After talking to sources at Google, I’ve come to the conclusion that these bad Gemini responses slipped through testing because everyone felt rushed to ship. An illustrative example: the photo generation in the Gemini app is not actually powered by the Gemini model. It’s an older, text-to-photo model that was tacked onto the Gemini user-facing experience to get the feature out the door faster.

I’m told that there’s also a lack of alignment between Demis Hassabis’s research team building the underlying models and Prabhakar Raghavan’s search organization that’s putting them into user-facing products. Perhaps the coming “structural changes” Pichai hinted at in his memo to employees last week will address that.