I enjoyed this - and like the Spy analogy very well. It's amusing to think of LLM responses to be the equivalent of someone with extreme imposter syndrome desperately trying to conjure an effective disguise for what is expected of them. I think some of the token prediction talk comes from the same place - trying to remind people to not be too trusting and double check for cracks in the outputs you receive.
I enjoyed this - and like the Spy analogy very well. It's amusing to think of LLM responses to be the equivalent of someone with extreme imposter syndrome desperately trying to conjure an effective disguise for what is expected of them. I think some of the token prediction talk comes from the same place - trying to remind people to not be too trusting and double check for cracks in the outputs you receive.