Why Generative AI is Extremely Bad: Trees, Poems and People.


NOT a human brain
(via Guru99.com)

Making an artificial brain is NOT the same thing as what all marketing is calling "artificial intelligence," what is marketed as "AI" is an incredibly inefficient SQL severely limited by the absurd volume of data dumped into it AND the bottlenecking of generative software technique.

Having scraped countless giga- if not terabytes of data to create images and text using a mere sentence or two is quite bad, but it is especially bad for visual art. 


NOT an SQL Server
(via HopkinsMedicine)

AND NOW IT'S TIME FOR A METAPHOR: 

If you wanted to create a generative AI that makes "trees," you can't just feed one tree into it, because it will just keep making variations on that one tree.

You can't just feed one type of tree into either, because again, it limits its tree-generation ability. 

You have to give it as much variety as possible to get it to successfully generate a believable tree. Not a tree that's realistic; not a tree that holds up to extended scrutiny, just a believable tree. 

I shall stretch this metaphor further; if you were able to get a good sample of every tree in North America, your generative system can only make believable North American trees.

Now, it is estimated that there are 73,000 distinct varieties of trees on earth at this time. Another estimate says that roughly 9,000 of them haven't been discovered. 
 

If you were to get one good version of the 64,000 known types of trees on planet Earth, the User Interphase for generative AI still needs to be improved because, in part, it's marketed to generate things. 

When a human types "show me a tree," then this Tree-generating AI can generate a tree but those parameters are entirely too vague and who knows what it'll spit out. A coniferous Baobab tree? An olive-pine tree? A lemon/maple tree? 

All of those sound fascinating, but in this metaphor, we've only 
been able to program trees' superficial traits into our Tree-QL. 

We haven't programmed the actual genetics of trees into another database so we have no clue whether or not these generated trees are even sustainable.

Sustainability is already a concern for generative AI because by 2029, if things remain on track, generative AI will account for 1.5% of the power consumption of the entire planet Earth

Can AI generate a tree lovely enough that Joyce Kilmer would write a poem about it with such a catchy first line it's quoted for nearly a century afterward? 

Not one that's sustainable. Not for the power and money it would cost. 


An Actual Tree (via Earth.com)




 

Comments

Popular posts from this blog

5 Things They Don't Tell You About Being In the U.S. Military

5 Ways A Deployment Turns You Into An Asshole

The Specifc Anger of America