Foundation Model

For most of AI's history, you built a separate model for each task. One model to translate languages. Another to summarise documents. Another to answer customer service questions. Each required its own data, its own training, its own maintenance. This was expensive and slow, and each model was brittle - good at one thing, useless at everything else.

Foundation models changed this. The idea is to train one very large model on an enormous, diverse dataset - text from books, websites, code, scientific papers, conversations - and develop general capabilities that can then be adapted to almost any task. Instead of starting from scratch for each application, you start from this shared foundation.

The scale of training involved is hard to overstate. Building a frontier foundation model requires months of computation across thousands of specialised chips, costing hundreds of millions of dollars. This is why only a handful of organisations - Anthropic, OpenAI, Google, Meta, a few others - are actually building foundation models. Everyone else is building on top of them.

Once a foundation model exists, it can be adapted in different ways: fine-tuned on specialised data, given access to specific tools, or simply prompted in different ways for different purposes. A single foundation model might power a coding assistant, a customer service bot, a document analysis tool, and a creative writing app - all at the same time, across different companies.

This architecture has concentrated enormous power at a small number of model providers, which is one of the defining tensions in the AI industry today. Access to foundation models - how much they cost, who can use them, what restrictions apply - shapes what AI products are even possible to build.