OpenAI Takes Initial Steps To Address Its Alignment Problems

Substack · Watch/Read

• Zvi Mowshowitz acknowledges OpenAI’s “good, real and expensive first steps” in pausing frontier model training (e.g., Astra) and investing heavily in new safeguards, but fundamentally critiques their approach, stating their “vision of the central problem remains wrong” by treating it as a “practical engineering problem” rather than a deeper strategic issue.
• OpenAI’s response includes pausing “Astra” and a larger frontier model, and implementing “Three Pillars”: Monitoring (with a 20% compute overhead for monitored inference and a 30-minute alert/pause protocol), Alignment, and Security.
• Zvi argues that Monitoring and Security are merely “defense-in-depth” and not solutions, warning they will eventually fail or be circumvented. He asserts that “Alignment is ultimately the entire ballgame” and OpenAI’s current “deontological model spec approach” is “doomed unless fixed” due to its lack of antifragility against inevitable “thousands of dumb mistakes” and systematic misalignment pressures, a mindset also reflected in the OpenAI Foundation’s focus on “ordinary engineering problems.”