I suspect AI getting good enough to create a Marvel movie is basically indistinguishable from it being able to create a good movie. The problem is LLM’s are basically just creating gibberish which mimics meaningful text well enough you don’t notice it’s meaningless as long as it’s a short enough sample.
Modern AI techniques are approaching human level capacity by combining a lifetime worth of information in their training sets with nearly human levels of computation. It’s not at the level of professionals, but being roughly as good as a teenager means crossing that threshold could happen well before people are ready for it.
A low end estimate of say 10 cycles per neuron / second * 90 billion neurons * 10,000 connections per neuron is already 90 peta operations per second.
Of course we also expect AI’s to be faster but conversely we aren’t trying to simulate a person that needs to walk and talk, just a rough equivalent of whatever is needed for some task.
Bigger models slow the rate of degradation, but that's it.
I know, it's trained with human data and its output is based on that. The same can be said of human authors, though. Arguably we never create an idea from scratch, originality is just putting ideas together in really non-standard ways.
I think trying to make a context free intelligent system is making it excessively hard for the AGI folks. Our brain definitely figures out what to do for pre-sentient reasons, or using pre-sentient hardware.
> A chasm between the third best and fourth best script writer, graphic designer, lawyer, etc. is just starting to open. ChatGPT might write the next Marvel movie but will never write the next Everything Everywhere All at Once.
Not my words. By Eric Peters CIO of One River Asset Management