Right, and we have to be careful not to misinterpret such declarations by LLMs as if they're the real thing. I will never forget the first wave of Chat GPT reporting from the likes of Kevin Roose, who took its conversations completely at face value, not realizing it was reacting to his cues, and frankly playing down to the quality of his input.

But bring to machine interfaces are an entirely different thing altogether and they test it with controls and training unknown examples and then proceeding to prediction based off of training.