I wrote an article about structure of self-improvment and all known ways how SI may be self-limiting. I will add you idea in it. I am still polishing the article and will send it to academic journal so I will not publish it openly now, but I had recently a presentation about the topic, and here are the slides: https://www.slideshare.net/avturchin/levels-of-the-selfimprovement-of-the-ai
Basically there are 5 main levels of SI—hardware, learning, changes of architecture, changes of code, changes of goal system and creation of copies of the AI. Each has around 5 sublevels, so an AI could improve on around 30 levels. However on almost each level it will level off.
Anyway even if SI on each level will be modest like 5 times, total SI will be like 10E20, so it is enough to create superintelligence.
Different ways how SI may slowdown also discussed. So the process will be slower and probably level off but even 1000 times improvement is enough to create rather dangerous system.
If anybody interested in the article I could send a draft privately.
I wrote an article about structure of self-improvment and all known ways how SI may be self-limiting. I will add you idea in it. I am still polishing the article and will send it to academic journal so I will not publish it openly now, but I had recently a presentation about the topic, and here are the slides: https://www.slideshare.net/avturchin/levels-of-the-selfimprovement-of-the-ai
Basically there are 5 main levels of SI—hardware, learning, changes of architecture, changes of code, changes of goal system and creation of copies of the AI. Each has around 5 sublevels, so an AI could improve on around 30 levels. However on almost each level it will level off.
Anyway even if SI on each level will be modest like 5 times, total SI will be like 10E20, so it is enough to create superintelligence.
Different ways how SI may slowdown also discussed. So the process will be slower and probably level off but even 1000 times improvement is enough to create rather dangerous system.
If anybody interested in the article I could send a draft privately.