Tech Videos — 2026-05-08#

Watch First#

How Transformers Finally Ate Vision – Isaac Robinson, Roboflow is the essential watch today for a clear, technical breakdown of how Vision Transformers (ViTs) defeated CNNs despite having a far worse O(n⁴) compute scaling factor. It perfectly illustrates how massive self-supervised pre-training strategies overrode the need for hand-crafted inductive biases in computer vision architecture.

Highlights by Theme#

Developer Tools & Platforms#

Over on the OpenAI channel, the demo for Codex can now use Chrome directly on macOS and Windows. shows the coding assistant operating within a user’s logged-in browser session, controlling multiple tabs in parallel without resorting to brittle screenshot-and-reason loops. For a front-end reality check, the Syntax crew in Date pickers suck, we fixed them benchmark custom natural-language date inputs against standard pickers, highlighting the hidden bloat of Lit web components when mistakenly shipping massive global holiday datasets to the client. Additionally, the Lex Clips segment Should programmers learn how to code in assembly? | Lex Fridman Podcast provides a sobering take on the “Rewrite it in Rust” trend, arguing that memory-safe wrappers fall apart when interacting with the highly optimized, hand-written assembly jumps required for video codecs like FFmpeg.

AI & Machine Learning#

The AI Engineer channel posted FLUX, Open Research, and the Future of Visual AI — Stephen Batifol, Black Forest Labs, outlining a “Selfflow” self-supervised training approach that drops external encoders (like DINO) to achieve better anatomical consistency and text rendering across unified image, video, and audio modalities. Also from AI Engineer, Agentic Search for Context Engineering — Leonie Monigatti, Elastic breaks down why “agentic RAG” often fails when relying on a single tool, arguing for a mix of specialized lookups for reliability and general-purpose shell or ESQL execution tools to handle complex, unexpected queries. Finally, Stripe’s A conversation with Manus AI’s cofounder and CPO Tao Zhang covers the underlying engineering of their autonomous agents, highlighting how giving the AI its own sandboxed cloud computer enables asynchronous execution without requiring constant human-in-the-loop approvals.

Hardware & Infrastructure#

The All-In Podcast episode Elon’s Anthropic Deal, The Next AI Monopoly?, “FDA for AI” Panic, Trading the AI Boom unpacks the market implications of Anthropic leasing compute from xAI’s Colossus data center to bypass their power constraints, effectively crowning “Elon Web Services” as a legitimate new hyperscaler. On the edge compute side, NVIDIA Developer’s DGX Spark Live: NYC Spark Hack Winner feature - 3D time machine for NYC demonstrates a Rust Bevy engine rendering a million polygons at 60 FPS while running the 30-billion parameter Neotron 3 Nano MoE model entirely in unified memory without costly device swapping.

Everything Else#

In a fascinating security post-mortem, Lex Clips’ How CIA spied on people using fake VLC video player | Lex Fridman Podcast details how Vault 7 exploits used targeted fake downloads to slip a malicious .dll into VLC, prompting ongoing engineering efforts to securely sandbox the media player’s sub-processes. For a broader perspective on human history, Dwarkesh Patel’s David Reich – Bronze Age shock, the Neanderthal puzzle, & farming’s sudden spread reveals that the human genome underwent its most intense period of natural selection during the high-density urbanization of the Bronze Age, challenging the assumption that our biology has been stagnant since the dawn of agriculture.