3 ms·

> And especially once AIs are drawing from the writing of other AIs, which themselves are quoting AI (dark, I know) it might become quite difficult to detect,

> And especially once AIs are drawing from the writing of other AIs, which themselves are quoting AI (dark, I know) it might become quite difficult to detect, No, it will be easier to detect, because the new AIs will learn to mimic earlier generations of their AI - and their faults - rather than mimicking human writing. Because that will score higher on a loss function. I'm going to assume Dead Internet theory - aka most text online is spambots. It's not strictly true, as humans are obstinately continuing to use the Internet and pouring content into them. But the Internet puts those humans on a level playing field with bots. And AI is basically perfect spambot material - it looks superficially different, meaning you can't remove it with simple patterns. So the training set, which is just scraped off the Internet, will be primarily composed of AI-generated material by less-sophisticated systems. Human-written content will be in the minority of the training set, meaning that no matter how good the optimizer statistics or model architecture is, the system will be spending most of its time remembering how to sound like a bad AI.

williamcotton 4y ago
There’s still a fitness function based on human interactions with reality: the cost of running these language models needs to keep up with the revenue generated. The revenue ultimately comes from the human economy concerned with housing, food, warmth, water, etc, so the language models need to cater to the concerns of the source of such economic activity by creating content useful enough to real people in order to cover the costs of production.
CuriouslyC 4y ago
The real problem isn't AI content generation, it's a lack of content filtering. There has been a mountain of human garbage pushed out by people trying to make money which is already too big to navigate, AI will make the mountain much larger but the problem won't change. We need AI discriminators that analyze content and filter out stuff that is poorly written, cliche, trite and derivative.

williamcotton 4y ago

There’s still a fitness function based on human interactions with reality: the cost of running these language models needs to keep up with the revenue generated. The revenue ultimately comes from the human economy concerned with housing, food, warmth, water, etc, so the language models need to cater to the concerns of the source of such economic activity by creating content useful enough to real people in order to cover the costs of production.

CuriouslyC 4y ago

The real problem isn't AI content generation, it's a lack of content filtering. There has been a mountain of human garbage pushed out by people trying to make money which is already too big to navigate, AI will make the mountain much larger but the problem won't change. We need AI discriminators that analyze content and filter out stuff that is poorly written, cliche, trite and derivative.