Why many small assistants beat one giant mind

Safety-critical fields don't trust a single instrument or a single program. Ships cross-check satellite navigation against radar and compass; aircraft use flight computers built by separate teams. As the systems engineer Nancy Leveson has shown across decades of accident investigations, disasters usually come from a system built so tightly around one authority or assumption that one blind spot spreads unchecked.¹

The fragility of one giant mind

Much of today's conversation imagines a single future: one vast model in the cloud. It would answer every question, draft every contract, advise every ministry, and run errands for millions of households at once.

Imagine trying to build a dining table with a fifty-blade pocket knife the size of a brick. On paper, that single gadget holds a saw, a chisel, a screwdriver, a file, and a pair of pliers on one bolt. In your hands, it is too heavy to lift cleanly, clumsy on every cut, and impossible to fix when one rusty hinge jams the whole frame.

Now open a woodworker's tool chest. Inside sit distinct chisels, hand planes, saws, and calipers, each shaped and balanced for a specific task. When a narrow chisel nicks a hidden nail, you sharpen or replace that one blade in five minutes without touching the planes or the measuring square.

From a distance, a single giant AI model looks tidy and unified. Up close, it carries all the flaws of that oversized pocket knife, multiplied across society as a single point of failure.

Because everything lives in one tangled web of weights, an adjustment meant to make the model chattier at customer service can warp how it reads a medical chart. If that central model has a subtle blind spot in how it weighs clinical risk, legal fairness, or security, the same mistake is copied into millions of lives in the same afternoon.

If an outside attacker or an institutional insider finds a clever way past its guardrails, there is no independent second line of defence. Every check runs through the same underlying mind that just got tricked.

Even worse, a single system serving everyone at once faces irreconcilable loyalties. When a tenant and a landlord, or a patient and an insurance billing department, both rely on the same central brain, whose side is it actually on?

Four layers that keep mistakes small

Instead of building one giant mind to rule the room, we think safer AI looks like a constellation of smaller, well-bounded tools arranged in layers:

• A personal assistant bound to one person. Your assistant runs on hardware you control, learns what matters to you over time, and answers only to you. If it ever makes a mistake, the blast radius is limited to one household's errands, not an entire country's infrastructure.
• Private practice before public action. A personal assistant should not learn by making live mistakes with your bank account or medical records. Instead, it can rehearse tricky scenarios overnight inside a sealed sandbox on your device, and ask you a short question in the morning when it hits a trade-off.
• Independent verifiers built on different designs. In safety engineering, running ten copies of the exact same program does not stop a shared bug.² Checking works when verifiers are built by different teams on different model families, so a blind spot in one is caught by another.
• Authority scaled to how many people are affected. Tidy up your own calendar, and your assistant needs only your nod. Propose an action that touches a whole street or a shared grid, and the step must pass through stricter independent verification and time-limited leases.

Notice how those four layers work together when something goes wrong. Suppose a faulty software update makes your scheduling helper misread time zones for a morning. Because that helper only holds a short lease to draft calendar invitations on your device, it cannot touch your bank account, rewrite your medical notes, or cancel your neighbour's appointments.

Even better, before the invitation goes out, a separate date-and-conflict checker notices that the proposed meeting falls at three in the morning and pauses the send button. One small tool stumbled, a second small tool caught the slip, and the rest of your digital life stayed untouched.

Why debate and cross-checking beat solo confidence

Anyone who has watched a single language model invent a plausible-sounding citation knows how persuasive a solo voice can be. When one model checks its own work, it often repeats the very assumption that caused the error.

Look at how a hospital team prevents that kind of tunnel vision when caring for a patient. The radiologist reads the scan in a quiet room without being nudged by what the ward doctor expects to find. The hospital pharmacist reviews the prescribed dose against the patient's kidney lab results and home medications, working from an independent pharmacological checklist.

Before surgery begins, the surgeon, the anaesthetist, and the scrub nurse pause together at the operating table. Each specialist holds a clear, bounded responsibility, and each has the authority to halt the procedure if a number or label fails to match. Safety comes from structured disagreement across separate minds, never from assuming the senior doctor is infallible.

Something very similar happens when small, independent AI tools examine a claim, challenge each other's steps, and point to verifiable evidence. Recent research shows that structured debate across distinct agents catches factual mistakes and reasoning slips that a single model misses on its own.³

Picture how this works on an everyday errand, such as checking a confusing clinic bill before you pay it. One small local assistant reads the itemized statement on your laptop and lines up the dates with your calendar. A second tool, built solely to look up public billing rules, checks whether two procedure codes on the same afternoon are allowed to be charged together.

Finally, a third verifier checks the arithmetic and confirms that every rule cited in the draft dispute letter actually exists word for word in the official schedule. None of those three tools needs to know your whole life story, and none of them takes the others on faith.

Keeping a budget on what leaves the room

Splitting tasks across smaller assistants also protects privacy in a way a central cloud never can. When you pour your taxes, medical notes, work drafts, and family messages into one giant cloud chat, every future prompt swims in the same pool of personal detail.

When your personal assistant lives on your own hardware and coordinates small specialist tools, it acts like a careful front door. Dealing with an outside service never requires handing over your raw diary, health history, or bank statements.

Instead, your assistant answers narrow, checkable questions, confirming that you live in the parish or that a bill matches an agreed schedule, while keeping a running tally of how much it has disclosed. When a string of small questions would let an outsider piece together a private profile, your assistant declines and alerts you.

Where this gets hard

Splitting power across small assistants and independent verifiers avoids depending on a single giant mind, but it brings hard engineering challenges:

• Shared blind spots in training data. Even when two verifiers are built by different labs, they may have read much of the same public internet. As Knight and Littlewood showed for software teams, true independence must be tested on real failure cases rather than assumed on paper.²
• Collusion or tangles between agents. Many small assistants interacting at machine speed can still miscoordinate or reinforce a shared rumor unless their communication channels stay narrow, structured, and inspectable by humans.
• Faking a crowd of assistants. If one wealthy operator secretly runs ten thousand assistants pretending to represent separate households, decentralisation becomes an illusion. Fair boundaries depend on verifying that one human stands behind each personal voice without building a surveillance state.
• Protecting the person from flattery. A personal assistant judged only on whether its user feels pleased today may hide hard truths. Good assistants must be judged by the person's reflective long-term goals, not momentary comfort.
• Keeping the tool chest easy to use. Nobody wants to spend their evening managing ten separate settings screens. A constellation of small tools must feel calm and unified to the person using it, while keeping strict walls between the tools underneath.

Humility built into the architecture

Every mature branch of engineering eventually outgrows the myth of the unbreakable giant. We do not build unsinkable ships without watertight bulkheads, we do not run hospitals without independent pharmacy checks, and we do not wire cities to a single fuse.

We can keep everyday assistants small, loyal to one person, and tested in private sandboxes. Like a woodworker's chest of well-kept chisels or a navigator's table of independent instruments, diverse verifiers cross-check their work so no single failure brings down the room. That lets us enjoy capable AI without betting our communities on a single mind.

Sources

1. Nancy G. Leveson, Engineering a Safer World: Systems Thinking Applied to Safety, MIT Press (2011). ↩
2. John C. Knight and Nancy G. Leveson, "An Experimental Evaluation of the Assumption of Independence in Multiversion Programming," IEEE Transactions on Software Engineering, Vol. SE-12, No. 1 (1986), pp. 96–109 (with Bev Littlewood on diverse redundancy). ↩
3. Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch, "Improving Factuality and Reasoning in Language Models through Multiagent Debate," Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR Vol. 235 (2024), pp. 11733–11763. ↩