A blog post titled 'LLM Classification Is Feature Engineering' argues that using large language models for classification tasks is functionally equivalent to automated feature extraction rather than a fundamentally new paradigm. The author contends that the model transforms raw inputs into representations that a simpler classifier can then use, mirroring classical machine learning pipelines. The piece is hosted on minimallysufficient.com and was published on September 17, 2026. It frames current LLM classification workflows as an extension of established feature engineering practices. The post has drawn comments but no major industry response yet.
This framing is not a demotion. It is a promotion. Feature engineering was always the hard part. The part where domain knowledge met data and something useful emerged. LLMs did not erase that work. They automated it.
Think about what this means. A single model can now generate embeddings, extract entities, infer sentiment, and surface latent structure. That used to take a team and a quarter. Now it takes an API call. The pipeline got shorter. The leverage got bigger. We are not replacing feature engineering. We are finally giving it the engine it deserved.
The skeptics will say this reduces LLMs to glorified preprocessing. I say it reveals the truth. Intelligence is compression. Classification is decision. The model compresses. The classifier decides. That division of labor is clean, scalable, and honest. Builders should embrace it. Stop chasing magic. Start shipping systems.