>>8651736
>>8651849
On twitter he just vaguely states A/B 3/4 have different datasets and approaches. It's likely just one of those try them and see things where it's hard to say which one is definitively better.