#AI/Machine learning and IRL

16 messages · Page 1 of 1 (latest)

safe swallow
#

Hey i just wanted to kick off with a thougth of mine.
Top points i want to think about:

  • How we raise kids
  • How we learn whats good and bad and how we act based on that
  • Time we spent for learning
  • The way AI learns or how we write "Smart scripts"
  • Other things that can connect to this

I want to hold this convo not so hard technically (like discussing in deepth how code works) more logical, so that fron what we all learned in life, can be applyed to this topic.

To start off:

I had the thought, that most of the ways people create AI (either in small or large scale) is half hearted many times.
The reason i think so is, that most super in deepth writers of this kinda code, have problems with IRL stuff (no offence!).
That said, sometimes they probably think "their normal" is "all normal" and other behaviour is wrong (or not smart/dumb).
I was thinking, that this big AI centers get their data from tons of user inputs... But!
The user inputs are mostly crap! Why?
Because when you think about raising a kid, you normaly wont let it play with bad kids rigth? So you prevent the kid from learning dumb/bad behaviour of others.
But for machine stuff, they take all inout not filtered. So just a simple example.

If you and most around you(like the people the kids are looking up to) are telling the kid while it raises the sky is green all time. The kid thinks thats true, even if some untrustworthy people said no its blue, the kid will believe the heaviest input.
I think we can apply this to AI too. If we habe 100.000ds of bad inputs compared to 100dreds of good inputs, they weight of the bad ones are throwing it off = the AI maybe learns wrong.

So why are there not filtered inputs, that make it smarter?

And also i thought, why a AI should be faster in learning than a well taught kid? The best pc there is, is our brain. We cant even grasp half of how it works so far 😂
And if we teach a kid smth, it takes years and daily exercise to master it. Like walking, talking and pooping. Its all hard to teach and based on how the kid is, we need to teach it sifferenr compared to other kids.
(Here we can think of thr AI how it takes in infos, we need to handle it different)

Does that make sense what i think?
What are your thougths? Anything were you say "yes absolutly" or "no you are wrong"? 🙂

latent condor
#

"Filtered inputs" as you call them, is done. I don't know why you think it isn't, but stuff like data preprocessing and cleaning is always done when implementing ML

safe swallow
#

whats ML?

#

machinen learning?

#

Ok then lets phrase it different,
im not sure if its filtered enough. Probably having many flaws.
Why? Because we cant even do that IRL well as humans and make a lot mistakes, because we are either having emotions messing stuff up or no emotions = forgetting cruical infos because we dont know them well

tawdry plume
#

Data quality is arguably the most work in most ML jobs

tawdry plume
#

Also who decides what is and isn't good training data?

safe swallow
# tawdry plume Also who decides what is and isn't good training data?

exactly.

I just think because we humans all think different, AI cant be an allrounder.
We all want smth. else, or prefer smth. else (lets say food, some like spicy some like sweet more) and TRUE == FALSE never 😄

But thinking about this they can run 100 years of "training" in a few days, why arent they outsmarting us till now?
Because the data or the learning concept is not good enough.

IMO it is possible to calculate everything that happens in the future, theoretically. The problem here is,
the Data we need to analyze. And analyzing this data is also an impact on smth, like energy use or heat production, which is new data which changes stuff. To calculate what happens in the future, we would need every data existing, even data we dont know that exists (like for human behaviour we would need to know all time exact body information, which includes diseases we cant see for now, which change human behaviour).
So what im on about here, is that if we look further around all data we gather is like a small drop of data and we dont know how to handle it for real.
Likewise, we say "I/we dont know why person XYZ was acting like that" the reason behind is that we dont had enough data or we cant access the data we need.

Im super curious about why we have so smart people in the world, it seems, but we dont have real working AI, or even if its at one specific field only, they are having many flaws. Which kinda is sad.
But i also see a big problem in how companys work, "Save money, make money", if there would be more "make your job good" then i think the world would look different allready pretty much.

latent condor
# safe swallow exactly. I just think because we humans all think different, AI cant be an allr...

It seems like your concept of AI is focused on only a tiny tiny subset of what AI is. What you are referring to is artificial general intelligence (AGI) which is still considered by most to be a hypothetical type of intelligence. Some people speculate that it is possible to achieve AGI soon but at the moment it hasn't been done.

There are narrow AI which have surpassed humans in a certain domain for decades already so I'm not sure what you mean by "we dont have real working AI, or even if its at one specific field only, they are having many flaws"

tawdry plume
#

(also the 100 years figure was just a random guesstimate, obviously depends on how complex your nn is and how much compute you have available)

tawdry plume
acoustic gust
#

ai gonna take over world pls dont study it

tawdry plume
#

😂

safe swallow
#

i knew i get some new input if i write smth like this here^^

acoustic gust
#

Ai gonna take my job fuck