SN 1003 AI Generated Homogeneity

[Badrod]

Is it a spoiler to comment on an item in the show notes before that episode of Security Now actually airs? If so, Spoiler Alert! ... but not really.

Richard's comment and the cartoon in the Show Notes for SN 1003 regarding Artificial Intelligence and Homogeneity reminded me of another AI related story from several months back. Computerphile's Dr. Mike Pound discussed a paper that predicts larger and larger generative AI models would require exponentially more and more training data--leading to series of diminishing returns. Has Generative AI Already Peaked? -- Computerphile (YT) The link to the paper, No "Zero-Shot" Without Exponential Data

The Nature article from 1003, AI models collapse when trained on recursively generated data

If you want to know what happened to the French Bulldogs? Just read the Health section of their wikipedia article.

 

[Cebess]

This whole issue reminds me of what happens with crops when there is not sufficient generic diversity. Having too much AI generated data just causes the whole environment to be weak.

 

[Fuzzball]

A big disappointment of mine regarding AI is that its most visible effect seems to be the generation of phony pictures.

This may come from the fact that I'm older and thus spent my first two decades in the concrete world, but I have noticed that the more effectively AI pretends to be human the more I seem to despise it.

 

[SeanBZA]

This whole issue reminds me of what happens with crops when there is not sufficient generic diversity. Having too much AI generated data just causes the whole environment to be weak.

Which is why there is a song called "yes we have no bananas today, and why the Cavendish is also under threat as well.