stance-and-level

A formal theory of trust developed at the limit-computable level, where beliefs are computationally omniscient (AIXI, reflective-oracle Solomonoff induction and related supercomputations), guides AI safety engineering work.

A formal theory of trust developed at the level of probability frames and Jeffrey–Bolker models, which are specified above the level where computability is quantified and are not AI designs, guides AI safety engineering work.

A formal theory of trust developed as whole-mind models of rationality at polynomial time or below, where the model is itself an implementable AI design, guides AI safety engineering work.

The phenomena that matter for trust between humans and AI systems, above all the computational part of human uncertainty about values, can be represented at the computationally unbounded level.

A formal theory of trust developed at the limit-computable level guides AI safety engineering work.

A formal theory of trust developed at the logical-induction level guides AI safety engineering work.

A formal theory of trust developed as whole-mind models at polynomial time or below would guide AI safety engineering work.

A formal theory of trust developed at the level of probability frames and Jeffrey–Bolker models guides AI safety engineering work.

Human uncertainty about human values is to a substantial degree computational uncertainty, uncertainty about the consequences of what one already accepts, rather than only uncertainty about which world obtains.

A computationally uncertain UDT exists over a logical-induction-like belief process, specified so that its policy evaluation uses the process's later beliefs, and it satisfies a tiling theorem.

At least one situation that arises between humans and AI systems, developers selecting systems on their predicted policies or a shutdown commitment a fully updated system would not renew, is one in which the updateful and the updateless prescriptions differ, so at least one engineering decision depends on which is followed.

The updateless side is high value to AI safety engineering work even where the theory of computational updatelessness is easy: the engineering decisions updateless reasoning changes are large enough that the programme's expected contribution to a formal understanding of trust exceeds the updateful plan's at equal cost.