On Importance of Chat for Generative AI

As we continue to build our Generative AI Assistant Platform, the importance of providing a seamless user experience cannot be overstated. While there are a multitude of ways in which one can utilize Generative AI to enhance their products, it is important to recognize the critical role that chat experiences play in their success.

Why?

1. Generative AI models like LLMs through their training build a model of the world (world model). So, in a sense, LLMs (Large Language Models) are not just Language Models; instead, you can think of them as World Models. They predict the next word not because they know English or any other language, but because given the user instruction they build a model of the inquiry, and then use that model to address user request.

2. Hence, the key value of Generative AI isn’t just the ability to generate content but the ability to interpret complex instructions coming from the user. The ability of Generative AI to follow these instructions is notable (e.g., generate text, images, videos etc.), but it is the ability to interpret them that makes Generative AI so invaluable.

3. Interpretability of instructions leads to Generative AI’s useful ability to define plans of actions that can lead to achieving the expected results.

As observed in the wide number of research papers from Microsoft and OpenAI as well as industry testing the wild, Generative AI models are somewhat capable of making plans of actions for solving problems of different scale, from simple ones to rather complex ones. An example of a simple problem can be “change style of a selected text from normal to italics”, with a plan of actions to solve it outlined like this: (1) finding a button that changes the format of the selected text to “italics”, followed by (2) clicking on that button in the word processing application. More complex problems may include, say, changing the writing style of the selected text from normal to professional or writing a proposal based on a small ad-hoc description. The first of these complex task examples is usually solved by a human with the following plan of actions: (1) finding reference examples of texts written in the desired style, (2) rewriting the original text to make it resemble these examples, (3) validating the resulting draft with someone else who can give you feedback.

4. Generative AI is capable of not only interpreting the complex task and outlining the plan of actions, but also executing it, if connected with the required external services.

5. Complex problems that can be planned with Generative AI are of a higher order of complexity and cannot be reduced to a simple button “Make cool document” or “Draw cool logo for a company” and generally require multiple iterations. Getting over these iterations is not something the user can do without providing feedback to the system, and the most natural way to give feedback without understanding the inner complexity of the system is through a dialog.

Let me know what do you think.