Transformer-based model with 150 M params, trained on 15 B purchases, and showed 10× lift in conversion,
where the line is between a "behavioral foundation model" and an LLM with behavioral pretraining. Is it the data, architecture, or something else?
The personalization examples are compelling — especially how one change in a cart shifts the recommendation path. Have you quantified how early in a sequence you can steer outcomes reliably? That could be a huge insight for anyone designing nudges or interventions.