I recommend The Meme Machine, it’s a shame it didn’t spawn a huge literature. I was thinking a lot about memetics before reading it, yet still I feel like I learned a few important things.
Anyhow, here’s an idea inspired by it:
First, here is my favorite right way to draw analogies between AI and evolution:
Evolution : AI research over time throughout the world
Gene : Bit of code on Github
Organism : The weights of a model
Past experiences of an organism : Training run of a model
With that as background context, I can now present the idea.
With humans, memetic evolution is a thing. It influences genetic evolution and even happens fast enough to influence the learning of a single organism over time. With AIs, memetic evolution is pretty much not a thing. Sure, the memetic environment will change somewhat between 2020 and whenever APS-AI is built, but the change will be much less than all the changes that happened over the course of human evolution. And the AI training run that produces the first APS-AI may literally involve no memetic change at all (e.g. if it’s trained like GPT-3).
So. Human genes must code for the construction of a brain + learning procedure that works for many different memetic environments, and isn’t overspecialized to any particular memetic environment. Whereas the first APS-AI might be super-specialized to the memetic environment it was trained in.
This might be a barrier to building APS-AI; maybe it’ll be hard to induce a neural net to have the right sort of generality/flexibility because we don’t have lots of different memetic environments for it to learn from (and even if we did, there’s the further issue that the memetic environments wouldn’t be responding to it simultaneously) and maybe this is somehow a major block to having APS capabilities.
More likely, I think, is that APS-AI will still happen but it’ll just lack the human memetic generality. It’ll be “overfit” to the current memetic landscape. Maybe.
I recommend The Meme Machine, it’s a shame it didn’t spawn a huge literature. I was thinking a lot about memetics before reading it, yet still I feel like I learned a few important things.
Anyhow, here’s an idea inspired by it:
First, here is my favorite right way to draw analogies between AI and evolution:
Evolution : AI research over time throughout the world
Gene : Bit of code on Github
Organism : The weights of a model
Past experiences of an organism : Training run of a model
With that as background context, I can now present the idea.
With humans, memetic evolution is a thing. It influences genetic evolution and even happens fast enough to influence the learning of a single organism over time. With AIs, memetic evolution is pretty much not a thing. Sure, the memetic environment will change somewhat between 2020 and whenever APS-AI is built, but the change will be much less than all the changes that happened over the course of human evolution. And the AI training run that produces the first APS-AI may literally involve no memetic change at all (e.g. if it’s trained like GPT-3).
So. Human genes must code for the construction of a brain + learning procedure that works for many different memetic environments, and isn’t overspecialized to any particular memetic environment. Whereas the first APS-AI might be super-specialized to the memetic environment it was trained in.
This might be a barrier to building APS-AI; maybe it’ll be hard to induce a neural net to have the right sort of generality/flexibility because we don’t have lots of different memetic environments for it to learn from (and even if we did, there’s the further issue that the memetic environments wouldn’t be responding to it simultaneously) and maybe this is somehow a major block to having APS capabilities.
More likely, I think, is that APS-AI will still happen but it’ll just lack the human memetic generality. It’ll be “overfit” to the current memetic landscape. Maybe.