Chat is Often the Wrong UI for AI Interactions

The Dominance of Chat in AI

Three years into the widespread adoption of AI, chat remains the primary user interface for interacting with large language models. This prevalence stems from its initial accessibility and versatility, as users explored various applications for AI without predefined structures.

Limitations of Chat for Task-Oriented AI

While chat serves as a universal solution when user intent is unknown, it becomes less effective once a user has a specific task in mind. Academic Steven Pinker notes that the initial large-scale implementation of AI as a chatbot was a 'gimmick,' suggesting that AI's true promise lies in task-oriented applications.

Introducing Customizable UIs: The Canvas Concept

A more efficient approach involves customizable user interfaces that can adapt to specific tasks. GitHub Copilot implements this concept with a feature called 'canvas.' A canvas functions as a full-stack application within the Copilot environment, operating without browser chrome.

How Canvas Enhances AI Interaction

The canvas allows an AI agent to communicate bi-directionally with the server component of the application. This setup provides a flexible surface that can perform any function of a standard computer program while also interacting with the GitHub Copilot agent. This enables more structured and efficient AI interactions for defined tasks.