Recently, it started returning even more verbose answers. The absolute bullshit research paper that Google Gemini gives you is what turned me away from using it there. Now, chatGPT also seems to go for more verbose filler rather than actually information. It is not as bad as Gemini, but I did notice.

It makes me wonder if people think the results are more credible with verbose reports like that. Even if it actually obfuscates the information you asked it to track down to begin with.

I do like how you worded it as a believable range, rather than an accurate one. One of the things that makes me hesitant to use deep research for anything but low impact non-critical stuff is exactly that. Some answers are easier to verify than others, but the way the sources are presented, it isn't always easy to verify the answers.

Another aspect is that my own skills in researching things are pretty good, if I may so myself. I don't want them to atrophy, which easily happens with lazy use of LLMs where they do all the hard work.

A final consideration came from me experimenting with MCPs and my own attempts at creating a deep research I could use with any model. No matter the approach I tried, it is extremely heavy on resources and will burn through tokens like no other.

Economically, it just doesn't make sense for me to run against APIs. Which in my mind means it is heavily subsidized by openAI as a sort of loss-leader. Something I don't want to depend on just to find myself facing a price hike in the future.

It even shows me a push notification at the top of my screen when the search task has finished.

But I’ve found that no matter the error - even if I disconnect from the internet entirely - I eventually get a push notification and opening up the thread a while later shows me the full response. (disclaimer: N=1)

One of the complications of your average query taking at least some number of seconds to complete - that is, long enough for the user to do something else while waiting.