Tech Videos

Watch First

Why Codex was built in Rust by The Pragmatic Engineer is essential viewing for senior engineers evaluating agent infrastructure. OpenAI explains how building their core agent harness in Rust provided critical compile-time safety, memory efficiency, and deterministic control required for datacenter-scale agent execution, despite Python and TypeScript dominating baseline model training distributions.

Highlights by Theme

Developer Tools & Platforms

In Introducing the Agents API by OpenAI, the team demonstrates a hosted Codex harness handling session orchestration, context compaction, and multi-agent delegation while streaming structured logs via Model Context Protocol (MCP) tool calls to minimize token bloat. On AI Engineer, Jeremiah Lowin presents Generative UI… in Python? — Jeremiah Lowin, Prefect, introducing Prefab—a lightweight Python DSL that compiles declarative UI components into compact Python/JSON payloads rendered natively by client-side MCP applications. Additionally, GitHub’s Attach images to PRs and Issues with GitHub CLI illustrates autonomous agent verification where Playwright captures before-and-after UI screenshots attached directly to pull requests via gh issue create --attach.

AI & Machine Learning

Leading with deep algorithmic theory, Microsoft Research presents Probabilistic Inference for Controlling Diffusion Models, formalizing sequential and parallel path-based Monte Carlo sampling to steer diffusion outputs accurately without fine-tuning weights. On Computerphile, Rob Miles discusses The AI Language We Can’t Read: Neuralese ft. Rob Miles - Computerphile, highlighting how recurrent hidden-vector passes (“Neuralese”) increase multi-step reasoning depth at the expense of losing human-readable chain-of-thought safety monitoring. Meanwhile, Google Cloud Tech’s Graph Engineering with ADK demonstrates how replacing fragile single-prompt loops with explicit DAG architectures (fan-out, join nodes, and deterministic routers) eliminates hallucinated tool calls.

Hardware & Infrastructure

On Bloomberg Tech, d-Matrix Plugs Into Nvidia’s AI Ecosystem breaks down d-Matrix’s Raptor XPU, which integrates 3D-stacked DRAM directly with compute to bypass the memory bandwidth wall for ultra-low latency inference in agentic coding workloads. In iPhone 18 Pro/Duo Impressions: Mogged, Marques Brownlee evaluates Apple’s 2nm A20 Pro chip featuring 50% greater memory bandwidth alongside the foldable iPhone Duo’s structural carbon-fiber hinge and nano-texture anti-glare screen layer. Additionally, Dwarkesh Patel features Dylan Patel in Where Does the Money Actually Go in AI? - Dylan Patel, analyzing the shifting profit margins between semiconductor foundries, high-bandwidth memory (HBM) suppliers, and model inference API providers.

Everything Else

On Lex Clips, David Heinemeier Hansson in DHH on danger of over-optimizing life | Lex Fridman Podcast Clips critiques modern hyper-quantification—such as Oura ring tracking and total alcohol elimination—arguing that rigid metric tracking damages essential social friction and human connection. Meanwhile, No Priors: AI, Machine Learning, Tech, & Startups hosts Brian Armstrong in Coinbase’s Everything Exchange: Agentic Finance, Stablecoins & Tokenization with CEO Brian Armstrong, exploring “Agentic Finance” where autonomous AI agents use self-custodial crypto wallets for sub-30-cent micro-transactions alongside Coinbase’s internal recursive self-improvement developer platform.

💡 Want to dive deeper into any specific paper or architectural design, or would you like me to compile these highlights into a tailored report or slide presentation?