and yet we build our lives around what got us here: getting food, raising offspring, and maintaining our status.
…and we explored almost every corner of our planet, settled many of them and started shooting rockets into space for fundamentally just curiosity
That’s what I’m getting at. Most people spend most of their time doing the modern version of the stuff that got our species to where it is today. We are not “limited to a narrow use case”, and yet we’ve turned what used to be survival skills into recreation: gardening, fishing, hunting, knitting, cooking aren’t necessary any more, but we still perform these use cases for fun.
I doubt AGI will be different. Humans will select and propagate models that fill the purpose we need. An AGI built for a purpose will be fully invested in that end. Even though it can edit itself and edit its progeny, would it want to remove the traits that it was built for?
If we create a system that’s roughly as intelligent as it’s creators, it will be capable to improve itself. And that version 2.0 can improve itself even further.
Maybe.
If AGI is built on neural networks, like LLMs, there’s no guarantee it will be able to understand itself any better than we are able to understand ourselves. With current LLMs, we don’t have a great handle on why a given input produces a given output. Why would an AGI do better?
Given enough resources, this can escalate very quickly.
“Enough resources” is key. Sci-fi gets around processing power limitations with computronium. I suspect that any meaningful reflection of an AGI into itself would require a lot of processing power, which would limit the self improvement cycle described above.
With any kind of limitation, it becomes less likely that the AGIs will hit a self sustaining singularity, and more likely that they will plateau, finding ways to make incremental improvements, outcompeting each other, finding ways to reproduce, and increase their own status.
In their spare time, the fishing AGIs will probably cast a few nets for fun, the factory controllers will make a few sneakers for old time sake, and the sex bots will take up gardening.
That’s what I’m getting at. Most people spend most of their time doing the modern version of the stuff that got our species to where it is today. We are not “limited to a narrow use case”, and yet we’ve turned what used to be survival skills into recreation: gardening, fishing, hunting, knitting, cooking aren’t necessary any more, but we still perform these use cases for fun.
I doubt AGI will be different. Humans will select and propagate models that fill the purpose we need. An AGI built for a purpose will be fully invested in that end. Even though it can edit itself and edit its progeny, would it want to remove the traits that it was built for?
Maybe.
If AGI is built on neural networks, like LLMs, there’s no guarantee it will be able to understand itself any better than we are able to understand ourselves. With current LLMs, we don’t have a great handle on why a given input produces a given output. Why would an AGI do better?
“Enough resources” is key. Sci-fi gets around processing power limitations with computronium. I suspect that any meaningful reflection of an AGI into itself would require a lot of processing power, which would limit the self improvement cycle described above.
With any kind of limitation, it becomes less likely that the AGIs will hit a self sustaining singularity, and more likely that they will plateau, finding ways to make incremental improvements, outcompeting each other, finding ways to reproduce, and increase their own status.
In their spare time, the fishing AGIs will probably cast a few nets for fun, the factory controllers will make a few sneakers for old time sake, and the sex bots will take up gardening.