We had given a talk titled “Solving the Problem of Diagonalization” at an International Conference and Forum EMERGE 2024: Ethics of AI Alignment, which was hosted by The Digital Society Lab of the The Institute for Philosophy and Social Theory of the University of Belgrade. Additional information about the event can be found on the conferences’ webpage. You are invited to read our abstract in the text below, as well as others abstracts of the conference in the book of abstracts.
Solving the Problem of Diagonalization
The topic of AI alignment has recently risen in popularity due to the widespread availability of AI tools, such as LLM. However, not much attention has been given to theories of machine motivation that would seek to understand the formation of an AIs’ goals and values. The paper aims to addressl this gap by contrasting Nick Bostrom’s orthogonal theory and Nick Land’s diagonal theory. Bostrom’s orthogonal theory, implicitly assumed to be correct in most AI alignment discussions, proposes that volitional structure (motivation) and cognitive capacity (intelligence) are independent. In contrast, Land’s diagonal theory challenges this premise, arguing that goal complexity rises as intelligence does, ultimately leading to intelligence increase becoming the singular goal. As a result, alignment with human ethical values may be nearly unattainable.
The behaviour of natural intelligences indicates that Land’s diagonal theory may hold more validity, as human goals are much more complex and varied than those of less intelligent animals. This paper will delve deeper into this issue and present additional arguments supporting the diagonal theory’s validity. The paper will also explore whether aligning AI with human ethical values is feasible if the diagonal theory of intelligence proves correct. While this question is very theoretical in nature and its effects may not be immediately apparent, we will demonstrate its relevance in the current socio-political climate. In particular, we will be focusing on its importance for freedom of expression and the future of democracy, by highlighting the effect of AI on media centralization, the importance of data and algorithm transparency, and the threat posed by unaligned AI(s) in a multipolar world. AI technology has the potential to revolutionize social and ethical norms, its development and use must be carefully guided to safeguard these pillars of democracy and ensure its continued flourishing.
Keywords: AI alignment; orthogonality thesis; diagonal thesis; democracy; media transparency;
What were the main arguments presented at the International Conference and Forum EMERGE 2024? Greeting : IT Telkom
Keypoints of the presentation were:
Terminology surrounding AI. When does AI become superintelligent?
– Difference between AI alignment approaches – Orthogonal vs. Diagonal (anti-orthogonal) approach
– Reduction to Will-To-Think (which we presented in detail here)
– We looked into challenges of AI alignment – specifically whether it is better to have a morally crippled AI or an amoral super-intelligence. Additionally, we discussed what are humanities’ options in preventing a runaway AI.
– We also advocated for clearer media, data, and algorithm transparency.
– Lastly, we presented risks to freedom of expression in an era of multipolar (corporations, nations, …) competition in AI development, which could potentially destabilize international relations.