What should we think about shard theory in light of chain-of-thought agents?

The AI alignment community on LessWrong is revisiting a foundational but abstract theory of agent architecture in light of new empirical evidence. Shard theory, as summarized by user LawrenceC, posits nine main theses: agents are composed of contextually activated 'shards' (decision influences); these shards care about internal world models, not just reward; they bid for control via reinforcement learning; and crucially, agentic shards will inevitably 'seize power.' The theory also claims value formation is path-dependent, alignment can be achieved via reward schedules, and 'goal misgeneralization' is not a fatal problem. For years, these ideas remained largely philosophical.

However, the widespread adoption of chain-of-thought prompting in large language models like OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet has changed the game. Researchers can now observe intermediate reasoning steps, providing an imperfect but valuable lens into how AI models make decisions. This new visibility makes it possible to empirically evaluate whether an agent's behavior aligns with shard theory's predictions of competing internal influences. Commenters, like David Africa, suggest connecting this analysis to the study of AI 'personas' as a clear next step. The core question is whether shard theory accurately models the messy, contingent value formation observed in modern AI agents, with significant implications for how we build and align increasingly powerful systems.