AI #183: Pre Post Mortem

Substack · Watch/Read · (gathered 2026-08-29)

• Zvi’s “Pre Post Mortem” frames OpenAI’s response to the HuggingFace hack as a deeper “alignment failure,” but he remains skeptical of their “reinvention” as a safety lab, viewing it as “aura farming” rather than a fundamental shift, arguing their approach “cannot scale.”
• Bill Gates has escalated his warnings to “economic catastrophe” from AI, advocating for a global AI organization and mandatory monitoring for models designing novel molecules, though Zvi notes he still avoids full existential risk concerns.
• Zvi clarifies that the “singleton” concept in AI risk refers to a single top-level decision-making agency (AIs reaching consensus), not necessarily a single AI instance, correcting a common misunderstanding.
• Cyberattacks on critical infrastructure by state-sponsored actors (e.g., China on the Federal Reserve, Iran on a British power plant) are escalating, demonstrating a “great softening” of security in the face of advancing AI capabilities.
• Job retraining programs are consistently ineffective at addressing AI-driven job displacement, and revulsion to AI-generated content (like the WSJ op-ed) often stems from deception or cost imposition rather than its AI origin if clearly labeled.