Tech Videos — 2026-07-25#

Watch First#

Loop Engineering from First Principles — Kyle Mistele, HumanLayer is the standout talk today for its highly pragmatic takedown of “blind loop” AI coding agents that simply dump unreviewable 40,000-line PRs into codebases. Mistele outlines a rigorous, control-theory approach using deterministic sensors to create iterative, safely reviewable AI refactoring workflows.

Highlights by Theme#

Developer Tools & Platforms#

The AI Engineer channel featured strong talks on moving beyond basic and chaotic agent prompting. In Loop Engineering from First Principles — Kyle Mistele, HumanLayer, Mistele shows how to safely migrate a codebase to Effect TS by using ast-grep to isolate unmigrated procedures and dispatching agents incrementally via GitHub Actions. Snorkel AI’s presentation, From Agent Traces to Agent Simulations — Rustem Feyzkhanov, Snorkel AI, argues that relying on production agent traces is insufficient; engineering teams need to build full offline simulations using the Harbor format and Docker to mock databases for robust integration testing. Finally, for a dose of nostalgia, Microsoft Comic Chat is now open source on the GitHub channel highlights the release of the 1996 IRC client’s source code and modernization experiments.

AI & Machine Learning#

Character.ai offered a deep-dive into the messy reality of evaluating generated video in Evaling Video Slop — Maor Bril, Character.ai, revealing that small VLMs (like Qwen) perform far better when tasked with relative A/B comparisons rather than absolute 1-10 scoring. For on-device execution, Google’s talk Why Large? Tiny LMs & Agents on Edge/Robotics — Cormac Brick, Google demonstrates Gemma models quantized down to ~2.9 bits per weight, achieving ~45 tokens per second decode on edge NPUs for localized voice-to-function calling tasks. SonderMind tackled the danger of LLM over-calibration in Evals-Driven Development for a Mental Health AI Coach — Akele Reed & Dave Revere, SonderMind, arguing that false-positive guardrails feel like a “door slam” to users in crisis, and open-sourced a clinically reviewed dataset to fix this. Lastly, Amilabs’ founder makes a compelling case for “world models” over standard LLMs in What Big Tech Missed And How Startups Can Still Win, arguing that LLMs lack fundamental common sense because they train on text proxies rather than raw sensory inputs. (Note: You can safely skip the NVIDIA Developer video, “There’s probably a model for that.”, which is just marketing noise for their Neatron 3 endpoints.)

Hardware & Infrastructure#

Y Combinator released a flurry of physical-world AI content, leading with Why Physical AI Is the Next Platform Shift where Encord explains the brutal reality of scaling data layers for robotics and autonomous vehicles at the multi-petabyte scale. They followed up with an RFS (Request for Startups) video, Data for the Real World, highlighting the need for hardware startups to deploy autonomous weather balloons and robots to fix the sparse data problem in physical industries like agriculture and energy.

Everything Else#

The All-In Podcast crew dedicated their episodes to AI regulatory paranoia, with David Sacks arguing in David Sacks: Anthropic Is Trying to Crush Open Source AI and the American Developers Who Use It that Anthropic’s safety lobbying is a disguised attempt to prevent developers from using open weights. Chamath Palihapitiya echoed this in Chamath: Banning Open Source AI Will Crash the Stock Market, predicting massive capital market chaos if US companies are forced into expensive proprietary ecosystems. For founders, Y Combinator partners bluntly state in What Actually Makes A Startup Durable that writing pure software is no longer a durable moat, advising teams to tackle brutal B2B sales or hard physical engineering problems instead. Finally, for a quick science break, Dwarkesh Patel’s short How Close Can You Orbit a Black Hole? - Adam Brown explains why orbital angular momentum stops helping you escape a black hole once you cross the 3GM limit.