Foundation Model

Before foundation models, you trained a separate model for each task — one for sentiment, one for translation, one for summarization. A foundation model flips that: train one large model on broad data once, then adapt it to many tasks through prompting, retrieval, or fine-tuning. The “foundation” is that everything else gets built on top of it.

That generality is the whole shift behind the current AI wave — and the strategic fault line the show keeps returning to. A handful of labs train the frontier models; everyone else builds on them, which raises the questions that matter for builders: how much to depend on a single provider, when open weights are worth the control, and where the durable value sits when the base model is a shared commodity.