Why scaling LLMs won't lead to real agency: A 3-tier Embodied AI architecture

Far-Start-1789 · reddit · 2026-08-27

The author argues that scaling frozen static models cannot create true agency or self-consciousness, proposing an Always-On Embodied AI Architecture with three tiers:

Key mechanics include using exponential decay ($\gamma < 1.0$) for a "fluid present," deducing agency via an Efference Copy auditing physical costs (Watts, thermal spikes), generating "pain" from homeostatic deviations to drive motivation, and an offline "Sleep Mode" using experience replay to prevent catastrophic forgetting without destabilizing real-time weights. The author seeks critique on hardware bottlenecks, specifically if Von Neumann architectures can support the $t-1$ loop or if neuromorphic hardware is required.

Original post →

More from Embodied

Embodied channel →