• TalkBass has been independent since 1998. Add your voice.
    Create a free account to reply to discussions, view embedded media, and browse with fewer display ads.
    Join freeLog in
    Want zero display ads or expanded classifieds tools? Compare plans.

Josh has a few thoughts about AI…

One of my tests is to ask about the saxophone solo for The (English) Beat. It's one of the few without any saxophone at all. Last time I tried it, about three months ago, the chat bots waxed lyrically and confidently about the solo.
That was a bit garbled. I'm laid up with a broken leg and pain meds. Hopefully, people can make some sense of it.

I'd love to see confidence scores for AI responses. When I worked in machine learning we were big on confidence scores or even refusing to return a response where confidence was low. Although I wasn't working on LLMs as this was before we had sufficient resources to efficiently implement such a thing.
 
  • Like
Reactions: agedhorse
AI or people learning from bad information end up regurgitating bad results based on the GIGO theory.
Our issue was frequently finding good quality data, or ways to take mediocre data and somehow filter it to improve it.

The only time we tried to use internet data was when bored and considered making an arse/elbow discriminator (UK expression: doesn't know his arse from his elbow) but we discovered that getting training material without losing our jobs was going to be challenging. So that project only lasted an afternoon.
 
  • Like
Reactions: agedhorse
Our issue was frequently finding good quality data, or ways to take mediocre data and somehow filter it to improve it.

The only time we tried to use internet data was when bored and considered making an arse/elbow discriminator (UK expression: doesn't know his arse from his elbow) but we discovered that getting training material without losing our jobs was going to be challenging. So that project only lasted an afternoon.
This is exactly what's driving the data grab scenario.
 
AI or people learning from bad information end up regurgitating bad results based on the GIGO theory.
And AI scrapes data a lot of which has been generated by AI anyway.
So in a sense it's feeding on its own excrement.
Thus the models get gradually worse over time
The whole system will crash as there is no route to profitability given the amount of debt taken on by AI companies
 
  • Like
Reactions: SDC1-ClickClack
And AI scrapes data a lot of which has been generated by AI anyway.
So in a sense it's feeding on its own excrement.
Thus the models get gradually worse over time
The whole system will crash as there is no route to profitability given the amount of debt taken on by AI companies
Correct, training on your own bad data that's then used to train itself to improve accuracy is just like eating your own babies...
 
  • Like
Reactions: Drredplate

And that is your summary?

[murder = (more gooder) than (profiteering / morality)^2] if [(culprit is allegedly moderately attractive to females) and (inaccurate media reframing of Robin Hood narrative) and (gets all the likes)] ?

Hmm.

On second thought, you may have just convinced me. Most humans' thought processes are worthless and should be removed from any decision of consequence.

Have any major firms reached out with a contract offer yet? You seem to be making the same mission statement as 'Big AI' is, only you are using different words.

Don't worry, the offers will come. True talent is hard to waste
 
The heady joys of browsing the WWW were squelched when the works of generous contributors were guarded behind pay walls. And targeted ads.

The fourth little pig built his house using Reddit posts. It fell down before the wolf arrived.

1777075754733.png


Imagine chatting with an ad generating machine disguised as an Oracle.
 
It's pretty interesting to decouple the mainstream PR machines from reality.

ChatGPT is full of hot garbage.

Gemini has been helpful to find source material faster, kind of like using old Google in a curated fashion, and then chase its cited sources for actual facts you can verify.

In this manner, Gemini has not been garbage in garbage out but filtering click bait ad infested sites and getting me to the relevant pages of online manuals in record time.
 
That brings me to my next point. I started my pedal board AI experiment because I am a historian by trade and wanted to test it out with something I would be able to see red flags with fairly quickly. Wrong pedal specs, incorrect features etc... What I have concluded, up to this point, is that Ai requires a semi-knowedgable, inter-active operator. Simply asking it to provide a history of X is probably gonna generate a pretty shakey answer.


TLDR; AI is great as long as you only ask questions for which you already know the answer
 
It's pretty interesting to decouple the mainstream PR machines from reality.

ChatGPT is full of hot garbage.

Gemini has been helpful to find source material faster, kind of like using old Google in a curated fashion, and then chase its cited sources for actual facts you can verify.

In this manner, Gemini has not been garbage in garbage out but filtering click bait ad infested sites and getting me to the relevant pages of online manuals in record time.
I used to use Google Scholar quite a bit; it now has an AI add-on ("Labs") which I haven't yet used for any questions I'm actually interested in, but it does at least provide links to the papers cited, and provides summaries, of what sort of quality I'm uncertain. At first blush it looks reasonably promising, though.
 
You'll find, if you wanna have a look, that they have a plan for this, much like kings of old.

They plan on a form of basic income and minimal housing for the masses, based of property taxes they want to define.

So basically, they're trying to reduce their cost exponentially.

Sam Altman even wrote a manifesto about this, called Moore's Law For Everything... And it reads like a dumped down version of the communist manifesto - minus all the pesky democratic control espoused by Marx, of course.
Earth was like that in the future world of The Expanse series.
 


It's only a matter of time before we have a new Luddite movement, driving vehicles packed with fertiliser bombs into data centres.

And only a little more time before data centres have heavily armed private militias guarding them.
 
  • Like
Reactions: DaBigBadaBoom
And AI scrapes data a lot of which has been generated by AI anyway.
So in a sense it's feeding on its own excrement.
Thus the models get gradually worse over time
The whole system will crash as there is no route to profitability given the amount of debt taken on by AI companies
Correct, training on your own bad data that's then used to train itself to improve accuracy is just like eating your own babies...
Yep- I saw this terrible AI-written post reposted on a Facebook bass group recently that misattributed Larry Graham's "slap story" to a broken amp. I searched for that story and was directed back by Google to basically a carbon copy of that same story in a Facebook post a few months back as the "source" for that info.
 
Last edited: