I have the feeling that the author has no idea how both models work, otherwise he would give more precise reasoning as why one or thw other performed better.

1. both models do not understand the text, they just transform it to an abstract representation, which allows them to predict the next token based on token frequencies in the training set.

2. Chat GPT has to be more natural, because it was finetuned in a reinforcement learning setup where humans give feedback about the responses made by the model in terms of grammer, expressions.

3. as stated by Robert Mazurowski, google cannot afford to used the full blown lambda model. I am pretty sure, if you look up the evaluation results of the full size lambda model it will perform as good as the GPT

4. I would argue that shorter, more precise answerds like presented from BARD are more helpful information gaining than naturally written texts by Chat GPT, which you have to read twice to get the valuable information.