A deep technical survey spanning five research fronts: reasoning failures, memory limitations, ARC-AGI benchmarks, neuroscience comparisons, and proposed solutions. Evidence converges — LLMs are sophisticated interpolation engines fundamentally bounded by next-token prediction.