The author says that although the AI experiment is actually only three years old, the perception of time makes it feel like 176 years. Noting that interaction with large language models is still predominantly conducted through conversation, the author argues that this may be the wrong user interface in many cases. The author also reports that academic Steven Pinker said it was unfortunate that the first large-scale application of AI was a chatbot, and that its real potential lies in task-focused systems.
According to the author, when users know what they want to do with AI, they need purpose-built interfaces that can be generated instantly. The “canvas” in the GitHub Copilot application is described as an example of this approach. Canvas is a small full-stack application that works without a browser interface, can communicate with a server, and can exchange data bidirectionally with the Copilot agent.
The author explains that with this setup, they created a game in which Connect 4 can be played against the agent within GitHub Copilot. While implying that they could not beat GPT-5.6 Sol, the author says they defeated GPT-5.6 Luna without reasoning. They say creating a canvas is as easy as asking Copilot for the required application.
These applications can call third-party APIs as well as run local code on the computer. As an example, the author presents a Winget interface that can display packages in the registry and install and remove local packages. According to the author, the goal is to avoid wasting tokens by creating tools that make future interactions free, rather than using the agent for every operation.
Why it matters
This approach moves the use of AI beyond the experience of merely generating responses and toward building tools for completing specific tasks. This allows users and developers to design small applications tailored to their needs, capable of exchanging data and, when necessary, working with third-party services or code on the computer, instead of asking for a chat interaction at every step. The practical goal is to reduce the need to instruct the agent again during repetitive operations and the associated token usage; however, the text does not explain what usage limits and access conditions apply to this feature. Therefore, the discussion shifts away from whether the chat interface is sufficient and toward how easily and comprehensively task-oriented applications can be used.
Term: agent
An AI agent is software that calls tools and performs multi-step tasks to achieve a goal rather than simply generating a single response.