| > GPT 'learning' from its own output Interesting implication. A system such as GPT learning from its own output is almost obviously absurd. A computer model getting fed its own output as training data feels like some kind of short circuit, or maybe just a circuit with a capacitor and resistor that simply slowly dissipates energy, until nothing happens/changes anymore. Yet, isn't what humans do mostly the same, just more sophisticated? Humans pick up experiences and references and information and, inspired from these, create "new" things. I'd argue that the "new" things humans create are really just new combinations/iterations of things that they picked up earlier in life. The difference between GPT and humans is mostly: the human recombination system is vastly more complicated/sophisticated than its GPT counterpart and produces more chaotic output. In the end, the _essential_ part of the, let's call it, cultural evolution, that's happening with humanity, maybe isn't so much the creation/recombination aspect (which produces this vast stream of new things, almost randomness), but rather the selection process.
Lots of random things are being created all the time, most are being discarded, except for those that somehow appear valuable or get selected for other reasons. The selection of things from the vast random stream might matter more than the process that creates the vast random stream. |