Superintelligence for Everyone: What Distributed Power Demands of Your Company

A recent op-ed proposed a philosophy for the next decade of artificial intelligence, resting on three pillars: individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety. The central argument is that trusting a single benevolent superintelligence, or a small group of experts who control it, would be more dangerous than distributing that capability to everyone, because there is no objective answer to what constitutes a good life and because concentrated power has historically limited choice.

You can agree or disagree with the philosophical thesis. What you cannot do is treat it as a distant debate, because the choice between concentrating and distributing AI power does not play out only among large labs and governments. It plays out, at a smaller scale and with immediate consequences, inside every company putting AI into production right now. And the decisions that make it real are quite concrete: which vendor you depend on, which use cases you prioritize and how much power each agent receives inside your systems.

Vendor concentration is an architecture decision, not an opinion

If your company's entire AI operation depends on a single API, a single provider and a contract you do not control, you have already adopted the concentrated model in practice, regardless of your position on the matter. Whoever sets pricing, usage policy, context limits and what happens when the model version changes without notice sits outside your organization.

The counterweight available to an ordinary company is not training your own model. It is portability: keeping an abstraction layer between your application and the provider, treating prompts, evaluations and context data as internal assets, and running an honest switching test. How much time and money would it cost to migrate your main agents to another model in thirty days? If the answer is "we do not know," the dependency runs deeper than the contract suggests.

One point follows directly from the thesis: if power lies in access to capability, your company's defensible asset is not the model, it is the proprietary context you organize around it. Structured data, documents and access rules are what remain when the vendor changes, and that is the work behind building tailored agents in AI Factory.

Invention or automation: the same technology, two different outcomes

Another part of the thesis deserves the attention of anyone approving budget: the distinction between using AI to automate and using AI to expand human capability. The argument is that superintelligence's greatest contribution will be invention, not substitution, and that the effect on jobs and the economy depends on which of the two uses prevails. The more the pendulum swings toward pure automation, the more negative the impact tends to be.

Translated into the day-to-day of a Brazilian company, that is a use case selection criterion, not an ideological debate. There are two kinds of AI project:

• The kind that reduces the cost of something the company already does: fewer hours checking documents, fewer people triaging tickets.
• The kind that expands what the company is capable of doing: serving a segment that never added up before, launching a product that used to require a whole team, analyzing a volume of information that was previously discarded.

Both are legitimate, and the first is easier to approve because the return is arithmetic. But a portfolio made up only of the first kind produces a leaner company doing exactly what it already did, a fragile position when a competitor uses the same technology to do something new. A useful test when approving each initiative: does this project shrink cost or increase capability? If nothing on the list answers the second question, the portfolio is out of balance.

Distributed power inside the company requires counterweights inside the company

The op-ed uses a strong analogy: if only one person had a superintelligent lawyer, they would hold an unfair advantage even when wrong on the merits; if everyone has one, the outcome is fairer. The logic works well for societies and becomes revealing when applied to your internal network, because it flips sign.

Distributing AI capability across the whole operation, with no counterweight, does not produce balance. It produces attack surface. An agent with broad credentials, access to sensitive databases and permission to execute actions is concentrated power in a component that has no judgment, can be manipulated through prompt injection and does not know when it is being used outside its purpose. Multiply that by dozens of agents spread across different departments, each one procured without going through security, and you have the shadow AI scenario that already costs dearly in audits and in exposure under LGPD.

The internal counterweight has a name and it is operational:

• A dedicated identity for every agent, never a shared human credential.
• Real least privilege, with data and action scope declared and reviewable.
• An audit trail of everything the agent read, answered and executed.
• Guardrails against prompt manipulation and sensitive data leakage.
• A clear policy on which data can feed which models and providers.

That is the set of controls an AI governance track must establish before scaling. Distributing power with no counterweight is not empowerment, it is giving up control.

Where the open-source analogy deserves caution

The thesis holds that for most risks, cybersecurity included, the history of open-source software shows that giving everyone full access to powerful systems is the best path to safety over time. Anyone who runs security operations daily has an important footnote to add here.

Open-source software did not become safer because the code was visible. It became safer because an operation organized itself around that visibility: maintainers, coordinated disclosure, a patching pipeline and a defensive community with an incentive to act. Openness was a condition, not the cause. In AI, wide access raises capability on both sides, with a difference in pace that matters: the attacker reaps immediate benefit in reconnaissance, social engineering and exploit development, while the defender only reaps it with an operation able to use that same capability, with detection, response and continuous coverage.

The claim that "wide access produces more security" is therefore true under a condition rarely stated out loud: that defensive maturity exists on the other side. For the average company, expanding AI access without expanding detection and response hands the advantage to the attacker. That is the difference between having a tool and having an operation, and it is the role a SOC with MDR plays in this equation.

Practical conclusion

Whether the philosophical bet is right, history will tell. What is already decidable today, inside your company, is simpler and more urgent: do not reproduce internally the concentration you criticize externally, and do not confuse distributing power with distributing risk.

In practice, it comes down to three questions worth a board meeting. How much would it cost to switch model vendors in thirty days? How many of our AI initiatives increase capability rather than merely cut cost? Does every agent already in production have an identity, least privilege and an audit trail? Those who answer all three with numbers in hand are building genuinely distributed power. The rest are simply outsourcing the decision to whoever sells the model.

If your company is defining architecture, use cases and governance to scale AI without giving up control, talk to the Intelliway team at /empresa#contato.