2 Comments

I've tried InstructPix2Pix but never saw a result nearly as spectacular as the examples. We'll see how quickly things improve. As an aside, what are your thoughts of GPTZero? Could make for an interesting article...

Expand full comment

Agree. It's definitely hit and miss. On some occasions, it failed to add pretty natural objects to a picture (I asked for a race car to be added to a picture of a highway). And sometimes it manages to convincingly rework the entire scene of an image. But like you said, we're at the early stages...I expect progress here to be as rapid as with the original AI art generators.

As for GPTZero, it looks a lot like a reimplement of this old GPT-2 detector (yup, from back in the GPT2 days):

https://openai-openai-detector.hf.space/

Curiously enough, the GPT-2 detector seems to still work pretty well at identifying text produced by ChatGPT. There are a slew of services like GPTZero coming out, which I guess is fully expected. I'm sure that Google has a version of this running in the background when ranking pages - Google explicitly doesn't want purely AI-generated content to rank well.

I feel like it's good we have these tools, and we'll probably see a bit of a "cat and mouse" game where LLMs get better and outsmart these detectors, so new ones are built to detect those, etc.

Expand full comment