[rockdoe]

It's explained in the Wired article that is linked that this was indeed the cause.

The independent reproduction that the IEEE article talks about also tested the method from the leaked paper from Chatterjee and concluded that his method indeed beats CT in almost all cases at a fraction of the computational cost.

The response from G&M to the independent reproduction emphasizes their method was used for production hardware at Google, was in a peer reviewed paper, and that their GitHub repo had a lot of forks and stars (I'm not making this up). Near the end it then raises several technical objections.

[warbaker]

It doesn't look like they didn't train CT to convergence, so idk why anyone would care about those results.

What is even the point of benchmarking against a learning based method of you don't let it finish learning? Quality of output matters way more than compute cost.

[warbaker]

EDIT (parity error): "didn't train" -> "trained"