On Fri, Apr 17, 2020, 4:41 PM James Bowery <[email protected]> wrote:
> What's the difference between "self-modifying AI" and algorithms that > search connection topologies, hyperparameter and parameter space for RNNs? > I wrote a minimal self improving agent (a 14 line C program) in http://mattmahoney.net/rsi.pdf Searching through a space of programs or parameters is learning, not self improving. A self improving agent recursively creates a more intelligent version of itself with no external input. For example, a chess program could improve by playing itself. I have argued that recursively self improving software is impossible because intelligence depends on knowledge and computing power, and a self modifying program gains neither. In my example, the program (which rewrites its own source code in C) only grows logarithmically in Kolmogorov complexity, and that gain is no more than the size of the input that tells it when to stop. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T1f1af8ac2c36937b-M8d12879723983f4c9d936ec9 Delivery options: https://agi.topicbox.com/groups/agi/subscription
