Stanford HAI's foundation model explainer is the piece that makes the current AI landscape make sense

The phrase foundation model appears in AI coverage constantly and is rarely explained properly. Stanford's Human-Centred AI institute published an explainer https://hai.stanford.edu/news/what-foundation-model-explainer-non-experts that fills that gap without requiring technical background.

The key insight: a foundation model is a large model trained on broad data that can be adapted to many different downstream tasks. GPT-4, Gemini, Claude, Llama, Stable Diffusion are all foundation models. The reason one organisation can build many different AI products on top of the same model is because the foundation model has learned generalisable representations that can be specialised through fine-tuning, prompting, or RAG without retraining from scratch.

The economic implication is worth sitting with. Training foundation models requires compute resources that only a few organisations in the world can afford. Every company building on top of those models is dependent on that small number of foundation model providers in ways that have no real precedent in software history.

Whether foundation models make AI more reusable, more centralised, or both is a question with genuinely different answers depending on your perspective on who controls the foundation and what conditions they attach to access.

What is the biggest risk you see from the current concentration of foundation model development in a small number of large organisations?