A both practical and philosophical discussion on LLM neural networks.

[hiccup]

(as many others) I am interested in the usefulness and development/progress of LLM's (large language models)
Especially in regards to it providing useful answers to questions that MusicBee users may have.

To start off the discussion: I have been testing ChatGPT (3.5 I think) to see if it could provide useful and working formulas for virtual tags.
And at this very moment it seems to be terrible and pretty much useless at it.
It is constructing incorrect syntax, and invents functions that are not even available for MusicBee.
It is nice that it keeps apologizing when confronted with that, but that doesn't improve on the results.

So for now carbon-based life forms such as you and me seem to have some purpose here.

But that will change in the near future without much doubt.
Today I read that Microsoft/OpenAI is going to build a new datacentre specifically for LLM that is estimated going to cost 100 billion dollars.
To put that in some perspective, that's not extremely far away from the annual budget for the complete military of the USA.
Interesting times indeed.

It is targeted to become operational in 2028.
Which gives me little doubt that our nice little forum will be extinct by then, and most of us will have retired from it.

So this topic is intended to be both practical (how useful is LLM at this moment, and what are your experiences with it), but also a bit philosophical.
(as in: 'the end is near')

[frankz]

AI tells me this is hiccup enjoying a Musicbee session with his assistant in the library of his estate.  I choose to believe in AI.

[hiccup]

Wow, it's as if you stole a photo from my photo album.
The only giveaway that this is not real, is the brand of computer I am using here.
So LLM's are still hallucinating sometimes.

[frankz]

My serious answer is that I have used it sparingly for things like TV show recommendations (I liked these 5 shows, what else will I like?), Google Sheets formulas, quick Python scripts to import and manipulate spreadsheet data, etc.  It is at once impressive and mildly frustrating. It's great at pointing you in the right direction.  You mostly have to get it across the finish line to a workable solution yourself, though.  It is a useful tool.

I think it's going to be much like any other direction-changing innovation.  The rise, fall, and rise of the Internet is a good example.  A quick burst of extreme excitement where everyone jumps in to build something around it and it seems like life will never be the same again because of it.  Then a crash where many, many participants fail and it seems like it was all a fad.  Then the real upswing starts and the participants that survived round 1 suck up all the talent from round 1 and really make it functional and essential to everyday life. In the end, life is improved but you're not living in an entirely different way that was inconceivable before the innovation arrived.

Because we don't know what that third phase will look like, we're probably excited about it for all the wrong reasons and fear it for all the wrong reasons.

[phred]

Something is very wrong with that graphic. While it may be hiccup at his desk (I've never met him so I can't tell for certain) that is NOT his assistant. I know that woman and she's making breakfast for me this very minute. Her name is Mrs. Phred.

[hiccup]

Because we don't know what that third phase will look like, we're probably excited about it for all the wrong reasons and fear it for all the wrong reasons.

I am not sure I agree with 'all', but I could go along with 'a lot'.


I also find it useful for applications such as Excel.
It often provides good solutions, explained in clear steps.

I am sure that has to do with the user base and the amount of info that is available on the internet about specific software.
That is probably why ChatGPT is doing so bad in offering formulas for MusicBee.
Or maybe it hasn't indexed the forum very well yet.

There are also some serious competitors to it on the horizon, such as Claude and Grok.
Maybe they will appreciate MusicBee more than ChatGPT ;-)

But that it is going to change things, also on our forum is for sure.
At the moment it is still kind of fun to think of 'clever' formulas or regexes.
But when these LLM's become good and better at it, I probably won't bother with it so much anymore.
The same will go for more regular questions.
And it would relieve some experienced users from telling new and lazy users to "Search the forum first. That has been asked and answered many times before".
Or maybe they will be replying with something like: "You should first ask Claude before posting here."

But it would be fun if a LLM also gets tired of that, and starts replying in the same manner ;-)
(I think Elon Musk's LLM has a sarcasm mode ;-)

[hiccup]

Something is very wrong with that graphic. While it may be hiccup at his desk (I've never met him so I can't tell for certain) that is NOT his assistant. I know that woman and she's making breakfast for me this very minute. Her name is Mrs. Phred.

So it's not only LLM's that are doing the hallucinating.
I'm sure Frankz will provide us with a reliable image of your living situation.

[frankz]

Something is very wrong with that graphic. While it may be hiccup at his desk (I've never met him so I can't tell for certain) that is NOT his assistant. I know that woman and she's making breakfast for me this very minute. Her name is Mrs. Phred.

So it's not only LLM's that are doing the hallucinating.
I'm sure Frankz will provide us with a reliable image of your living situation.

I was able to locate an image of phred training the next generation of forum moderators at the forum HQ if that helps.

[hiccup]

It's 3.6.8856P, ma'am.

[phred]

YIKES!!

I'll bet we see an uptick in forum members volunteering to become a moderator.

@frankz- what site or engine are you using for those AI-generated graphics?

[frankz]

I used Ideogram.  I read that it's one of the only ones that can put actual text into the photo, although it's not always great at that either.  Took several passes to get it right.  I liked this version better - the girl looks more upset / annoyed by what she's reading and the guy watching her is funnier - but it got the text on the shirt wrong.

I originally had it as the trainee watching you, but I couldn't get it to make a nice picture of your character unless you were the one watching.  Weird.

[phred]

Even with the misspelling, I like the second one better also. But then again, who is looking at the letters??

Thanks. I've played with ChatGPT and have been looking for something that does a halfway decent job at images/graphics. I can't remember what I tried previously, but it would've had to get better just to suck.

I will try Ideogram.

[phred]

The life of a MusicBee moderator...




And when you said it's likely to make mistakes, take a look at this one with four arms.

[frankz]

Nice.

Either AI isn't replacing us any time soon or it knows we're getting additional arms and fingers before we do and we're totally screwed.

[phred]

Either AI isn't replacing us any time soon or it knows we're getting additional arms and fingers before we do and we're totally screwed.

Perhaps we'll be required to use two keyboards simultaneously.