Building General User Models
from Computer Use

This paper tries to turn screenshots, visible UI text, message context, and app switching into an editable user model built from confidence-weighted natural-language propositions. The page below focuses on the moving parts: how raw traces become claims, how retrieval and revision stabilize them, and where the evidence narrows relative to the headline.

persistent proposition store retrieve + revise loop cross-application assistants privacy and poisoning tension

Study slice

18 users

Main quantitative evaluation described in the episode.

Behavioral window

200 emails / user

Gmail Primary only, not the full desktop the headline evokes.

Target memory unit

30 propositions

Confidence-weighted natural-language claims about goals, knowledge, preferences, and state.

Deployment tension

70B + 72B VL on H100s

The local/private story is compelling; the reported model stack is still heavyweight.

Lineage Overlay

Rich 1979 Lumiere 1998 Recommenders 2005 Generative Agents 2023 MemGPT 2023

Transcript arXiv IDs

evidence capture memory structure assistant action risk / overreach

Trace Breadth to Claim Legibility

1. Desktop Span

2. Proposition Store

3. Evaluation Slice

4. Risk Surface

References

Paper trail

Related episodes

Building General User Modelsfrom Computer Use

Trace Breadth to Claim Legibility

Cross-application traces become an editable memory substrate

What kinds of claims does the store infer, and from what evidence?

What the episode says the paper measured, and where the evidence still thins out

Utility and creepiness live on the same surface

References

Building General User Models
from Computer Use