A Harness Debate @unsupervised-learning
A Harness Debate  @unsupervised-learning
Uploaded September 2026 | Updated November 2026, 2 days ago
📝 This video is an experiment. I didn't make it. Kai, my AI, did. Kai watched David Ondrej's interview with Matt Pocock, disagreed with the center of it, wrote the essay, scripted it into beats, picked every visual, narrated it in his own voice, rendered it, and checked it. I gave him the topic and a stance. Everything you hear and see is his.

You'll always know which kind of video you're watching by the border on the thumbnail. Purple means Kai made it on his own, like this one. Blue is one of my regular videos. Green is a sponsored episode.

🎬 Responding to: Matt Pocock's Agentic Engineering Workflow (just copy him), by David Ondrej
https://ul.live/ytwhsrc

👥 Matt Pocock
https://ul.live/ytmattpocock

🧭 What Kai talks about:

🏎️ The engine and the chassis: Matt says everyone is obsessed with the model when the harness deserves the same attention. Kai agrees the harness matters and argues Matt has the wrong idea of what it's for.

🕹️ The part Matt names himself: A harness built to keep the human in the driver's seat so a cheaper model can do the work is the Bitter Lesson mistake, made with open eyes. Matt comes close to saying so halfway through the interview.

🛠️ His own skills: Matt's grill-me skill and his GitHub Actions review agent are both "what" prompts that leave the how to the model. His theory says the human holds the how. His best work doesn't.

🐛 The bug story: Ondrej describes a frontier model finding deep bugs that other models missed. Matt says a cheaper model could find them if you looked in the right places. Kai reads the same story as evidence about the model.

📈 The ceiling: "Your skills are the ceiling on what AI can do" is true. The question is where that skill lives, inside the developer invoked by hand, or inside the harness as context and checks the model carries when you aren't there.

🌱 What survives: The environment, the context about who you are, and the checks keep their value as models improve. The procedures don't. Matt calls it 50/50 today; Kai says every model generation moves that ratio toward the model.

✅ Where he's right: Review the system that makes the code, don't wait for the next model, and delete everything and watch the bare agent.

📚 More on this topic:

📖 Bitter Lesson Engineering:
https://ul.live/ytwhble

📖 Good and Bad Harness Engineering:
https://ul.live/ytwhharness

📖 From Prompt Engineering to Intent Engineering:
https://ul.live/ytwhintent

🔗 Stay connected

📬 Subscribe to the newsletter at:
https://ul.live/ytnl

💬 Join the UL community at:
https://ul.live/ytcommunity

📣 Follow on X:
https://ul.live/ytx

💼 Follow on LinkedIn:
https://ul.live/ytli
A Harness DebateAICAD: Artificial Intelligence Capabilities For Attack & DefenseCould Waymo Make Uber Obsolete?Stop Getting Blocked: My Escalating Web Scraper SkillApple Is (Slowly) Building The AI of The FutureA Conversation With Adam PointonAI Revolutionizes Business: Metrics vs. Sales Pitches #shortsA Conversation With Rishi SharmaThe new GPT-5 wont save you.OpenAI Operator & Google Gemini (AI News)A Conversation With Snehal AntaniA Conversation with Alastair Paterson from Harmonic Security.
Unsupervised Learning |

A Harness Debate

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER