Re: Running out of data to train AI programs.

On 12/11/24 3:02 PM, RonO wrote:

On 12/11/2024 4:45 PM, erik simpson wrote:

On 12/11/24 2:17 PM, RonO wrote:

https://www.nature.com/articles/d41586-024-03990-2
>
The claim in this article is that soon the AI programers will run out of data to train their AI on.  If they want to improve their AI they will have to create their own data, but how are they going to do that?
>
My guess is that they will identify what data would be most beneficial to have and try to generate it.  It could direct medical research into generating useful data.
>
They could also spend decades weeding through the data that has already been used and throw out the trash data.  They could also go through the scientific experiments in a field and use the good data, but remove the conclusions of the researhers, and see what conclusions the AI can come up with, cross check the conclusions to see if there was anything missed by the original researchers and use that to train some other AI.
>
Ron Okimoto
>

Why not let AI create its own data.  Then we wouldn't have to worry about it.  And think of the great literature and movies!
>

That results in increased bogus output by the AI.
 The AI start to "hallucinate" when fed data generated by other AI according to one article that I recall reading.  I think one poster responded that AI hallucination can be caused by other factors too. Maybe the AI generate data in a format that they were not trained to deal with, and that causes issues in assimilating AI generated data.
 Ron Okimoto
 

I was joking about AI -> AI generating anything sensible.  AI is very useful detecting patterns in all kinds of data.  Applying in to weather data could save millions if not billions of dollars.  AI navel-gazing isn't going to work.