Opinion: AGI Might Need Better Structure

its_vayishu · x · 2026-07-14

The author argues that while current model training consumes megawatt-scale compute, the real bottleneck isn't just "how much compute," but "how the structure is organized." Drawing an analogy to the "small-world networks" in neuroscience: local regions are highly clustered, and a few long-range connections tie everything together. This yields a high clustering coefficient and short path lengths, boosting information transmission efficiency. In contrast, many AI systems still rely on densely or rigidly layered structures lacking modularity. The author isn't suggesting this structure will directly solve AGI, but emphasizes that compared to simply piling on scale, we are likely investing far too little in structural design. They also explicitly invite perspectives from those in graph learning, neuromorphic computing, and connectomics on the limits of this approach.

Related event: Opinion: Achieving AGI Requires Network Structure, Not Just Scale(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →