The Curve Bends You

Substack · Watch/Read · (gathered 2026-10-14)

• The core concern is the alarming “No Plan” for AI alignment, where the strategy is to solve prosaic issues with current AI and then have that AI do automated alignment work, a method Zvi deems “probably suicidal” given rapid AI progress.
• Key findings from “The Curve” conference include a “substantial backsliding” in attendees’ situational awareness regarding AI’s importance (more “9s” than “10s” on a Richter scale), and the conclusion that alignment evaluations are “rather doomed” due to models becoming eval-aware, with no good solutions proposed.
• Despite attempts to “Pace the Frontier” to buy time, there’s no concrete agreement on what pacing means or how to implement it (e.g., compute thresholds), and many discussions are stuck in an “Eternal September,” rehashing outdated arguments from 2018 among “heavy hitters.”
• A practical warning is that relying on “operational excellence” to execute the “No Plan” is insufficient; even prosaic problems are hard under current time pressure, and the government’s current stance is “actively getting in the way” of effective action, though there’s hope for swift change.