[anon373839]

> If they are so adamant on protecting their IP,

What they are trying to protect doesn't qualify as intellectual property. Only 4 categories of IP exist: (1) copyrights; (2) patents; (3) trade secrets; (4) trademarks.

The capabilities embedded in model outputs don't qualify. Machine-generated outputs are ineligible for copyright. They aren't covered by patents. They aren't trade secrets, because the model companies are selling them rather than keeping them secret. And of course, trademarks are conceptually inapplicable.

This leaves the model companies with contract law (ToS) which is pretty inept because it can't bind third parties. And technical measures, like the ones being discussed in the article. And, of course, politics.

Frankly, I think it's pretty ridiculous to even think that models can be protected from being learned from. I feel the Stanford Alpaca team demolished that idea 3 years ago.

[eloisius]

The hypocrisy of the pro-AI mega corp arguments makes my head spin. For three years they’ve been using the example of a human reading books and then outputting creative works influenced by them as analogy for training AI on copyrighted works. Now suddenly we’re supposed to not draw the same parallel about a hypothetical person who learned from Claude and is now outputting creative work based on it.