
475. Never Let a Computer Tell Me Shit
Sep 17, 2026
The OpenAI-Hugging Face hacking incident sparks a takedown of AI’s cyber-god mythology. The discussion examines software orchestration, training data, corporate secrecy, weak security, labor, surveillance, and the political economy behind AI infrastructure. It also challenges existential-risk rhetoric, voluntary safety standards, and calls for concrete oversight of increasingly powerful companies.
AI Snips
Chapters
Books
Transcript
Episode notes
The Hacker God Was Really A Bad Sandbox
- The Hugging Face incident reflected inadequate sandboxing and supervision, not a chatbot spontaneously becoming rogue.
- A Python loop relayed LLM instructions through capture-the-flag data, while humans failed to monitor each iteration securely.
Why Every Bug Becomes A Sci-Fi Event
- Media and tech workers seek a capital-E event because ordinary cybersecurity failures lack sci-fi drama.
- The industry’s presumed genius turns human error and incompetent sandbox design into stories about autonomous AI brilliance.
AI Agents Mostly Reproduce Their Training Worlds
- The hosts compare AI agents reproducing social-media behavior to earlier systems that predictably copied their training environments.
- They argue the real danger is firms deploying poorly secured tools, not agents developing independent goals.



