Jargon:Foundation model
1 min read
Say this at the party
“One huge, general AI model trained on masses of data, which other companies then build their own apps and tools on top of.”
In plain English
Training a big AI model from scratch costs a fortune, so only a handful of companies do it.
The model they make is general. It can write, summarise, translate and more. Lots of other businesses then take that model and adapt it for their own job, like a chatbot for a bank or a helper for lawyers.
It's called a foundation because everything else gets built on top of it, which also means its flaws get built in too.
Where it came from
Researchers at Stanford's new Center for Research on Foundation Models coined the term in August 2021, in a long report on the opportunities and risks of these models.
See it in action
A small UK startup wants an AI that answers questions about tenancy law. Instead of building one from nothing, it takes an existing model such as Claude or Gemini and builds on top of it.
What people get wrong
- Foundation model and large language model overlap a lot, but they aren't the same. Foundation models can also work with images, sound or other data.
- Building on a foundation model doesn't make an app safe or accurate by default. Problems in the base model carry through.
Where Ugenie fits
Ugenie is built on top of foundation models. Its AI Council asks several of them (Claude, ChatGPT and Gemini) the same question and flags where they disagree.
Get Ugenie →Go deeper
- PaperBommasani et al., Stanford, 2021On the Opportunities and Risks of Foundation ModelsThe report that coined the term. Long, but the opening sections are readable.
- ArticleWikipediaFoundation modelA tidy summary of the definition, the history and the debate around the name.
- GuideIBMWhat Are Foundation Models?A plain explainer of how foundation models are built and adapted.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.