Creating General User Models from Computer Use
Source: https://generalusermodels.github.io/ Author: GUM Research Date: 2025-09-16
Summary
Research on “General User Models” (GUM) — AI systems that watch how users interact with their computers and build persistent models of user preferences, workflows, and goals. The vision: an AI that understands you by watching, not asking.
Key Claims
- GUM architecture: watches screen content, mouse/keyboard events, application usage patterns — builds behavioral model.
- From observation to preference: the system infers preferences (likes terminal over GUI, prefers certain coding patterns) from behavior, not explicit input.
- Personalization use: the user model drives personalized AI assistance — same query gets different responses based on user model.
- Privacy concern: watching everything is powerful but invasive. GUM requires deep trust and transparent data handling.
- The goal: an AI assistant that “knows you” without constant explicit instruction.
Concepts
- Agent Memory — user model as persistent agent memory
- Context Engineering — user model enriches context for every interaction