My students keep using calculators to do their homework. How can I make sure they're using abacuses instead?

a consequence of authority is fewer people get to force other people to make mistakes and try things that won't work

Yeah, speed/cost does seem like the most likely bottleneck (assuming perfect secrecy).

GAN-like, I guess. One example: train a model to modify the generated image such that it fools both a model that detects the steganographic signature and a model that checks for human-perceptible differences from the original generated image.

I've been meaning to learn more about steganography. Even if the algorithm weren't public, you could probably use a GAN setup to beat it, right?

Incredible theory of mouse-mind. A mouse would indeed expect a sentient potato to, like itself, fall for a trap of cheese.

Oh, I meant they strip the metadata automatically (I think). Otherwise it would be trivial to detect an AI image, because Midjourney adds some information to the image file explicitly stating that it was generated with AI.

Nice, nailed it!

I didn't strip the metadata, but it looks like @bsky.app does that automatically, so I think it's legit that good (I don't think the metadata would reveal that it was Midjourney, either).

Thanks, this seems to work very well; it's got me curious how they built their discriminator.

~59 human vs. ~46 AI seems like a reasonable guess to me tbh. It's not like AI is going to get easier to distinguish over time.

Lol. This one was obvious, but they're not always.

Interesting. What service are you using to detect AI-generated images?

Being a Christian does not imply being an infernalist. In particular, I find theologian David Bentley Hart's case for universalism extremely compelling.

Amen!

based and moravecpilled

There's no alpha in normality

FWIW, he routinely deletes all of his tweets, so that's not big news in and of itself. I've considered adopting the same practice as a sort of sand mandala exercise; can't speak to why he does it, though.

(better as in truer to its limitations. "alphago everywhere" would obviously be better as in more performant, but good luck learning the world simulator necessary to train that)

I think Karpathy was trying to make a point about the limits to achieving superhuman performance with this approach; "average data labeler" is a better framework than "AlphaGo for all language-expressible thought", but I agree that it has a lot of holes, including everything that you mentioned here.