Further Reading

Two independent formalisms — one clinical, one computational — reach the same claim about what any goal-directed mind must do.

Alpha Drive Theory (ADT) is, at its foundation, a theory about the architecture of a goal-directed mind. It holds that behavior is organized around the continuous defense of perceived agency: the Alpha Drive System works to regain and maintain the felt capacity to act on the world, and that maintenance function is what makes ADT an account of all behavior rather than a repair mechanism for distress alone.

The study of how to make artificial intelligence safe, a field devoted to building minds from the ground up, arrived at a structurally identical result from the opposite direction. ADT reverse-engineers the human minds that already exist. AI safety specifies what would have to be true of any mind we construct. They’re two blueprints of the same machine, one drawn by deconstructing a working model and the other by asking what you’d have to use to construct one. They meet at the same horizon.

Steve Omohundro, “The Basic AI Drives” (2008). Omohundro argued that almost any sufficiently capable, goal-directed AI would, unless deliberately prevented, develop a predictable set of instrumental drives regardless of its actual goals: self-preservation, protection of its goals from alteration, resource acquisition, and self-improvement. The logic is hard to argue. Whatever your objective, you cannot achieve it switched off, with your objective quietly changed, or without the means to act. This means defending your capacity to act becomes a near-universal necessity.

Nick Bostrom, Superintelligence (2014). Bostrom sharpened this into two theses. The first is the orthogonality thesis which states that intelligence and final goals are independent — capability tells you nothing about what a system will value (his illustration is a superintelligence that wants only to make paperclips). The second is the instrumental convergence thesiswhich states that whatever their ultimate goals, a wide range of agents converge on the same intermediate subgoals of self-preservation, goal-content integrity, resource acquisition, and cognitive enhancement. They’re different destinations with the same road for most of the trip, and the shared stretch is always the defense of the system’s own capacity to act.

What they support in ADT. Omohundro and Bostrom support this: that a goal-directed system defending its own capacity to act is not a quirk of human psychology but instead a structural property of goal-directed agency as such. That is exactly what ADT’s central claim requires in order to generalize past the clinic door. Omohundro and Bostrom establish, from first principles about minds we might construct, the universality that ADT infers from the minds that are already constructed and running. They’re two derivations from opposite directions with the same structure — a floor that neither field can be accused of manufacturing out of its own assumptions.

Where ADT goes further. ADT doesn’t merely restate the AI result. This is the part that answers the obvious objection: “defends its capacity to act” is loose enough to fit almost anything. The convergent drives in Omohundro and Bostrom are instrumental…they’re means to a terminal goal, arising because a goal is being pursued. ADT makes a stronger, more specific claim. Agency-maintenance is not a subgoal in service of something else; it is the regulated target itself — continuous, on at all times, and the setpoint the system defends whether or not any discrete goal or threat is active. The AI theories locate the drive; ADT commits to its status. A generic control account does notcommit to the drive being terminal and continuous. ADT does. That’s a narrower claim than the AI work makes, not a wider one. That’s where any “wouldn’t any control theory predict this?” objection loses its argument.

What this is and isn’t. This isn’t a claim that ADT solves AI alignment or that the safety field needs it. Those are results that stand on their own. The evidence runs the other way: a model built to explain human minds independently reproduces the structural core of a theory based on existing ones, and that corroborates ADT’s generality without annexing anyone else’s field. One area of overlap is that the drive is structural, a property of the control architecture and not of the particular goal whether anything is consciously experienced or not. On the AI side, that’s the orthogonality thesis. In the Alpha Drive Theory, it’s the claim that agency-maintenance is content-independent. Same feature…two blueprints.

Further reading

Omohundro, S. M. (2008). The Basic AI Drives. In Artificial General Intelligence 2008: Proceedings of the First AGI Conference. IOS Press.

Bostrom, N. (2014). Superintelligence: Paths, Dangers, and Strategies. Oxford University Press. (Instrumental convergence and orthogonality are developed in Chapter 7, “The Superintelligent Will.”)