#AI#GenerativeAI#LLMs#AITraining#ChatGPT#Chatbots: "The people who build generative AI have a huge influence on what it is good at, and who does and doesn’t benefit from it. Understanding how generative AI is shaped by the objectives, intentions, and values of its creators demystifies the technology, and helps us to focus on questions of accountability and regulation. In this explainer, we tackle one of the most basic questions: What are some of the key moments of human decision-making in the development of generative AI products? This question forms the basis of our current research investigation at Mozilla to better understand the motivations and values that guide this development process. For simplicity, let’s focus on text-generators like ChatGPT.
We can roughly distinguish between two phases in the production process of generative AI. In the pre-training phase, the goal is usually to create a Large Language Model (LLM) that is good at predicting the next word in a sequence (which can be words in a sentence, whole sentences, or paragraphs) by training it on large amounts of data. The resulting pre-trained model “learns” how to imitate the patterns found in the language(s) it was trained on.
This capability is then utilized by adapting the model to perform different tasks in the fine-tuning phase. This adjusting of pre-trained models for specific tasks is how new products are created. For example, OpenAI’s ChatGPT was created by “teaching” a pre-trained model — called GPT-3 — how to respond to user prompts and instructions. GitHub Copilot, a service for software developers that uses generative AI to make code suggestions, also builds on a version of GPT-3 that was fine-tuned on “billions of lines of code.”"
