Google’s latest artificial intelligence model, Gemini 3, represents a significant step beyond conversational chatbots. The model functions as an agent that can understand complex goals and execute multi-step tasks on a computer. Ethan Mollick writes for his Substack newsletter “One Useful Thing” about his tests of the new system. He concludes that AI is evolving from a tool that describes concepts to one that can build and run them.
Mollick illustrates this progress with an example. Three years ago, an AI could write a poem about a fantasy concept. Gemini 3, given the same prompt, created a fully playable computer game based on the idea. This is possible because the model can write and execute code to perform tasks like accessing files, searching the web, or building websites based on simple English commands. In one test, Mollick asked Gemini 3 to analyze all his past articles, identify his predictions, research their accuracy online, and build a website to display the findings. The AI planned and executed the project, checking in for approval at key stages, which Mollick compared to managing a human teammate.
In another experiment, the model analyzed a folder of messy research data and produced a 14-page academic paper with original hypotheses and statistical analysis.
Mollick notes that while the AI still requires human oversight, the user’s role is shifting from correcting errors to directing the work of a capable digital assistant.
