It doesn't know it's wrong afterwards, just as it doesn't know it was wrong before. It's just the training of the model suggests to the generative process that apologising is the most likely thing to be said after being called out. The architecture of large language models isn't really fit for trustworthiness, just plausible-sounding responses.

Can you give a reference to the Architecture paper? I'm researching around this at the moment.

> a prime example of the human mind finding connections between random data points

This is what poetry is working with. Calling it a "figment of your imagination" belies how powerful our imagination is in how we live. Seeing Jesus in burnt toast is a testament to the power of the image of Jesus in the viewers unconscious. Otherwise you may interpret it as Chuck Norris. Similarly people don't tend to go to war on the basis of a personal cost-benefit analysis, they think of Duty, Valour, Heroism, and Patriotism.

Poems are not literal representations. They evoke feelings that are concretely "real" in our lives, but cannot be indexed directly with words.

I'm glad to see people calling "bullshit" on this. I'm big into developments of PhotoSynth/Bundler/PMVS and other interesting 3D photogrammetry, so this is close to my heart. I'd just like to clear up a confusion though: "You can do a better job of 3D scene reconstruction with three or four Kinects in different places[...]" Unfortunately you can't, as the Kinects use structured light reconstruction, and the IR light patterns from multiple Kinects would confuse each other.