An AI That Learns to Use a Computer
This is a piece I originally posted to my Instagram story in March 2026.
markov-ai has released the computer-use-large dataset. It’s a massive dataset made up of 48,478 screen-recording videos, totaling roughly 12,300 hours. It was designed to train and evaluate computer-use agents—models that interact with desktop software through GUI actions like clicking, typing, and scrolling. In short, it’s a large-scale collection of screen recordings built so that an AI can learn to operate computer software the way a real person does. It contains footage of actual tasks performed in six widely used professional applications: AutoCAD, Blender, Excel, Photoshop, Salesforce, and VS Code.
What caught my eye here is that the dataset also includes recordings of people using Photoshop, AutoCAD, and Blender.
These days I keep telling people I don’t really code anymore. The agents do all of it for me. The software-development world is in an uproar right now. Caught between the threat of being replaced by agents and the sheer convenience of having something do all your work for you, it’s become a chaotic scene of restless thoughts and anxious glances.
When I talk to people in the art world, on the other hand, most of them are still amazed by Nano Banana—let alone agents.
Okdalto
한국어
Comments