Recreating the Apollo AI adoption rate chart with GPT-5, Python and Pyodide (simonwillison.net)

Simon Willison used GPT-5, Python, and Pyodide to successfully recreate Apollo Global Management economist Dr. Torsten Sløk’s AI adoption rate chart, which showed a recent slowdown in AI adoption among large U.S. companies (250+ employees). The data originated from a biweekly US Census Bureau survey tracking whether firms used AI tools—like machine learning, NLP, virtual agents, or voice recognition—in the prior two weeks. Despite initial discrepancies caused by Apollo’s use of a six-survey moving average, GPT-5’s code interpreter handled complex data parsing, applied the correct rolling average, and generated a nearly identical matplotlib chart, validating the Census figures and revealing a subtle recent uptick in adoption. This exercise highlights GPT-5’s prowess in navigating challenging datasets and converting ambiguous descriptions plus raw spreadsheets into polished, actionable visualizations using pandas and matplotlib. Furthermore, Willison demonstrated the feasibility of rendering the same charts entirely client-side in-browser via Pyodide, a Python runtime for browsers, overcoming package and compatibility hurdles. This approach enables dynamic, interactive data analysis and visualization within web interfaces without server-side dependencies—an important step for democratizing AI and data science workflows. For the AI/ML community, these advancements illustrate how large language models can streamline challenging data discovery and visualization tasks, while tools like Pyodide open new possibilities for lightweight, browser-native data exploration. The blend of GPT-5’s search and coding capabilities with Python’s data ecosystem marks a practical evolution toward more accessible, reproducible AI-driven analytics.