This doesn't engage with the article in the precise way that transformer-based improvements and fundamental research that way are not being incorporated into models, at Microsoft AI, which is concerning.