AI General Thread

[the-pi-guy]

Gemma 4 -26b version, fails your pet rock challenge. It does understand the 1 in 5 Chinese odds.

Generating text, it seems fine. Not sure how else to test it. 

[Legend]

So better than the other local models or not?

[the-pi-guy]

So better than the other local models or not?

I'd say it's definitely better at being useful. 

Asked it to generate some C# code, did a good job. 

Design a game, that was pretty good. 

I asked it to write some stories and something feels off to me, and I'm not sure what it is. It just feels a little less natural than the other models I've used for some reason. 



I've also tried some uncensored models that people made. All of them have been different kinds of bad. 
A lot of them have the issue of using non-words. Weird stuff like putting numbers in the middle of words. 

[Legend]

I'd say it's definitely better at being useful.

Asked it to generate some C# code, did a good job.

Design a game, that was pretty good.

I asked it to write some stories and something feels off to me, and I'm not sure what it is. It just feels a little less natural than the other models I've used for some reason.



I've also tried some uncensored models that people made. All of them have been different kinds of bad.
A lot of them have the issue of using non-words. Weird stuff like putting numbers in the middle of words.

Oh that's really weird. You think it's glitching or just actually got screwed up with whatever fine tuning they did?

[kitler53]

so i just don't know what to say about my companies AI tool.   

I tried to use it yesterday to solve the hangman question.   I also used it to write some SQL queries.   

My hangman question has a little flag on it that claims it took "2 seconds" to process the request but in real time it took over 10 minutes to get a response to my question.   

one of my questions about a SQL query took "41 seconds" but it was over 20 minutes to get an answer.  while I was waiting for that answer I did a google search.   Google's AI gave me a response in way less than 41 seconds.   At a glance I thought it was pretty good but then i looked at the source and it was almost straight up plagiarism of a MS help page.   when i finally got an answer from my companies AI it was also a straight up plagiarism of the same MS help page.

the overall theme though is by the time my companies AI gives me an answer I have already finish with my work and moved onto the next thing.  i just don't understand why it is sooooo slow.   google's public (and free) AI is just soo much faster and frankly gives better answers.   ...and i appreciate that it provides sources.

[the-pi-guy]

so i just don't know what to say about my companies AI tool. 

I tried to use it yesterday to solve the hangman question.  I also used it to write some SQL queries. 

My hangman question has a little flag on it that claims it took "2 seconds" to process the request but in real time it took over 10 minutes to get a response to my question. 

one of my questions about a SQL query took "41 seconds" but it was over 20 minutes to get an answer.  while I was waiting for that answer I did a google search.  Google's AI gave me a response in way less than 41 seconds.  At a glance I thought it was pretty good but then i looked at the source and it was almost straight up plagiarism of a MS help page.  when i finally got an answer from my companies AI it was also a straight up plagiarism of the same MS help page.

the overall theme though is by the time my companies AI gives me an answer I have already finish with my work and moved onto the next thing.  i just don't understand why it is sooooo slow.  google's public (and free) AI is just soo much faster and frankly gives better answers.  ...and i appreciate that it provides sources.

That makes little sense to me. What service were you using when you got these long response times?

I could understand if your company has a single queue, and you're waiting for other employees (without even being aware of it). But I haven't seen anything like that.  

Although Copilot goes down sometimes for me...  

[kitler53]

That makes little sense to me. What service were you using when you got these long response times?

I could understand if your company has a single queue, and you're waiting for other employees (without even being aware of it). But I haven't seen anything like that. 

Although Copilot goes down sometimes for me... 

"relavance AI" but as I mentioned before it seems like that is just a "business security layer".   The actual AI is:
- Gemini 2.5 Pro
- Claude Haiku 4.5, Sonnet 4.6, or Opus 4.6
- OpenAI GPT 5.2, 5 mini, 4.1

[the-pi-guy]

I knew about Relevance AI, I was just wondering if you had any more specifics. 

Asking Copilot about Relevance AI  (Asking AI to learn about AI, sure is a funny thing). 

And it also thinks you're getting rate limited by the company, probably per model.

We don't have any official system at work yet. Not sure if that'll change.
But a lot of use Copilot in the Edge Browser.

Sometimes I use Duck AI:
https://duck.ai/

I think two of my coworkers (one of which is my boss) uses the actual claude.ai on occasion. 

(I'm sure it's against your company policy to use either of these.)

[kitler53]

yeah, policy wise i'm only suppose to use the company AI tool.  

...but if my question doesn't contain any company specific information i ignore the policy.  my SQL question was pretty generic.   I have a table the represents a heirarchy.   So there are basically 3 columns:
- ID
- ParentID
- Name

So the topmost items has a Null ParentID and all of the other records have a parent.   I needed to calculate the full path:
- <<Name Top>>\<<Name Middle>>\<<Name Bottom>>

there is a recursive SQL query using "Union All" that i didn't know existed but it was very well documented use case and nothing about what i'm doing is IP.  

[the-pi-guy]

I don't know if this is actually the case, but I almost feel like a lot of people look at AI as if it's this box, and they're trying to get the box to be capable of everything they want it to be. 

But I've been viewing it under a lens that each LLM is it's own personality. 

If you try 10 different LLMs, they might be generally capable of most of the same things. One might be better at X and worse at Y. But you also find that they generally like different phrasings, they use different names when you ask them to come up with names.

There's different model temperament, even if you're asking them to do the same things.