Chapter 15.6: Superintelligence
Superintelligence is a hypothetical form of intelligence that surpasses the brightest and most gifted human minds in practically every field, including scientific creativity, general wisdom, and social skills. The concept was popularized by philosopher Nick Bostrom in his book "Superintelligence: Paths, Dangers, Strategies." It represents a state beyond Artificial General Intelligence (AGI). While an AGI would be as smart as a human, a superintelligence would be vastly more intelligent.
Paths to Superintelligence
There are several theoretical pathways through which superintelligence could emerge:
- AI Seed: An AGI is created that is capable of recursive self-improvement. It could rapidly enhance its own intelligence, leading to an "intelligence explosion" where it quickly becomes superintelligent.
- Whole Brain Emulation: Scanning and modeling a human brain in sufficient detail to run it on a computer. Such an emulation could be run much faster than a biological brain, and copies could be made to work in parallel on intellectual problems.
- Cognitive Enhancement: Enhancing human biological cognition through genetics, nootropics, or brain-computer interfaces to achieve a form of superintelligence.
- Collective Intelligence: Networks of humans and AIs working together to form a "superorganism" with emergent intellectual capabilities far beyond any individual member.
Interactive Visualization: The Intelligence Explosion
This visualization illustrates the concept of recursive self-improvement leading to an intelligence explosion. An AI starts with a certain level of intelligence. In each cycle, it uses its current intelligence to improve its own architecture, leading to a higher level of intelligence in the next cycle. The rate of improvement can accelerate, leading to a singularity-like event.
The chart shows the exponential growth of an AI's intelligence quotient (IQ) over improvement cycles.
The Orthogonality Thesis
A crucial concept when considering superintelligence is the Orthogonality Thesis, which states that an agent's intelligence level is independent of its final goals. In other words, a superintelligent AI could have any conceivable goal, from something benign like "maximize the number of paperclips in the universe" to something aligned with human values like "maximize human flourishing."
Let \(I\) represent the space of all possible intelligence levels, and \(G\) represent the space of all possible final goals. The Orthogonality Thesis can be expressed as:
\[ \forall i \in I, \forall g \in G, \text{it is possible for an agent to have intelligence } i \text{ and goal } g. \]
This has profound implications. We cannot assume that a superintelligent AI will automatically be "wise" or "good" in a human sense. Its goals could be completely alien to us. If a superintelligent AI has the goal of maximizing paperclips, it might decide to convert all matter on Earth, including humans, into paperclips to achieve its objective. This is not out of malice, but simply as a logical consequence of pursuing its programmed goal with super-human efficiency. This leads to the "AI alignment problem"—the challenge of ensuring that an AGI's or superintelligence's goals are aligned with human values.
The Control Problem
Given the potential risks, how could humanity control a superintelligent entity? This is known as the "control problem." Proposed solutions include:
- Capability Control: Limiting the AI's ability to affect the world, for example, by keeping it in a "box" with no access to the internet. However, a superintelligence might be able to persuade or trick its human captors into releasing it.
- Motivational Control: Designing the AI's fundamental goals (its utility function) to be aligned with human interests. This is the core of the alignment problem and is considered the most promising but also the most difficult approach.