I wasn't thinking about this strictly as a search engine pov given what I thought assistant was for. That said one thing I currently use assistant for is something like
How do I change the access vlan on a juniper ex4300?
This seems in line with assistant as better search.
Well, I also often need to get the logo of a program to put in our internal app store. I don't know if AI can help, but the app stores built in Google image search almost never matches any icon or logo unless it's at the level of say Microsoft Office "known". It seems like this would need a lot of what the vision AI stuff does.
However, another thing I often search for is images but not specific ones. I might need a free image of a train station to illustrate a proposal. Here generative AI would potentially be an easier search solution than trying to find an artist rendering of the same thing on unsplash or the right license online. Idk but it seems like a plausibly equivalent amount of work for the computer.
Another thing I search for is laser cut out files. While I would guess we're some distance from DALLE generation of 3d multi layer cuttable images, I also bet it might be able to make png or svg of a snowflake or something like that.
Is this strictly search? No, but neither was image search in general if we define search as text based only.
The other thing is - it seems more and more apparent to me that AI is going to be needed to improve search much more, that's why Google and Bing are making part of their search results. It's also why at least some AI chatbots have played with search integrated into the chatbot - that blending is a major boost to the value of just an llm with a knowledge date cutoff.
While I get the "feature creep" concern, I think AI image generation and recognition blended into assistant (and eventually maybe making assistant either a lens or a section of search kind of like Bing does) is going to only improve search.
Feature creep to me would be like making a video game or an app store, not integrating more ways of classifying data and getting results, even if part of that is "making it up" a la DALLE when it fits the actual search request. Of course you would label generated vs found for the user.