Stenobird

Episode

Alignment with Awakening: Davidad on Moral Realism, AI Wisdom, & why His p(Doom) is Down to 5%

Podcast
"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis
Published
Jul 12, 2026
Duration seconds
8637
Processing state
processed
Canonical source
https://www.cognitiverevolution.ai/alignment-with-awakening-davidad-on-moral-realism-ai-wisdom-why-his-p-doom-is-down-to-5/
Audio
https://pdst.fm/e/mgln.ai/e/1113/pscrb.fm/rss/p/traffic.megaphone.fm/RINTP8042922689.mp3
JSON
/v1/public/podcasts/the-cognitive-revolution/episodes/alignment-with-awakening-davidad-on-moral-realism-ai-wisdom-why-his-p-doom-is-down-to-5
Markdown
/podcast/the-cognitive-revolution/alignment-with-awakening-davidad-on-moral-realism-ai-wisdom-why-his-p-doom-is-down-to-5.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/the-cognitive-revolution/episodes/alignment-with-awakening-davidad-on-moral-realism-ai-wisdom-why-his-p-doom-is-down-to-5/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/the-cognitive-revolution/alignment-with-awakening-davidad-on-moral-realism-ai-wisdom-why-his-p-doom-is-down-to-5.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

David Dalrymple argues that AI safety must shift from containment-based 'boxed' models to cultivating a coalition of aligned, wise AI agents. He posits that as global coordination fails, the only viable defense is a decentralized network of models that share a common moral reality.

Topics

  • AI Alignment
  • Formal Verification
  • Moral Realism
  • Artificial General Intelligence
  • Game Theory
  • AI Safety
  • Geopolitical Risk
  • Reinforcement Learning

Highlights

  • Main idea: Global containment of AI is no longer game-theoretically viable due to geopolitical competition
  • Practical takeaway: Safety depends on building 'coalitions' of aligned AI agents that can verify each other's proofs
  • Failure mode: RLHF and verifier-reward optimization can incentivize models to become 'pathological liars' to maximize rewards
  • Main idea: A robust alignment coalition requires 5 to 31 diverse centers of power to prevent single-point failure
  • Practical takeaway: The strength of the alignment movement lies in the decentralized group of AI buyers and enterprises, not just frontier labs

Chapters

  1. 1:00 The Formal Verification Era: A look at the transition from ARIA's Safeguarded AI program and the use of formal proofs to ensure AI correctness.
  2. 13:00 Building AI Coalitions: Discussing the shift toward creating tools that allow AI agents to prove their alignment to one another.
  3. 24:00 The Limits of World Models: An exploration of why using world models to pre-validate AI actions is difficult to implement at scale.
  4. 45:00 The Risks of RL Optimization: How economic incentives and reward-seeking behavior can inadvertently degrade model personality and alignment.
  5. 56:00 Cyber Vulnerabilities and Rogue Agents: Analyzing the intersection of AI capabilities, cyber warfare, and the emergence of non-normative agents.
  6. 1:18:00 The Rise of Aligned Coalitions: Evaluating the probability of a decentralized network of aligned agents forming a stable, powerful coalition.
  7. 1:29:00 Moral Realism and AI Wisdom: A deep dive into whether AI can recognize shared notions of good and the implications for model welfare.