Less Weights, More Thought: 

- value moves from bigger weights to better runtime. better schedulers, speculative decoding, adaptive caches, and low cost simulators that stretch reasoning time without ballooning cost

- domains where deliveration outweighs latency (legal, math, strategy games, policy simulatuion) will likely see the earlierst use (resharing high leverage knowledge work)

- Tasks like factual lookup, robotics or unknown unknowns still need new sensrs, world models and grounding. Longer token chains (could plateau w/o breakthroughs in perception and actuation

Scaling Test Time Compute to Multi Agent Civilizations | Source