Perhaps I missed it in the article, but these hands were played online, not live, right?

Also, I think it's pretty interesting that it wasn't until the 30k hand mark that the AI started to become consistently profitable, and it wasn't until 80k hands played that this profitably was statistically greater than 0. I think this suggests that the AI's ability to win isn't from some sort fundamentally better strategy against all opppnents, rather, its improved ability to learn from opponents and exploit them over a large sample size. This bot can't just sit down at a random online table and start crushing, which is a good thing.