> The problem with "reverse engineering people" is that you don't need to, the things they like in any era follows predictable and generic patterns. These patterns can be encoded into AI provided we find enough curated training data.
That only works to a certain degree. For the infamous example, try making GPT output the correct number of asterisks, you will get mixed results. You don't have any problem with typing exactly 1589 asterisks because you run the stateful counting algorithm in your head. GPT has no idea about the algorithm - it has to reverse engineer it from the text, and can only extract the vague correspondence between a number and a string about this or that length. You don't give humans examples to reverse engineer, you teach them to count.
This is a simplest example, it might even learn to count eventually, as it's far more capable in certain aspects. But as the dimensionality of the task grows, the amount of resources and training data required to reverse engineer it grows much faster.
Sure, it can spot some patterns and that can look good, but some things are just plain invisible in the result - you will have a hard time making it learn higher level concepts because they highly depend on hardwired things like the dumber part of neural circuitry and biochemistry in humans, which the model doesn't have.
It's like trying to make a photo in a dark room - no matter how you improve the sensitivity of your camera, you might not have a single photon in it.
> This is because the artist can't read your mind.
Yes, this is what I mean by the limited capacity of a simple textual description. It's a fundamental limitation - natural language is just poorly suited for the detailed conceptualization. A sketch, or a conceptual diagram, or other higher order control methods have far more capacity to explain your intent, and that's the direction those models move to. At which point their usage is nothing like "type something simple and receive the result".
The asterisks thing is another issue. LLMs don't need to do this to replace directors.
>Yes, this is what I mean by the limited capacity of a simple textual description. It's a fundamental limitation - natural language is just poorly suited for the detailed conceptualization.
Except LLMs can accept sketches as input. The higher order methods of communication are covered by encoders.
That only works to a certain degree. For the infamous example, try making GPT output the correct number of asterisks, you will get mixed results. You don't have any problem with typing exactly 1589 asterisks because you run the stateful counting algorithm in your head. GPT has no idea about the algorithm - it has to reverse engineer it from the text, and can only extract the vague correspondence between a number and a string about this or that length. You don't give humans examples to reverse engineer, you teach them to count.
This is a simplest example, it might even learn to count eventually, as it's far more capable in certain aspects. But as the dimensionality of the task grows, the amount of resources and training data required to reverse engineer it grows much faster.
Sure, it can spot some patterns and that can look good, but some things are just plain invisible in the result - you will have a hard time making it learn higher level concepts because they highly depend on hardwired things like the dumber part of neural circuitry and biochemistry in humans, which the model doesn't have.
It's like trying to make a photo in a dark room - no matter how you improve the sensitivity of your camera, you might not have a single photon in it.
> This is because the artist can't read your mind.
Yes, this is what I mean by the limited capacity of a simple textual description. It's a fundamental limitation - natural language is just poorly suited for the detailed conceptualization. A sketch, or a conceptual diagram, or other higher order control methods have far more capacity to explain your intent, and that's the direction those models move to. At which point their usage is nothing like "type something simple and receive the result".