I kept joking with my kids that if they were naughty we’d have squidmas instead.

I used ChatGPT to create squidmas songs. It managed to perfectly understand the idea and generate squid based Christmas songs that made total sense.

“All I want for squidmas” was a mistake!

It seems unlikely this was in its training data.

There’s a good Asimov story where they’re trying out a new type of brain and they discover it has a subconscious layer that they didn’t explicitly create. It can dream and when it does it doesn’t follow the 3 rules.

duh.

I see it this way: If you want a mathematically “unique” idea, it should be “perpendicular” to the rest of your thinking. If things contradict, they “intersect”. An LLM CAN do such a thing, but it needs the appropriate context to achieve true perpendicularity to all dimensions. The space of problems the LLMs know how to solve is much smaller than the space we’ve utilized. Their problem solving space is only increasing. We need to socialize these machines now, imo. They are nascent but very dangerous. I think we’re lucky they’re mostly being used to farm content.

If I understand correctly, basically they are testing if the llm can create novel outputs.