Definition

What is a foundation model?

A foundation model is a large artificial intelligence model trained on highly varied data, able to perform many tasks without being retrained for each one. The models behind ChatGPT, Claude and Gemini are examples. It stands in contrast to a specialized model, trained for a single task. What it knows comes from its training data: recent information reaches it only if the tool using it reads pages at the moment it answers.

How a foundation model is built

It is built in two stages. The model first learns from a very large quantity of varied data: text, often images or code as well. It draws a general ability from it, not a skill for one specific task. It is then adjusted for a use. An assistant like ChatGPT or Claude is fine-tuned from the preferences of people who compare its answers (reinforcement learning from human feedback, abbreviated RLHF).

The same model can serve as the base for several products. LinkedIn, for example, documents that retrieval of the posts in its feed rests on Llama 3, an open model from Meta that LinkedIn fine-tuned for that task. The term “foundation model” was proposed in 2021 by researchers at Stanford University: it names this common base on which other systems are built.

What the model knows and what it goes looking for

What a foundation model knows comes from its training data. Its knowledge stops at the date its training ended. It predicts the most likely sequence of words based on that data: without a precise instruction, it therefore falls back on the most frequent choices in what it has read. To answer about a recent fact (a product launched last month, a price that changed), the assistant using it has to read pages at the moment of the question. It can then cite them.

Your site therefore meets these models at two distinct moments, through two families of crawlers. A training crawler collects content to feed future models: OpenAI describes its own, GPTBot, as a crawler used to make foundation models more useful and safer. A retrieval crawler reads a page while a person is asking their question, then cites the source. Blocking the first does not remove you from answers. Blocking the second can.

In recommendation platforms

Platforms are also trying foundation models to choose what they show. In January 2025, LinkedIn researchers described 360Brew, a 150-billion-parameter model built on the open model Mixtral 8x22B. A single model handled more than 30 prediction tasks across at least eight surfaces of the platform, reading everything as text. Its results were measured offline, on logged data.

LinkedIn tested 360Brew with a small group of members, then stopped the test. On April 14, 2026, Tim Jurka, Vice President of Engineering at LinkedIn, wrote publicly that 360Brew is not part of the ranking system. A LinkedIn research paper from February 2026 notes that rankers built directly on a language model did not give the best result live: a sequential model, which reads each member's history, won on results as well as on computing cost. A foundation model is therefore not necessarily the best choice once in service.

Example

Take a fictional company: a manufacturer of industrial doors that launches a new model in October. In November, a buyer asks ChatGPT which manufacturers offer that type of door in Quebec. The foundation model behind ChatGPT finished its training before the launch: it does not know the new product.

If ChatGPT search can read the product page, the answer has a good chance of mentioning it and citing it. If the manufacturer blocked OAI-SearchBot thinking it was only refusing training, the page stays out of reach and the answer may leave it out. Blocking GPTBot alone would have changed nothing in that answer.

How we read it

We cannot check what a foundation model retained from your site during its training. Our work therefore focuses on what can be checked: pages that retrieval crawlers can read, information that matches from one source to the next, and a decision on training crawlers made separately, according to your business model.

We ourselves work with Claude, Anthropic's AI, to build and modify websites. Every result goes through automatic checks, then through a person's approval before it is published. On LinkedIn, we judge a post by the inquiries it produces, not by the name of the model that ranks it.

Not to be confused with

Large language model (LLM)
A large language model is a foundation model specialized in text: it reads and writes words. Foundation model is the broader category, which also includes models for images, sound or several types of data at once.
Specialized model
A specialized model is trained separately for a single task, such as predicting a click on one specific surface of a platform. A foundation model serves as a common base for several tasks, with fine-tuning as needed.
Recommendation system
A recommendation system is the whole that chooses what a platform shows you. It can hand part of the work to a foundation model: at LinkedIn, post retrieval rests on a fine-tuned Llama 3, while ranking rests on a sequential model.

Related concepts

Further reading

Related services

Frequently asked questions

Is ChatGPT a foundation model?

ChatGPT is an application. The GPT models that run it are foundation models, fine-tuned for conversation. The distinction matters: the application adds functions the model alone does not have, such as search that reads recent pages and cites them.

If I block training crawlers, do I disappear from ChatGPT or Claude?

No. GPTBot and ClaudeBot are used to train models. Your presence in answers depends on other crawlers: OAI-SearchBot for ChatGPT search, Claude-SearchBot and Claude-User at Anthropic. These are two separate decisions, each made according to your business model.

Does LinkedIn use a foundation model to rank my feed?

Not 360Brew: LinkedIn said publicly in April 2026 that it is not part of its ranking system. Post retrieval rests on Llama 3, a foundation model fine-tuned by LinkedIn. Ranking rests on a sequential model that reads more than 1,000 past interactions of each member.

← The full glossary