Why LLMs cannot reach GenAI, but why it looked like they could (haversine.substack.com)

This piece argues that the “spark” of reasoning in large language models (LLMs) is real but misleading: scaling laws and the 2022 Google findings revealed surprising emergent behaviors—reasoning-like answers, novel modality outputs (e.g., drawing from text)—which fueled AGI hype. However, those capabilities reflect LLMs learning the gestalts or statistical patterns of human communication, not building genuine internal models of the world. Trained only on what humans bother to say, LLMs miss the tacit, embodied, and subconscious information that underpins human cognition, so they can mimic reasoning without the underlying world-model, and remain brittle and prone to overfitting. The significance for AI/ML is practical and conceptual: language uniquely encodes traces of human thought (motivated by evidence like the Flynn Effect and language’s role in meta-cognition), which made emergent capabilities more likely in text than in vision or game domains. But this also implies scaling alone won’t produce AGI — you don’t get grounding, causal models, or self-updating internal representations just from more text. The technical takeaway: move beyond pure-scaling of text-only models toward multimodal, embodied, or causal-learning paradigms and new evaluation protocols that distinguish pattern mimicry from true world-model-based reasoning.