Large Language Models, or LLMs, are the computational heart of the modern AI assistants. ChatGPT by OpenAI popularized the use of LLMs back in December 2022, and demonstrated to everyone what they can do. However, in addition to being really good at creating nice stories, writing songs and poems, AI models such as ChatGPT possess an inherent ability that can be considered a latent superpower - the capacity to acquire new capabilities through example. Giving these examples to LLMs make them power AI assistants to accomplish different tasks. A combination of an explanation of the task and examples is an instruction, also known as prompt. This technique, which is commonly known as few-shot learning within the research community, is a remarkable feat of artificial intelligence.

There are plenty of models available today, with some being open-source like GPT-J, and others being closed-source, like GPT-3.5 or ChatGPT. The other way to think of these models is to distinguish them as those that can simply continue the words based on a given input text - base models, and as those that can try to accomplish the task represented as an instruction - instruction-tuned models.

As you design your AI assistant, you may find it useful to decide which model or models to use in it. For example, ChatGPT is better optimized for supporting chat experiences, while GPT-3.5 is designed to do completions. You may also want to choose a model based on its price at the cost of spending more time tuning prompt for it. In any case, our goal is to give you flexibility in picking the model or models you want to use in your AI assistant.

Changing the model or models you want your AI assistant to use is easy. Each Skill can use its own language model. You just have to pick the one you want.

Build an AI Assistant with different models