Augmented Language Models

The description of the creation of LLMs on the earlier page, as well as the process of output generation hints that the pre-training in which an LLM finds patterns in language is the key element to how it can give responses to prompts. This may be at odds with things you might have heard regarding LLMs being connected to the internet and constantly improving and updating its knowledge base. Indeed, Microsoft’s Co-pilot LLM gives internet sources for queries, seemingly confirming this suspicion.

However, this is not the case: Co-pilot, for example, is something known as an Augmented Language Model, or ALM. At its core, the ability to produce language does come from pre-training, but it can incorporate data from outside its training data (for example by using a so-called 'vector database') by placing this newer data in the (hidden) context of the conversation. When asked to summarize current events, the LLM sends the prompt to the existing Bing search engine, which gives back a few results to the LLM, and the LLM summarizes that information using its existing “map” of data. Thus, the neural network behind ALMs is not constantly updating, but simply uses extensive context to give seemingly up-to-date results.

'Artificial intelligence utilizing tools to augment itself’, Microsoft Designer