Testing a new feature: AI responses (Page 1/2)
Cliff Pennock AUG 31, 06:05 PM
Ok, this is very, very, very experimental. Also, this new feature only works in the TD&Q section.

Whenever you post a new topic in the TD&Q section, the system automatically adds an AI generated answer. It will only do this once. So you can respond to it, but it will not answer back.

Also, be aware of the following points:

  • The answers given by the AI Bot can be incomplete, incorrect or even completely fabricated. Like I said, this is very experimental
  • Although the AI "understands" that all questions are related to the Pontiac Fiero, it might give a very generic answer which might not apply to the Fiero at all
  • The AI bot can not "see" images. So if you ask to identify a part on an image. It will probably tell you that it needs more information (which you can't give since it will only respond once in a thread), or give some weird answer
  • Did I mention that the answers given can be incomplete, incorrect or even completely fabricated?

I understand that everybody is now very curious to see it working and wants to post a question in TD&Q. And since I understand, I'll allow it for a while. I will however delete those threads after a while because I don't want TD&Q cluttered with test questions, causing the real questions to get buried.

Also, please be aware that this is not a free service for me. Every question answered by the AI Bot costs me real money. It isn't much, we're talking parts of a dollar cent per question, but it can add up really quickly if the questions or answers get elaborate.
Raydar SEP 02, 11:36 AM
I just noticed this, this morning. (Surprise!)
But all the bot did was quote my original post, verbatim, with additional spacing between phrases. (I did edit my original post, just a bit, for clarity. I'm guessing the bot wont't pick that up. But it was only a minor detail/clarification, so no big.)

At least it didn't seem as "rambling" as my original post did.

If it was supposed to provide a possible solution, it seems I "stumped it".
Cliff Pennock SEP 02, 06:48 PM
Yes, like I said, it's very, very experimental. If it turns out AI doesn't really offer helpful answer, I can just as easy disable it again.

I had test it with a few existing threads in TD&Q and it gave good, sensible answers. But that of course says nothing. It could very well be that those were the exceptions.
sdgdf SEP 02, 07:37 PM
If you could feed the entire PFF archive to ChatGPT this could be really cool. Does the response have to be in a link though?
Raydar SEP 02, 09:44 PM

quote
Originally posted by sdgdf:

If you could feed the entire PFF archive to ChatGPT this could be really cool.
...



Heh... I didn't even think about that.
Cliff Pennock SEP 03, 05:46 AM

quote
Originally posted by sdgdf:

If you could feed the entire PFF archive to ChatGPT this could be really cool. Does the response have to be in a link though?



That is my plan eventually. Train it with all the threads on PFF. And I made the response inside a link because it sometimes gives very long-winded answers.
Cliff Pennock SEP 03, 05:48 AM
Oh I see that it's just repeating the question now. I think that's a mistake on my side. Let me have a look at that.
Raydar SEP 03, 05:03 PM
Thanks Cliff!
I'm watching. Maybe this will eliminate all the "did you do a search?" responses. Since the AI will have already done the searching.

Now... if you can just get it to filter out the wrong answers that people have given, over the years. (Am I failing to understand? AI is uncharted territory for me.
Although some people might suggest that even regular "I" is a stretch, in my case. )

[This message has been edited by Raydar (edited 09-03-2023).]

Cliff Pennock SEP 03, 07:15 PM
Well, that is exactly the problem... I can train it with all data on PFF, but it will "learn" the wrong answers too.
Cliff Pennock SEP 03, 07:21 PM
But right now this is just an experiment. I changed the model to GPT-4 now to see if it gives better answers. But GPT-4 is almost 20x more expensive than GPT-3.5. So I'm keeping a close eye on the costs. If it gets out of hand, I'll have no choice than to revert back to GPT-3.5. And if it turns out GPT-3.5 just continuously gives wrong answers, I'll unfortunately have to disable it again.

The reason I wanted this, is because often people are stuck on a project and need an answer fast. The PFF AI Bot will give an immediate answer and although it might not be 100% correct, it can give the asker some ideas how to continue.