I have been using the regular RA for a while and when the experimental one was released to Ultimate Subscribers I decided to switch over for a while and compare.
OBSERVATIONS:
- As another user previously reported, it does not seem to work with custom prompts. Hopefully this will get sorted at some point.
- The librarian limits seem to be much higher which is awesome!
- Also awesome is that it seems to use integrations like Wolfram Alpha, Python Sandboxes, etc., more often.
- It is DEFINITELY MORE COSTLY! More than the 1-star difference would lead you to believe, but I'm not complaining as I have found it to be worth the cost increase.
- The summary shows that the regular RA gets a 5/5 Privacy Score while the ERA gets a 4/5. This leads me to my request:
REQUEST/SUGGESTION:
The last observation above on the privacy score has got me thinking. I know that some models are queried directly reducing privacy so it would follow that the ERA uses at least one or more of those models whereas the regular RA does not. However, I can't really seem to find any solid documentation on the differences between the two. I have observed things (above), but would like to know what it is doing different. Is it just using higher models with higher librarian limits and increased searches? Does it use different criteria to select which model to utilize for various functions? Why does it seemingly cost almost twice as much? Most importantly why is it not as private? If queries are being directly submitted to providers, which providers are these as they have different privacy practices and I don't think that just showing a lower privacy score works with these meta models as privacy can be compromised in different ways than if you were just using a model directly.
My concerns about privacy has me thinking of moving back to the regular RA. I think that if you just added some information as to how these models differ and most importantly why the ERA is less private, more people would be inclined to use it. Proper documentation comparing the RA to the ERA would also help the user to decide when each may be more appropriate. Some things may get answered just as well with the regular RA, but for others the ERA definitely provides better results. I don't really know when to use the ERA due to this lack of documentation. While for me personally the cost isn't a big concern, (I favor quality above all else but privacy), but I'm sure it would be for others. THE PRIVACY FACTOR DOES CONCERN ME SUBSTANTIALLY! Especially considering that privacy is a core tenant of KAGI, transparency about it in your best performing AI Assistant should be a no-brainer.
