The real treasure was the Markov chain samples we made along the way

• 2x faster on GPU, due to parallelizing better the tree sampling step.
• Uses less memory, now can do $n=100,000$ with $10,000$ trees on a V100. This was mostly an excessively large batch size for counting datapoints per leaf.
• The Metropolis-Hastings ratio is saved only for the proposed move.
• The grow and prune moves are merged into one object.