Systems, Methods, Applications: Three Paper Types

If you read a systems paper as if it were a methods paper, you will be frustrated. Each type has its own rhythm, and good readers know the difference.

Systems papers (e.g., TPU, Megatron, vLLM)

These papers describe real systems at scale. Megatron-LM described how NVIDIA trained huge language models across thousands of GPUs. Judge these on engineering insight and practical trade-offs, not theoretical novelty.

Methods papers (e.g., Transformer, LoRA, Chain-of-Thought)

These propose a new algorithm, architecture, or training trick. Judge them on whether the ablation isolates the contribution and whether the baseline is fair.

Application papers (e.g., AlphaFold, AI diagnostics)

These apply AI to a specific domain — protein folding, climate, medicine. Judge them on domain impact, not architecture novelty. AlphaFold's science was the point, not the neural network itself.

“All models are wrong, but some are useful.”

The big idea: match your reading lens to the paper's goal. You will save time and make fairer judgments.