John Schulman

I was also happy to see that this paper, and an earlier one by Hase et al. also on counterfactual simulatability (but more focused on training) used tinker for their fine-tuning experiments https://x.com/tinkerapi/status/2095677662291472494