This post was rejected for the following reason(s):
Low Quality or 101-Level AI Content. There’ve been a lot of new users coming to LessWrong recently interested in AI. To keep the site’s quality high and ensure stuff posted is interesting to the site’s users, we’re currently only accepting posts that meets a pretty high bar. We look for good reasoning, making a new and interesting point, bringing new evidence, and/or building upon prior discussion. If you were rejected for this reason, possibly a good thing to do is read more existing material. The AI Intro Material wiki-tag is a good place, for example. You're welcome to post questions in the latest AI Questions Open Thread.
Introduction
Artificial Intelligence (AI) has made significant strides over the past few decades, evolving from simple rule-based systems to complex models capable of performing tasks that were once thought to be the exclusive domain of human intelligence. However, as AI systems become more powerful and autonomous, the need for ensuring their alignment with human values and ethical principles becomes increasingly critical. This article delves into the emerging field of Artificial Wisdom, a concept that seeks to address the multifaceted challenges of AI alignment by integrating ethical and meta-ethical considerations into the design and deployment of AI systems.
The discussion is based on a detailed presentation I gave at VAISU in May 2024, the video of the presentation is linked below. I am very passionate about this topic and love to talk to people and other AI Safety researchers about it. I generally tend to record them and put them on YouTube, but you can find articles about those discussions as well on my profile. The presentation covers a wide range of topics, from the philosophical underpinnings of wisdom to the practical challenges of aligning AI systems with human values. This article aims to provide a comprehensive and technical overview of the key ideas presented, while also offering additional context and explanations to make the content accessible to a broader audience.
The Concept of Artificial Wisdom
Defining Wisdom in the Context of AI
Wisdom, in the context of AI, is not merely about intelligence or the ability to perform tasks efficiently. It is about the generation of ethical systems that allow machines to make decisions that align with human values and promote human flourishing. The presenter defines wisdom as an "ethics generator," analogous to how intelligence is a "logic generator." Just as intelligence allows machines to generate the logic needed to solve problems, wisdom enables machines to generate ethical systems that guide their behaviour in a way that is aligned with human values.
This distinction is crucial because it highlights the difference between instrumental goals (the means to achieve an end) and terminal goals (the end itself). Intelligence is concerned with choosing the right instrumental goals to achieve a given outcome, while wisdom is concerned with determining what the terminal goals should be in the first place. This idea is rooted in the orthogonality thesis by Nick Bostrom, which posits that the two types of goals are independent of each other—a highly intelligent system can have any set of goals, and wisdom is what ensures that those goals are aligned with human values.
The Role of Meta-Ethics in Artificial Wisdom
Meta-ethics is the branch of philosophy that explores the nature of ethical statements, their meaning, and how we come to know what is right or wrong. In the context of artificial wisdom, meta-ethics provides the framework for understanding and modeling ethical systems. However, wisdom goes beyond mere theoretical understanding—it involves the practical application of ethical principles in real-world scenarios.
One of the ways to look at wisdom is as the implementation of meta-ethics in a way that allows machines to navigate complex ethical dilemmas. This involves not only understanding the ethical principles that guide human behaviour but also being able to adapt those principles to different contexts and situations. For example, a wise AI system should be able to generate ethical systems that are appropriate for different cultures, time periods, and social contexts, much like the adage "When in Rome, do as the Romans do."
The Challenges of AI Alignment
The Broader Problem of Alignment
One of the key challenges in AI safety is the alignment problem, which refers to the difficulty of ensuring that AI systems act in ways that are aligned with human values. I believe that the current approaches to AI alignment, such as inner alignment, outer alignment, and goal misgeneralization, are insufficient because they focus on solving specific instances of the problem rather than addressing the broader issue or the bigger picture.
The alignment problem is not new—it has existed throughout evolutionary and human history in various forms. For example, there has always been a need to align the actions of individuals and groups with the broader goals of society. However, the alignment problem becomes significantly more complex when dealing with AI systems, particularly artificial superintelligence (ASI), which has the potential to surpass human intelligence and act autonomously.
The Rabbit Hole of Alignment
Not to introduce a lot of new words but I term the concept of recurring need for alignment as an "alignment rabbit hole," which refers to the idea that alignment is an emergent issue that has always existed in some form. Initially, there was an alignment problem between nature and animals, then between animals and humans, and finally between humans and other humans. Now, the alignment problem exists between humans and machines.
The key insight here is that we cannot simply align AI systems with humans and expect everything else to fall into place. We need to recursively align AI systems with all elements of the chain, from nature to animals to humans. If AI systems are aligned with humans but not with nature, for example, they could inadvertently destroy the environment, which would ultimately harm humans as well. This recursive alignment is necessary to ensure that AI systems act in ways that are beneficial to all elements of the ecosystem.
The Dynamic Nature of Human Values
Another major challenge in AI alignment is the dynamic nature of human values. Human ethics and values are not static—they change over time and vary across different cultures and societies. This poses a significant challenge for AI systems, which need to be able to adapt to these changes in order to remain aligned with human values.
Current approaches, such as coherent extrapolated volition (CEV) and predictive models, attempt to address this issue by extrapolating human values. However, these approaches are limited because they do not fully account for the complexity and interrelatedness of human values, reality isn't made of independent identical distributions (iids). Artificial wisdom, on the other hand, offers a more holistic solution by addressing both the spatial and temporal dimensions of human values. This allows AI systems to generate ethical systems that are appropriate for different contexts and time periods.
The Data Structure of Wisdom (DSW)
Overview of DSW
The Data Structure of Wisdom (DSW), is one of my initial projects in this field that covers very abstract and fundamental groundings of my initial thoughts about artificial wisdom, it is basically a conceptual framework that outlines the path from intelligence to wisdom. DSW is an abstract representation of how humans have developed intelligence in machines, and based on a rough approximation of that, it provides a blueprint for developing wisdom for alignment.
The framework is divided into two main parts: the data-to-intelligence path and the verisimilitude-to-wisdom path. The first one covers the development of AI from observation to understanding, while the second one covers the development of wisdom from verisimilitude to the impact of alignment. The development of wisdom will likely take a significant amount of time, just as the development of intelligence has taken several decades.
Key Elements of DSW
Except for the initial half where the DSW framework deals with the ongoing process and blueprint of intelligence, it advances them to include several key elements that are essential for the development of artificial wisdom:
The Role of Abstraction in DSW
Abstraction is one of the initial abilities that need to be worked on for a proper segue from intelligence to wisdom. It plays a crucial role in the development of both intelligence and wisdom. In the context of intelligence, abstraction allows machines to generalize from specific instances and apply their knowledge to new situations. Similarly, in the context of wisdom, abstraction allows machines to generate generalised ethical systems that are applicable across different contexts and scenarios. Wisdom is not merely a plugin that can fit or evolve over intelligence, it surely does share many similar groundings and abilities but is a different paradigm. Similar to how current learning and intelligence models are based on Algorithms and Data, but they are a different paradigm than the classical data processing algorithmic programs.
The Future of Artificial Wisdom
The Path Forward
The development of artificial wisdom is still in its early stages, and there are many challenges that need to be addressed, for instance, there isn't a consensus on a specific name for most of the topics discussed above, it goes by various names like Artificial Metaethics, Computational Ethics, Aristotelian Virtue Ethics for Machines, etc. There is a huge importance of patience in this process, as it took humanity several decades to develop the intelligence we have today. Similarly, it will likely take several decades to develop artificial wisdom.
One of the key challenges is the integration of wisdom and intelligence. While intelligence allows machines to perform tasks efficiently, wisdom ensures that those tasks are aligned with human values. The ultimate goal is to create self-sustaining artificially wise machines that can generate ethical systems and make decisions that promote human flourishing.
The Role of Human Agency
There is a very significant role of human agency in the development of artificial wisdom. As AI systems become more autonomous, there is a risk that humans may outsource their agency to machines, leading to a loss of control over important decisions. This highlights the need for governance and ethical oversight in the development and deployment of AI systems.
Therefore artificial wisdom is not just about aligning machines with human values—it is also about ensuring that humans retain their agency and control over the decisions that affect their lives. This requires a holistic approach that takes into account the interplay between human agency, machine autonomy, and ethical principles.
Conclusion
Artificial wisdom represents a promising new direction in the field of AI safety, offering a holistic solution to the complex challenges of AI alignment. By integrating ethical and meta-ethical considerations into the design and deployment of AI systems, artificial wisdom has the potential to ensure that AI systems act in ways that are aligned with human values and promote human flourishing.
However, the development of artificial wisdom is still in its early stages, and there are many challenges that need to be addressed. These include the dynamic nature of human values, the need for recursive alignment, and the integration of wisdom and intelligence. Despite these challenges, I remain optimistic about the future of artificial wisdom and its potential to create a safer and more ethical future for humanity.
As we continue to explore the possibilities of artificial wisdom, it is essential to remain mindful of the ethical implications of our work and to ensure that we are always striving to create AI systems that are aligned with human values. Only by doing so can we hope to create a future in which AI systems act as partners in promoting human flourishing, rather than as threats to our existence.