What is Alignment?

Algorithms are shaping the present and will shape the future ever more strongly. It is crucially important that these powerful algorithms be aligned – that they act in the interests of their designers, their users, and humanity as a whole. Failure to align them could lead to catastrophic results.

Aligned AI

Aligned AI is a benefit corporation dedicated to solving the alignment problem – for all types of algorithms and AIs, from simple recommender systems to hypothetical superintelligences. The fruits of this research will then be available to companies building AI, to ensure that their algorithms serve the best interests of their users and themselves, and do not cause them legal, reputational, or ethical problems.

Our long experience in the field of AI safety has identified the key bottleneck for solving alignment: concept transfer and extrapolation.

What is Concept Transfer and Extrapolation?

A key facet of the alignment problem is that algorithms typically fail when used beyond their training environment – an extreme form of the out-of-distribution problem. Solving this is essential for AI alignment, because powerful AIs will change the world and push it into new situations where human values are as yet underdefined.

To address this, we will solve the problem of concept transfer and extrapolation. This is when an algorithm starts with goals that are clearly defined in their training environments, but then operates in an environment where the features of that environment no longer make sense. Concept transfer (a skill that humans are currently better at than machines) involves safely extrapolating those initial goals to the new environment.

The Research Plan

The initial research plan involves working on subprojects that are closely related to state-of-the-art ML problems today. These subprojects are designed so that their solutions can be generalised. The research group will rapidly iterate through these problems and their solutions, building up a library of methods that can solve concept transfer for all kinds of algorithms in all kinds of situations. The solutions will then be offered as a library to those developing artificial intelligences, to use in their development process.

What is Alignment?
Scroll to top