AI alignment boundaries
- Konstantyn Spasokukotskiy
Abstract
The article presents a theoretical study that highlights applicability
boundaries for the contemporary AI algorithms. Past the boundaries the
technology presents an existential threat to humanity. A discussion how
to extend the safety margin concludes the article. In particular, the
article analyzes various AI alignment classes, which are differentiated
by algorithmic principles. The applicability constraints are being
considered. To quantify the phenomenon, AI alignment limits are compared
against cognitive task complexity and mapped onto the same scale. It
reveals safe operations ranges for the algorithmic approaches. Another
insight is that the AI alignment limits are forming a distinct data row.
An improved alignment criterion is being proposed as a result of
extrapolation. Respectively, a new class of AI alignment is being
identified. It resembles being failsafe for all actual cognitive tasks.
An algorithm feature to implement the alignment class is proposed.