Summary: Training modern LLMs is a costly process that shapes the model’s outputs and involves unsupervised, supervised, and reinforcement learning.
By this point, you’ve undoubtedly heard that the large language model (LLM) behind your favorite AI tool has been “trained on the whole internet.” To some extent, that’s true, but after training hundreds of UX professionals on how to use AI in their work , it’s clear that many don’t understand how the AI is trained. This is crucial for forming an accurate mental model of how these LLMs work, their limitations, and their capabilities.
This article discusses four basic types of training, when these are performed within an LLM, and how they impact the role of AI in user experience.
1. The Pretraining Phase: Unsupervised Learning
When you’ve heard that large language models have been “trained on the whole internet,” people are typically talking about the pretraining phase, which involves unsupervised learning.