ChatGPT Canvas Executes Python in Browser Now! But Who Is It Really For?

Key Concepts

• Canvas: A collaborative writing feature within ChatGPT that now supports Python execution.
• Code Interpreter: A ChatGPT tool that executes Python code in a sandboxed environment, allowing for data analysis and file manipulation.
• Pyodide: A Python distribution for the browser based on WebAssembly, used by Canvas for Python execution.
• Artifacts: Interactive outputs in Claude that can include code execution and visualizations.
• WebAssembly: A binary instruction format for a stack-based virtual machine. Wasm is designed as a portable target for compilation of high-level languages like C/C++/Rust, enabling deployment on the web for client and server applications.
• OCR (Optical Character Recognition): The process of converting an image of text into machine-readable text.

Canvas with Python Execution: A Deep Dive

Introduction

The video explores the new Canvas feature in ChatGPT, focusing on its Python execution capabilities and comparing it to existing tools like Code Interpreter and Claude's Artifacts. The speaker shares his initial excitement and subsequent confusion after testing Canvas, particularly regarding its utility compared to Code Interpreter.

Canvas Overview

• Availability: Canvas is now available to all ChatGPT users, not just Plus subscribers.
• Python Support: Canvas allows users to write and execute Python code directly within the chat interface. Code blocks are highlighted and editable.
• Custom GPT Integration: Canvas can be integrated into custom GPTs.

Comparison with Code Interpreter

• Workflow: In Canvas, users co-write code with ChatGPT, whereas Code Interpreter acts as a "worker" that generates and executes code independently.
• Result Handling: Code Interpreter returns the results of code execution directly to the chat, allowing for iterative analysis. Canvas, however, does not automatically feed results back into the chat, requiring manual copy-pasting.
• Environment: Canvas runs Python code in the browser using Pyodide, while Code Interpreter uses a server-side environment.
• File Access: Code Interpreter can access files uploaded to the chat, while Canvas cannot directly access these files.
• Internet Access: Code Interpreter is isolated and cannot make network requests, while Canvas can make network requests but has limitations on using complex libraries.

Technical Details of Canvas's Python Environment

• Pyodide: Canvas utilizes Pyodide, a Python distribution compiled to WebAssembly, to run Python code in the browser.
• Limitations: The Pyodide environment in Canvas has limitations, including the inability to use certain libraries that rely on processes.
• Environment Information: The speaker demonstrates how to retrieve environment information within Canvas, revealing that it runs in a CPython environment using Emscripten.

File Handling and Data Analysis

• File Uploads: The speaker attempts to analyze a CSV file using Canvas but encounters an error because Canvas cannot access the uploaded file.
• Code Interpreter's Superiority: Code Interpreter successfully analyzes the same CSV file, calculates the median, and generates a chart, highlighting its superiority for data analysis tasks.
• Claude's Artifacts: The speaker also tests Claude's Artifacts with the same CSV file but experiences issues with data loading and visualization.

Network Requests and Library Usage

• Image Loading: Canvas can load images from URLs using network requests.
• Library Limitations: The speaker attempts to extract text from an image using the pytesseract library in Canvas but fails because the environment does not support processes.
• Code Interpreter's Success: Code Interpreter successfully extracts text from the same image using pytesseract after the image is uploaded to the chat.
• Claude's Limitations: Claude's Artifacts encounter content security policy issues when attempting to load images from external URLs and fail to load libraries.

Limitations of Claude

• Content Security Policy: Claude's Artifacts are restricted by content security policies that prevent them from accessing external URLs unless they are whitelisted.
• Library Loading Issues: Claude's Artifacts also encounter issues when attempting to load complex libraries.

Conclusion

The speaker concludes that while Canvas is faster and more shareable than Code Interpreter, it is currently less useful for real-world tasks due to its limitations in file access, library usage, and result handling. Code Interpreter remains the better tool for data analysis, despite its lack of network access and shareability. Claude's Artifacts are deemed less effective for these tasks. The speaker questions the target audience for Canvas, suggesting that its primary use case may be for learning Python with small, limited scripts.

Notable Quotes

• "This is what's different about Canvas: you co-write with it, it's a coworker, while this [Code Interpreter] is just a worker, which sometimes you actually would prefer."
• "For real work, the Code Interpreter that's by now more than a year old is a better tool."
• "Who is it for? This is a question I asked at the start of the article, and after this exploration, I still am wondering because there are better tools for this, one of which is already in ChatGPT, and it's Code Interpreter."