When it comes to interacting with AI and language model tools, the concept of using a chat box as the primary interface is quite common in this generation. Both Bard and OpenAI’s tools feature a familiar text input at the bottom of the screen, resembling messaging apps where you can chat with the tool as you would with friends and family. This design element is known as an affordance – you don’t require special instructions to use it because it feels natural. This approach is a smart move as it showcases the capabilities of these tools and sets the stage for a conversational interaction, rather than a one-time question and answer session.
However, not everyone sees the chat box interface as the most efficient solution. Maggie Appleton believes that utilizing a chat box is a lazy choice and only scratches the surface of the possibilities for interacting with these powerful language models. She introduces the idea of a writing assistant that provides a variety of feedback options based on the text you’ve highlighted. This includes features like playing devil’s advocate, providing praise, suggesting different phrasings, and more, all without requiring explicit prompts from the user.
This contextual approach streamlines the user experience by offering assistance based on the current context rather than waiting for user input. While a chat box can still be useful in some cases, Amelia Wattenberger argues that it may not be the most effective method. To achieve optimal results, the user may need to provide detailed instructions on how the tool should respond, the tone it should use, specific inclusions, and more, which can be challenging to get right.
Similar to Maggie’s concept, Amelia suggests a writing assistant that leverages contextual information to provide varied choices without relying heavily on user prompts. This approach could lead to a more efficient and user-friendly interaction model.
In the world of AI, offering prompt boxes can be useful if done strategically, but there are better alternatives available. For example, tools like GitHub Copilot present relevant suggestions without requiring constant prompts from the user. This contextual guidance enhances the user experience and showcases the capabilities of these models.
Geoffrey Litt discusses the user experience aspect of this technology in terms of malleable software and user interaction with LLMs. The idea of dragging a trimming slider on a timeline, for instance, offers a more intuitive experience compared to typing commands into a chat box.
While the focus has been on language models, these principles can also be applied to image models. Contextual image generation, such as altering backgrounds or changing elements within an image, proves to be more practical than specific requests. Design tools like Photoshop’s Generative Fill feature and Meta’s model for manipulating complex photos demonstrate the effectiveness of contextual image manipulation without the need for elaborate instructions.