scaling hypothesis

The scaling hypothesis proposes that simply increasing model size, training data, and compute leads to emergent capabilities and improved performance in language models, without requiring fundamental architectural changes. Core claim: large language models exhibit predictable performance improvements following power-law relationships as scale increases, and qualitatively new abilities emerge at sufficient scale that are absent in smaller models. Evidence supporting: (1) GPT series progression—GPT-2 (1.5B) → GPT-3 (175B) → GPT-4 showed dramatic capability jumps; (2) Smooth loss scaling—test loss decreases predictably as power law of parameters, data, and compute; (3) Emergent abilities—few-shot learning, chain-of-thought reasoning, code generation appeared at scale thresholds; (4) Cross-task transfer—larger models generalize better across diverse tasks. Key scaling dimensions: (1) Parameters (N)—model size/capacity; (2) Training data (D)—tokens seen during training; (3) Compute (C)—total FLOPs ≈ 6ND for transformer training. Nuances and debates: (1) Diminishing returns—each doubling yields smaller absolute improvement; (2) Emergence vs. measurement—some "emergent" abilities may be artifacts of evaluation metrics; (3) Data quality vs. quantity—curation and deduplication can substitute for raw scale; (4) Architecture matters—efficient architectures achieve same performance at lower scale; (5) Chinchilla finding—previous models were under-trained relative to their size. Practical implications: (1) Predictability—can estimate performance before expensive training runs; (2) Resource planning—calculate compute budget needed for target capability; (3) Investment thesis—justified billions in AI compute infrastructure. Limitations: scaling alone may not solve alignment, reasoning depth, or factual accuracy—motivating complementary approaches like RLHF, tool use, and retrieval augmentation.