Tech Videos — 2026-06-25#

Watch First#

The Miranda Hypothesis: How Hamilton Poisoned Persona Evals is a rigorous, must-watch takedown of current LLM benchmarks, demonstrating that persona evaluations measure fluency and pop-culture composites rather than historical accuracy.

Highlights by Theme#

Developer Tools & Platforms#

In an excellent architectural argument on the AI Engineer channel, The Log Is The Agent argues that an agent’s state should be treated purely as an append-only event log rather than an active process, making resume, failover, and multi-agent orchestration structurally trivial. On the NVIDIA Developer channel, Query Your Codebase with DeepSeek V4 and vLLM provides a practical demo of serving a 284B DeepSeek V4 Flash model with a 1M context window via vLLM on two B200 GPUs to generate architecture maps from an 80k-token repository. Additionally, the GitHub channel explores the shift from manual prompting to automated agent loops using work trees, cron jobs, and skills in Understanding Loop Engineering.

AI & Machine Learning#

Google Cloud Tech’s Implementing DeepMind innovation: Deep research API showcases a substantive new API that handles long-running, multi-step research tasks, complete with remote Model Context Protocol (MCP) server support and visual chart generation. Over at Microsoft Research, Plenary Talk 1: Navigating the AI Horizon: Promises, Perils, and the Power of Collaboration details the training behind the Pharaoh 1.5 computer-use agent and suggests that agent performance scales differently than standard LLM parameter curves.

Hardware & Infrastructure#

For Kubernetes operators, Engineering the future of Kubernetes for AI at scale from Google Cloud Tech is highly relevant, detailing Dynamic Resource Allocation (DRA) for mapping workloads to underlying hardware topologies and non-disruptive, in-place pod resizing to optimize resource bin-packing. On the hardware front, NVIDIA & Coherent: Reindustrializing America, Manufacturing for the AI Era highlights the scaling of a 6-inch indium phosphide fab in Texas, directly addressing the silicon photonics interconnect bottlenecks in massive AI clusters. Lastly, cryptography engineers should check out Microsoft Research’s Provable Security and Privacy Analysis of WPA3’s SAE and SAE-PK Protocol, which exposes role confusion and offline dictionary vulnerabilities in Wi-Fi 3 authentication logic.

Everything Else#

Microsoft Research dives into the biological constraints of cognitive decline and the search for reliable blood biomarkers using multi-modal data in Plenary Talk 4: Aging Brain Research: Challenges and Opportunities. On the cultural side of engineering, What AI coding costs engineers on Lenny’s Podcast briefly touches on the isolation caused by heavy agent reliance and the deliberate reintroduction of pair programming to maintain team knowledge transfer.