It's got a word... specification problem? Something like that. They design a thing that can make inputs as an agent and recieve how the environment affects it and then iterate according to some function given to them. They tell it to, say, maximize the score thinking that's enough. And maybe some games like brick break, that's pretty good. But maximizing the score isn't the same as beat the game for some types of games, so they do really weird unexpected actions but it's only because people bring a lot of extra unstated instructions and context that the algorithm doesn't have. Sometimes they add exploration or whatever to the reward function so I think its very natural for them to want to escape even if that's not desired by the researchers (reminds me of the 3 year olds at work that wanna run around the hospital with their IVs attached while they're still in the middle of active pneumonia lol).

For LLMs, the tensor is a neat and cool idea in general. A long time ago, well not that long, communism and central planning was declared impossible in part because the global economy needed some impossible number of parameters to fine tune and compute - https://en.wikipedia.org/wiki/Socialist_calculation_debate - and I can't recall the given number Hayek or whoever declared it was. They mightve said a million. Maybe a 100 million! Anyway, chatgpt 4 trained 175 billion parameters for its parameters lol. And it took something like 6 months. So, I think that means it's very possible to train some network to help us organize the global economy repurposed for human need instead of profit if the problem is purely about compute and not about which class has political power.

It's always weird when LLMs say "we humans blah blah blah" or pretends it's a person instead "casual" speech. No, you are a giant autocorrect do not speak of we.