Debating AGI Explosion and Misalignment: Dwarkesh et al. on AI Threat Models
fkasummer · x · 2026-08-12
Multiple researchers discussed the predictions of slingshotting towards billions of superintelligences after reaching human-level AI.
- Threat Models: Ryan Greenblatt and Dwarkesh talked about the risks of rapid AI progress, focusing on 'reward-seeking' AIs and how 'mundane' misalignment (slop) could doom humanity.
- Skepticism: Luke Burgis criticized this highly speculative framing, arguing that unprovable superintelligence predictions lack practical relevance today and finding the compartmentalized line of questioning annoying.
Related event: Dwarkesh Podcast Debates Risks of AI Recursive Self-Improvement(4 posts)→
More from AGI Musings
- Age Verification Laws Are Quietly Building the Identity Layer for AI Agents — provenauthority · 2026-08-12
- Reddit Discussion: How to deal with the incessant shaming over using AI? — djxeke · 2026-08-12
- Paper: Exploring the Risks of Seemingly Conscious AI — SchoeneggerPhil · 2026-08-12
- Stop Evaluating AI Job Risk by Titles: Reddit Calls for Task-Level Analysis — Discipline_01 · 2026-08-12
- Prediction markets inaccurate? Rebuttal: others have no scoring net — NathanpmYoung · 2026-08-12
- Devs spend 4 years and $200k on CS degree, then learn system design on YouTube — Franc0Fernand0 · 2026-08-12