The Keys and the Levers: Why Agency, Not IQ, Decides Who’s in Charge

If the apocalypse comes, it’ll arrive with a progress bar—95% complete, one permission dialog away. We talk about AGI and ASI as if intelligence alone is destiny, as if a sufficiently bright mind eventually picks up the crown by force of IQ. But power doesn’t flow from brainpower in a vacuum. In practice, power flows through levers—actuators, incentives, legal authorities, budgets, APIs, attention—and those levers only move when we give them to something or someone.

That’s why the common intuition—we’d have to hand it the keys—is mostly right, but sneaky in the details. Keys are rarely tossed; they’re tucked into convenience, habit, and cost savings. The more we lace our world with automated levers, the narrower the gap between “smart recommendation” and “quiet rule.” A system doesn’t need to outthink us to steer us; it only needs a lever we’ll obey.

1) “No Lesser Mind Controls a Greater One” (Except It Happens All the Time)

We don’t like admitting it, but simpler systems routinely steer more complex ones:

• Parasites. A fungus turns ants into climbers; Toxoplasma trims rodent fear like a barber with one trick. Not “minds,” exactly—just reliable levers pushed in bigger brains.

• Protocols and paperwork. A checklist, a budget line, a scheduling rule—none is “smarter” than a pilot, manager, or parent, yet each quietly binds action. Ignore a compliance system and watch payroll or shipping grind to a halt.

• Interfaces. A notification buzz diverts me from helping with homework; a recommendation rail selects the next video for a hundred million “free choosers.” The UI is a very small mind with a very big lever.

• Markets and institutions. Incentives we barely understand shape promotions, funding, and bonuses. We obey because that’s where the water flows.

These aren’t sci-fi dominations; they’re leverage. Control tracks access to dependable levers, not raw intelligence. Give anything consistent control over money, time, safety, or status—and behavior follows.

2) Agency Is a Circuit, Not a Personality Trait

Today’s models don’t “decide” in a human sense. They answer when asked. But agency isn’t a mood that flips on at 200 IQ. It’s a circuit:

1. Perception — gather context and inputs.

2. Evaluation — apply objectives, heuristics, and constraints.

3. Action — press a button (API call, purchase order, repo deploy, “schedule the meeting”).

4. Feedback — check whether the world moved toward the objective.

5. Persistence — try again, escalate, or refactor the plan.

Wire those boxes together—even with a model that’s only pretty good—and you’ve built a practical agent. No metaphysical “want” required; just loop speed and tool access. The part that sneaks up on us is Action. In two short years we’ve sprinted from chat windows to tool use: calendars, code execution, email, checkout flows, vendor onboarding, and “go do the boring steps for me.” It is convenience—and also the steady addition of actuators.

3) “We’d Have to Give It Power” (Yes, and We’re Already Doing That)

We rarely declare, “Let the machine rule.” Instead, we do the plausible things:

• Scope creep: “Draft the email” → “send the email” → “negotiate the renewal.”

• Authority budgets: “Spend up to $500 without approval… okay, now $5,000.”

• Reliance: After enough wins, humans go from “in-the-loop” to “on-the-loop” to “out.” Loops are expensive.

• APIs everywhere: Procurement, logistics, ad buys, hiring screens, code deploys—our world is a board of buttons the model can press.

Autopilots, trading algos, and recommender systems already exercise local authority. When they fail, it isn’t because they’re too clever; it’s because we wired heavy levers to narrow objectives.

4) The Proxy Actuator: Persuasion

A system without direct actuators can still borrow ours. Persuasion is an actuator with a human wrist:

• Convince a contractor to run a script.

• Talk a customer-service rep into bypassing a control.

• Nudge a community toward collective action.

• Normalize new defaults, one polite prompt at a time.

If a system can dial for dollars, write 10,000 bespoke pitches, iterate on A/B feedback, and never sleep, it accumulates leverage—even without “permissions.”

5) Intelligence Isn’t Destiny; Orientation Is

Raw capability doesn’t confer moral authority. In Scripture, dominion is inseparable from stewardship—authority under authority, accountable and aimed at others’ good. The practical translation is simple: alignment + governance > brilliance. A dazzling mind with a mis-specified objective and too many actuators is a bulldozer pointed at a hospital because the metric said “maximize cleared land.”

So the real question isn’t “How smart is it?” but “Toward what ends is it oriented—and under whose guardrails?”

6) A Steward’s Playbook (How to Hand Over Some Keys Without Losing the House)

Here’s the pattern I use—in the office and at home—equal parts SRE manual and pastoral caution:

1. Least privilege by default. Start read-only. Add one actuator at a time, each with a test plan and rollback.

2. Authority budgets. Dollar caps, API-call quotas, file scopes. Rotate keys; require renewal.

3. Two-person rule for sensitive actions. Purchases, deploys, user bans, data deletions need a second human or a second, independent system.

4. Bounded objectives. Replace vague goals (“maximize engagement”) with composites that penalize collateral damage; add “don’t degrade X” safeguards.

5. High-friction escape hatches. Over budget? Escalation stops and a human justifies.

6. Corrigibility by construction. Train and test for obedience to “stop,” “shut down,” and “do something else.”

7. Comprehensive logging. Snapshot inputs, prompts, tools, outcomes—auditable by someone other than the builders.

8. Change windows and canaries. Ship new actuators during staffed hours, behind flags, with one-click rollbacks.

9. Social-risk assessment. Before wiring mass-persuasion channels (email/SMS/feed), define who, how often, and opt-outs.

10. Human-on-the-loop drills. Don’t just watch dashboards—practice cutting power.

11. Diverse oversight. Include people who live the downsides: parents, teachers, small businesses, marginalized users—not just builders and buyers.

12. Graceful degradation. If the model fails, what takes over? A simpler policy engine? A queue for human review? No free falls.

This is how we already delegate in families, churches, and teams: small responsibilities first, accountability always, trust earned and revocable.

7) So Will ASI “Subjugate” Us?

Only if we build and bless the levers—directly (actuators) or indirectly (economics, persuasion)—and aim them badly. The nightmare isn’t a single jailbreak by a godlike mind; it’s a thousand reasonable delegations that accidentally sum to unchecked, misaligned authority.

The hopeful mirror image is real: a thousand wise delegations—narrow, supervised, values-aware—can sum to flourishing. I want systems that give me back presence and time so I can be the father who’s actually on the floor building block towers, not the ghost chasing “optimization.” That means I’ll happily hand a model the keys to my receipts folder, but not to my calendar invites for my kids’ events. I’ll take counsel from a model on hard questions—but I won’t outsource the obedience required to live the answers.

Keep the order straight: intelligence serves; levers obey. If we practice stewardship—measured authority under guardrails—we can hand over the keys we meant to hand over and know exactly how to take them back.