Machine Learning#

Hung-yi Lee’s lecture 人工智慧能否自我成長 · Hung-yi Lee provides a rigorous, math-heavy breakdown of self-improving AI, specifically exploring how models can generate their own proxy reward functions (Unsupervised RL) and use entropy minimization for test-time training. He warns that while models can bootstrap themselves to some extent, performance eventually plateaus without external human signals or stronger teacher models. Meanwhile, a brief clip from the All-In Podcast, David Sacks: Is Anthropic Just Standard Oil With Better PR?, cynically evaluates Anthropic’s exponential growth, suggesting their focus on “AI safety” is essentially brilliant PR to construct a massive regulatory moat around a rapidly expanding monopoly.

Hardware & Infrastructure#

Several clips from Lex Fridman’s FFmpeg interviews highlight the gritty reality of low-level infrastructure, with developers detailing FATE (FFmpeg Automated Testing Environment)—a massive pivot-table of volunteer-hosted tests catching miscompilations across a staggering matrix of OS, compiler, and CPU architecture combinations in The impossible task of testing FFmpeg code. They also throw cold water on the “Rewrite it in Rust” meme in Rust vs C vs Assembly programming languages, pointing out that because FFmpeg relies heavily on handwritten assembly for performance, rewriting the C wrappers in Rust breaks the security model anyway unless you implement compile-time assembly instrumentation. On the data center front, Elon was quietly building THIS BUSINESS all along… highlights how xAI is quietly pivoting to become a hyperscale compute provider by monetizing their massive 1.2 gigawatt GPU cluster to subsidize internal model training.

Everything Else#

Eric Ries joins Lenny’s Podcast in How to build a company that withstands any era to discuss corporate governance as a technical defense mechanism against “financial gravity” and misaligned incentives. He highlights how structures like Public Benefit Corporations (PBCs) and Anthropic’s Long-Term Benefit Trust are deployed to legally protect product quality and engineering integrity against hostile takeovers or activist investors pushing for short-term extraction.