Joel Z. Leibo [1], Alexander Sasha Vezhnevets [1], William A. Cunningham [1, 2], Sébastien Krier [1], Manfred Diaz [3], Simon Osindero [1]

[1] Google DeepMind, [2] University of Toronto, [3] Mila Québec AI Institute

We published a more academically oriented version of this post on arxiv, available here.

Disclaimer: These are our own opinions; they do not represent the views of Google DeepMind as a whole or its broader community of safety researchers.

Beyond Alignment: The Patchwork Quilt of Human Coexistence

"We pragmatists think of moral progress as more like sewing together a very large, elaborate, polychrome quilt, than like getting a clearer vision of something true and deep." – Richard Rorty (2021) pg. 141

Quite a lot of thinking in AI alignment, particularly within the rationalist tradition, implicitly or explicitly appears to rest upon something we might call an 'Axiom of Rational Convergence'. This is the powerful idea that under sufficiently ideal epistemic conditions – ample time, information, reasoning ability, freedom from bias or coercion – rational agents will ultimately converge on a single, correct set of beliefs, values, or plans, effectively identifying 'the truth'. We suspect that the axiom of rational convergence is independent of the rest of AI safety and ethics. By this, we mean something analogous to the way different geometries can be constructed by selecting different "parallel axioms" e.g. Euclid's parallel axiom gives you the geometry of a plane, postulating instead that parallel lines converge gives you elliptic geometry, and postulating that parallel lines diverge generates hyperbolic geometry. The choice of which parallel axiom to adopt is like the choice of what to study (e.g. a plane or a sphere). Just as selecting different parallel axioms generates distinct yet internally consistent geometries, the choice of whether to accept the axiom of rational convergence shapes our theories about AI alignment and ethics. Could the Axiom of Rational Convergence be similarly independent of other foundational commitments? Its appeal lies in its offering a potential objective grounding for AI goals, "something true and deep" towards which alignment can aim. However, what if this axiom doesn't hold, or what if we simply choose not to adopt it as the basis for navigating the complexities of advanced AI? This pos...

Posts

Wikitag Contributions

Comments