This sequence attempts to formulate (and resolve) the problems of agent foundations using the tools of algorithmic information theory, particularly by investigating the properties of the AIXI agent, its variants, and proposed generalizations. The central thesis is that “universal artificial intelligence” (~the theory of AIXI) should serve as a central paradigm for agent foundations. In particular, the informally-stated limitations attributed to AIXI (as a bounded/embedded agent) should be made rigorous within the paradigm. I expect that many can also be repaired within the paradigm, or do not arise in practice at all.
I probably will not get around to writing my own proper introduction to the AIXI model and its philosophical motivation until I need it for my thesis, so in the meantime this sequence requires prerequisite knowledge.
Here is a quick intuitive explanation from Thomas Larsen, which I endorse except for its position on objective randomness according to AIXI.
The original paper:
https://arxiv.org/pdf/cs/0004001
For a longer and more complete introduction, read one of Hutter (et al.’s) books: