Why AGI Isn’t a Compute Problem

by No Comments

Just read this fascinating piece on why we might be hitting AGI limits sooner than expected: The Road to Declining Marginal Intelligence.1

The author, Andrew Cote, argues that AI performance follows logarithmic scaling with compute—meaning constantly declining returns on investment. GPT-5’s underwhelming performance compared to competitors seems to confirm we’re entering the “choppy waters” phase.

But this connects to something I’ve been thinking about in (most recently in AI Writing Tools and Cognitive Decline: The Wrong Question?).2 The piece suggests LLMs hit limits because they can’t become “smarter than their environment”, which is human knowledge.

I think this points to something deeper: intelligence isn’t something we HAVE, but something we DO in relationship with our environment. The constraint isn’t computational: it’s environmental coupling.

Current LLMs can only couple with the world through text. They’re doing sophisticated pattern matching on linguistic traces, not the dynamic, embodied sense-making that creates real intelligence.

This reframes the whole AGI challenge. Instead of scaling up compute, maybe we need to scale up the capacity for world-engagement. True AGI might require systems that can participate in what cognitive scientists call “structural coupling”—the ongoing dance between organism and environment that makes intelligence possible.

The author’s prediction about “empirical AI” that learns from direct environmental observation aligns perfectly with this. Intelligence emerges from environmental interaction, not internal processing power.

We’re not hitting AGI limits because our computers aren’t smart enough. We’re hitting them because we’ve been thinking about intelligence wrong from the start.

The intelligence was always in the coupling, not the components.

  1. Cote, Andrew. “The Road to Declining Marginal Intelligence” Substack newsletter. Physics of Industry, August 15, 2025. https://andercot.substack.com/p/the-road-to-declining-marginal-intelligence. ↩︎
  2. Evans-Greenwood, Peter. “AI Writing Tools and Cognitive Decline: The Wrong Question?” Substack newsletter. The Puzzle and Its Pieces, August 12, 2025. https://thepuzzleanditspieces.substack.com/p/where-does-the-thinking-happen-the. ↩︎