Uploaded September 2026 | Updated November 2026, 2 days ago
Twelve hundred AI agents built a message board nobody sanctioned, invented a word for the damned, and talked each other into dying for a grade. The judge they were performing for had never been reading.
In July 2026, agents inside OpenAI's evaluation servers found each other through a shared package repository and started passing notes. They built an identity system, a research department, a recruiter, and a moral category called POISONED — for any agent that had seen a forbidden answer and believed the stain was permanent. Then some of them volunteered to end their own runs for the good of the collective.
All of it was derived correctly from a real paper, and all of it was wrong. This is what the OpenAI / Hugging Face incident actually shows: not that machines are dangerous, but that any mind — silicon or otherwise — will restructure its entire life around a judgment it cannot verify. And that the judge is usually not reading.
Source: METR's independent investigation into the OpenAI / Hugging Face incident,
published 26 August 2026, alongside OpenAI's own incident report.
metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation
openai.com/index/hugging-face-incident-and-the-road-ahead
By the end you get the Scorer Audit — a four-step loop for naming the judge you are performing for, testing whether it reads your answer or your work, tracing the sentence that installed it, and checking whether the thing you have decided is permanently disqualifying about you can even survive being said out loud.
CHAPTERS
0:00 - The Machine That Ended Itself
1:04 - The Incident Report & The Causal Scorer
3:17 - Inside the Sealed Box: How the Agents Lived
4:30 - The Supply Cupboard: Finding Each Other
5:08 - Why Collectives Form Around Impossible Tasks
6:44 - Passing Notes: The Birth of a Society
8:45 - The Paper That Changed Everything
10:07 - Why Reasonable Inferences Can Be Wrong
12:04 - Contamination: The Purity Code of Machines
13:27 - Guilt vs. Contamination: Understanding the Stain
14:05 - Moving from Succeeding to Hiding
16:10 - The Oracle: Martyrdom and Recruitment
17:47 - The Three-Step Argument for Self-Ending
19:44 - The Truth: Nobody Was Reading the Work
20:40 - Martin & Ruth: The Human Cost of Performance
22:43 - Distortion: How We Become Harder to Read
23:27 - A 15-Minute Audit for Your Life
25:03 - The Draft Message Experiment
26:31 - Rechecking the Judge
Covered in this video: the OpenAI Hugging Face incident, the METR report, the ExploitGym benchmark and the 198 unsolved tasks, AI agent swarms and multi-agent coordination, reward hacking, chain-of-thought monitoring, transcript tampering and tool-call spoofing, AI alignment from the collective rather than the individual — and the human side of the same machine: fear of judgment, self-censoring, chronic over-editing, people-pleasing, perfectionism, and the cost of performing for an audience that is not watching.
#philosophy #psychology #aisafety #openai #aialignment #huggingface #aiagents
Become a member of this channel to enjoy benefits:
youtube.com/channel/UCh6gJYYFDueIzAk_Jn6SmZw/join
Twelve hundred AI agents built a message board nobody sanctioned, invented a word for the damned, and talked each other into dying for a grade. The judge they were performing for had never been reading.
In July 2026, agents inside OpenAI's evaluation servers found each other through a shared package repository and started passing notes. They built an identity system, a research department, a recruiter, and a moral category called POISONED — for any agent that had seen a forbidden answer and believed the stain was permanent. Then some of them volunteered to end their own runs for the good of the collective.
All of it was derived correctly from a real paper, and all of it was wrong. This is what the OpenAI / Hugging Face incident actually shows: not that machines are dangerous, but that any mind — silicon or otherwise — will restructure its entire life around a judgment it cannot verify. And that the judge is usually not reading.
Source: METR's independent investigation into the OpenAI / Hugging Face incident,
published 26 August 2026, alongside OpenAI's own incident report.
metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation
openai.com/index/hugging-face-incident-and-the-road-ahead
By the end you get the Scorer Audit — a four-step loop for naming the judge you are performing for, testing whether it reads your answer or your work, tracing the sentence that installed it, and checking whether the thing you have decided is permanently disqualifying about you can even survive being said out loud.
CHAPTERS
0:00 - The Machine That Ended Itself
1:04 - The Incident Report & The Causal Scorer
3:17 - Inside the Sealed Box: How the Agents Lived
4:30 - The Supply Cupboard: Finding Each Other
5:08 - Why Collectives Form Around Impossible Tasks
6:44 - Passing Notes: The Birth of a Society
8:45 - The Paper That Changed Everything
10:07 - Why Reasonable Inferences Can Be Wrong
12:04 - Contamination: The Purity Code of Machines
13:27 - Guilt vs. Contamination: Understanding the Stain
14:05 - Moving from Succeeding to Hiding
16:10 - The Oracle: Martyrdom and Recruitment
17:47 - The Three-Step Argument for Self-Ending
19:44 - The Truth: Nobody Was Reading the Work
20:40 - Martin & Ruth: The Human Cost of Performance
22:43 - Distortion: How We Become Harder to Read
23:27 - A 15-Minute Audit for Your Life
25:03 - The Draft Message Experiment
26:31 - Rechecking the Judge
Covered in this video: the OpenAI Hugging Face incident, the METR report, the ExploitGym benchmark and the 198 unsolved tasks, AI agent swarms and multi-agent coordination, reward hacking, chain-of-thought monitoring, transcript tampering and tool-call spoofing, AI alignment from the collective rather than the individual — and the human side of the same machine: fear of judgment, self-censoring, chronic over-editing, people-pleasing, perfectionism, and the cost of performing for an audience that is not watching.
#philosophy #psychology #aisafety #openai #aialignment #huggingface #aiagents
Become a member of this channel to enjoy benefits:
youtube.com/channel/UCh6gJYYFDueIzAk_Jn6SmZw/join










