As AI becomes more interwoven into our day-to-day lives, we will want to understand how it has been trained to perceive the world so we know what to expect. We may even want to have multiple personas available to us, as I might expect a different tone and voice from my life coach than I would from my personal assistant.
Personas can be defined up-front in the AI's initial training, or at the surface level in the form of subsequent prompts.
Foundational prompt
Each LLM on the market has a foundational prompt that guides the way it thinks about the world and interacts with users. Most of these are private, though Anthropic has shared theirs publicly. It's surprisingly short!
Because of this, the choice of your underlying model will have downstream impacts on the AI. Understanding how each LLM thinks is the first step towards defining the persona for your AI overall.
Subsequent training
When you think about AI personalities, you're probably thinking about voice and tone, contextual awareness, and so on. These are traits that can be trained later on. When creating AIs for your company, think about the different scenarios where it would be playing, and what characteristics are helpful in each. A human-centered approach helps here. If I needed technical support, I want my agent to be helpful and not overly personal. That changes if the person is my professor, or my friend. Take a similar approach to developing AI agents. Some apps allow users to define these themselves, and examining the consumer side of this, like at Character.ai, can help us think through what this looks like in a B2B context.
