Raj Thaker

Engineer at Bland · San Francisco

I'm an engineer at Bland, where I built our in-house media server and our omnichannel product. Before that I was the first hire at Open Ledger, an embedded accounting API that Collective acquired, and co-founder and CTO of Boondoggle AI.

On the side I build Langbot and do independent research on how language models simulate people.

Research

What does a language model need to know about a person to predict what they do?

I gave six models from four developers the personality scores of 1,190 people from the Human Connectome Project and asked them to predict each person's questionnaire answers and measured task performance. The models came close to a trained model on questionnaires but barely predicted behavior. Persona simulations built from self-report reproduce how people describe themselves, not what they do.

Questionnaires Measured behavior Trained model Pearson r between prediction and the real person. Opus 5 ran on 600 people, the rest on 1,190.

Paper (PDF)

Work

Projects