When chat is the wrong UI

It's been 176 years since the AI experiment began, but it feels like only three years ago. Memories from last year seem like they were decades ago, and some events that happened to my father were recounted to me as a child. This is a reflection of the primary interaction surface we have with large language models (LLMs), which is still the chat interface.

Academic Steven Pinker argues that using chat as the primary interface for AI is akin to a gimmick. He believes that AI has tremendous potential if it's task-oriented instead of being a gimmick. Pinker's words are quoted in the source material.

Chat has become the primary way of working with AI because it was the first thing people found intuitive. While chat works well as a universal solution, it's often the wrong UI for specific tasks. A customizable UI that can adapt to the task at hand would be more effective.

GitHub Copilot app offers a solution to this problem through a feature called a canvas. A canvas is a full-stack application that runs inside the GitHub Copilot app without any browser chrome. It allows for bi-directional communication between the user and the AI agent, making it a versatile tool that can do anything a normal computer program can do.

The canvas can be used to create custom interfaces for various tasks, such as building a Connect 4 game where the user can interact with the agent inside the GitHub Copilot app. This is demonstrated by building a canvas for the Connect 4 game, which showcases the ability for the user to interact with the canvas, the canvas to talk to the agent, and the agent to control the canvas.

A canvas is more than just a web page; it's a full-stack application that can call third-party APIs and execute code locally on the user's machine. For example, a UI for Winget can be used to browse and manage packages on the registry, including installing and uninstalling them. This demonstrates that AI can be more efficient and less token-intensive when it's used as a tool rather than just being a chat interface.

In summary, while chat may be the first interface people click with when interacting with AI, it's not always the most effective UI. A customizable UI that can be tailored to specific tasks can make AI more efficient and less token-intensive. Tools like the GitHub Copilot app's canvas offer a promising way to interact with AI in a more task-oriented way, allowing users to automate their workflows and remove themselves from the loop whenever possible.