Can you explain what you think makes today's LLMs more than glorified Markov chains?

I know they're not technically Markov chains, but conceptually, it's still close enough. There's nothing revolutionarily new. No amount of loops and harness tricks is going to change that fact.