Trying out the new AI models

Community
1

Hello :wave:

I make this topic for share experiences with the new AI models using on Discourse.

I set up a few days ago the grok-2-1212 for topic summaries. It works really well. The language detection and quality is well enough. I tried it also with AI bot but it most of time failed I guess it can’t handle the tools well yet.

I also tried the Gemini Flash 2 for AI bot. It works fast and add great answers however sometimes it seems it can’t handle well the tools and broke the answer. Sometimes it is a simple markdowns formatting issue, sometimes it can’t search. On my forum most of time says didn’t find anything on the forum but I know more topic about that subject…

4 Likes
2

For the search problem, do you think it could be related to the AI not fully indexing the forum content or there might be a mismatch in query understanding?

2 Likes
3

If have no idea. Most of time it searches for nothing “” or failed with timeout… but sometimes do the search correctly and linked the correct topics. It will be good I think but it’s strongly experimental yet.

1 Like
4

Have you tried xml tools? I found that on grok they work quite well

3 Likes
5

Thanks, I tried it now. Yeah looks better, the problem is mostly happening now when I create a new conversation. It starts in English something like: I am searching for…in sitename… and stop replying. Sometimes it continue the answer after the English I am searching… sentence on the correct Hungarian language and add good answer. However if I reply the grok response after that in the conversation it will works great.

2 Likes
6

This is really interesting, I kind of want to allow “grounding” examples as an option for personas, it could totally solve this

2 Likes
7

Bingo! But do you think it’ll slow things down, especially with a lot of data? Could it affect response times for AI queries or search results, or is it all good?

1 Like
8

The big problem examples have are “contamination”

The model learns shape, but also can mistakenly think a user said something they did not.

Ideally carefully crafting system messages can do the trick , that would be my first resort

Examples in a system message can lead to less leakage cause it can be clearer to a model it is just an example

A minimal thing I would recommend Don, is writing your system message in Hungarian, it could help

Maybe even try giving an xml tool example or two in the system message?

4 Likes
9

This genuinely sounds good, thanks for sharing : )

10

I tried it but same result with grok-2-1212 then I switched it to grok-beta and it works perfect but it works with English system message too…

4 Likes