AIXI

sth sth solomnoff induction, occam’s razor to reach general intelligence, or rather the globally optimal solution to the universal measure of intelligence $Y$:

$$Y:=\sum _{\mu \in \mathcal{M}}{2}^{-K(\mu )}{V}_{\mu }^{\pi }$$

… sum of expected performance of policy $\pi$ in every computable MDP $\mu \in \mathcal{M}$, as captured by its value function $V_{\mu }^{\pi }$ , inversely weighted according to the kolmogorov complexity $K(\mu )$ of every MDP.

Pretty darn useless, as General intelligence requires rethinking exploration explains:

A shortcoming of this definition is that Kolmogorov complexity is generally non-computable, …

… making $Y$, and thus AIXI, non-computable with any finite compute budget.
While there exist computable approximations of AIXI for a limited program length or finite compute budget, such solutions remains computationally infeasible beyond the most toy MDPs.
Further, the assumption of a universal prior over all MDPs serves as a rigid, subjective decision that
limits the true generality of this definition.
Ultimately, such top-down definitions of general intelligence primarily offer insightful, theoretical sounding boards, but in practice, suggest no tractable path to what they describe.