Canvas is a new interface for working with ChatGPT on writing and coding projects, offering a collaborative environment beyond simple chat. It opens in a separate window, allowing users to create and refine ideas alongside ChatGPT. Built with GPT-4o, Canvas can be manually selected in the model picker during its beta phase. Initially available to ChatGPT Plus and Team users, it will soon be accessible to Enterprise, Edu, and eventually all ChatGPT Free users.

Better Collaboration with ChatGPT

Canvas enhances collaboration by allowing ChatGPT to understand the context of projects better. Users can highlight specific sections for ChatGPT to focus on, providing inline feedback and suggestions. The interface allows direct editing of text or code, with shortcuts for adjusting writing length, debugging code, and other actions. Canvas can be triggered automatically or by including "use canvas" in prompts.

Writing Shortcuts

• Suggest edits: Inline suggestions and feedback.
• Adjust the length: Modify document length.
• Change reading level: Adjust reading level from Kindergarten to Graduate School.
• Add final polish: Check for grammar, clarity, and consistency.
• Add emojis: Insert relevant emojis for emphasis.

Coding in Canvas

Canvas simplifies tracking and understanding code revisions, offering transparency into edits.

• Review code: Inline suggestions to improve code.
• Add logs: Insert print statements for debugging.
• Add comments: Add comments for clarity.
• Fix bugs: Detect and resolve errors.
• Port to a language: Translate code into JavaScript, TypeScript, Python, Java, C++, or PHP.

Training the Model to Collaborate

GPT-4o was trained to collaborate as a creative partner, triggering Canvas for writing and coding tasks, generating diverse content, making targeted edits, rewriting documents, and providing inline critique. The model's performance was measured through over 20 automated evaluations and novel synthetic data generation techniques.

Challenges and Improvements

• Triggering Canvas: The model was trained to open Canvas for appropriate prompts, achieving 83% correct triggers for writing and 94% for coding.
• Editing Behavior: The model was tuned to perform targeted edits when users select text, otherwise favoring rewrites.
• Comment Quality: Human evaluations showed Canvas outperforms zero-shot GPT-4o by 30% in accuracy and 16% in quality.
• Canvas represents a new approach to making AI more useful and accessible, marking the first major update to ChatGPT’s visual interface. Currently in early beta, its capabilities will be rapidly improved.