Stanford paper argues LLMs already have AGI-level skills — they just need a coordination layer

A new paper out of Stanford, highlighted by @rohanpaul_ai, argues that current large language models already possess the raw cognitive skills required for artificial general intelligence — but that these skills fail to produce reliable, general-purpose reasoning because they lack a coordination layer. The paper introduces two concepts: "anchoring strength," a metric for measuring how reliably a model can maintain consistent reasoning across a task, and MACI (Multi-Agent Coordinated Intelligence), a framework where multiple agents debate to resolve ambiguity.