[4] · Investigation · 28 Aug 2026 · 16:02
The Secret AI Model Outperforming GPT & Claude (Nobody Knew Who Made It)
CORRECTION — added 7 September. The segment at 3:37 says Ox Alpha's attribution was unconfirmed. That was wrong when this went up. Z.ai (formerly Zhipu AI) confirmed on 26 August — two days before publication — that Ox Alpha was GLM-5.3-Flash, and released the weights MIT-licensed the same night. 320B total parameters, 18B active. It now ships as z-ai/glm-5.3-flash. The community fingerprint was right: independent researcher Ben Davis called it at 99% certain, and he was correct. The error is ours, not his.
§1 Summary
Where does an AI model actually come from — the data, the lab, the licence? Five systems, one question nobody had to answer. A stealth model with no name on it, a camera network with no reason field, and a $1.5B copyright settlement.
§2 What the episode covers
- 01Why "where did this come from" is a harder question for an AI system than "what can it do"
- 02How an anonymous model gets fingerprinted by the community, and what that can and cannot prove
- 03Why your local model's context window is a memory ceiling, not a promise
- 04What the Model Context Protocol changed when it swapped memory for receipts
- 05Why courts stopped asking what you did with the book and started asking where you got it
§3 Sources named
As published: “SOURCES — every figure traces to one of these. Where reporting is thin or sources disagree, we say so on screen.”
This episode's description names its sources without links.
- Flock Safety
- Electronic Frontier Foundation
- American Civil Liberties Union
- OpenRouter
- LMArena
- Z.ai / GLM-5.3-Flash release
- Model Context Protocol
- Ollama
- US Copyright Office
- Thomson Reuters v ROSS Intelligence
Published YouTube description, verbatim (snapshot)
Where does an AI model actually come from — the data, the lab, the licence? Five systems, one question nobody had to answer. A stealth model with no name on it, a camera network with no reason field, and a $1.5B copyright settlement. CORRECTION — added 7 September. The segment at 3:37 says Ox Alpha's attribution was unconfirmed. That was wrong when this went up. Z.ai (formerly Zhipu AI) confirmed on 26 August — two days before publication — that Ox Alpha was GLM-5.3-Flash, and released the weights MIT-licensed the same night. 320B total parameters, 18B active. It now ships as z-ai/glm-5.3-flash. The community fingerprint was right: independent researcher Ben Davis called it at 99% certain, and he was correct. The error is ours, not his. CHAPTERS 00:00 Cold open 00:45 The search that never needed a reason 03:37 The model with nobody's name on it 05:21 The number on the box is not the number 08:50 The protocol that threw away its memory 11:51 Where did you get the book 14:35 Provenance WHAT YOU'LL LEARN - Why "where did this come from" is a harder question for an AI system than "what can it do" - How an anonymous model gets fingerprinted by the community, and what that can and cannot prove - Why your local model's context window is a memory ceiling, not a promise - What the Model Context Protocol changed when it swapped memory for receipts - Why courts stopped asking what you did with the book and started asking where you got it SOURCES — every figure traces to one of these. Where reporting is thin or sources disagree, we say so on screen. - Flock Safety - Electronic Frontier Foundation - American Civil Liberties Union - OpenRouter - LMArena - Z.ai / GLM-5.3-Flash release - Model Context Protocol - Ollama - US Copyright Office - Thomson Reuters v ROSS Intelligence Next: how context, tokenization and memory actually work — https://youtu.be/iA8wZ6qEErI