Superintelligence Can Pursue Your Goals. But Who Decides What Matters?
Superintelligence can pursue almost any goal. The harder question is who defines what matters, how success is measured, and whether people own the map their agents follow.
Mark Zuckerberg has published a 6,500-word case for putting superintelligence into everyone’s hands.
The obvious arguments are about jobs, safety, open-source models, and whether Meta should be trusted with this future.
One sentence stopped me.
Meta wants to build a personal agent that understands you, your goals, and everything you care about.
That sentence carries much of the argument.
Humans do not have a clean goals file.
What matters to us is often vague, contradictory, and constantly changing.
Even when a goal sounds clear, we may disagree about what achieving it would mean.
An AI can help us pursue a goal. But first, somebody has to define it, decide when it counts as achieved, and understand what it depends on.
That is the problem behind matters.global.
Matters.global
Matters gives people and AI a shared map of what needs to change. Instead of trusting an agent to reconstruct your intentions from a long chat history, you write down the outcomes, open questions, and constraints that matter.
The AI can read this map before it acts. It can see what is ready to move, what remains blocked, and which open matter would unlock the most progress. Resolution conditions describe what success looks like, so the agent cannot confuse activity with completion.
Some steps can be started by the AI. Others need a human decision, evidence from outside the system, permission, or collaboration. That boundary stays visible. Matters does not decide what should matter. It helps the AI navigate the structure people have defined.
The model is simple.
A matter is something worth revisiting: a goal, concern, decision, risk, responsibility, or unresolved question.
Each matter has conditions that define what must become true for it to count as resolved. Matters can also depend on other matters.
That creates a map.
The map does not tell people what they should want. It makes different perspectives visible. It shows what is unresolved, what is blocking progress, and where one intervention could unlock several other matters.
This is not task management. A task tells an agent what to do. A matter explains why something needs attention and what reality should look like when it is resolved.
That difference becomes important once agents can execute almost anything.
The Meta essay is itself a hidden matters graph
Its central argument depends on several chains of assumptions:
Broad access to AI leads to individual empowerment, which leads to invention, prosperity, and employment.
Distributed superintelligence creates a balance of power, which produces safety.
Open models give defenders an advantage over attackers.
American leadership supports a democratic technological order.
These are not facts. They are propositions about how one outcome depends on another.
Some may be right. Others may be wrong. They need conditions, evidence, and room for competing interpretations.
Zuckerberg asks us to imagine that everyone has a superintelligent lawyer. He argues that this would make justice fairer and more efficient.
It could. It could also produce a superintelligent litigation arms race. Giving everyone a weapon does not resolve the underlying conflict. Sometimes it only increases the speed and cost at which the conflict is fought.
The question is not only who has access to intelligence. It is whether we can make our assumptions, dependencies, and definitions of success visible enough to coordinate that intelligence.
Alignment is not only a relationship between one person and one agent
My goals depend on other people.
A company’s goals depend on employees, customers, suppliers, regulators, and communities. Scientific progress relies on previous findings, physical experiments, funding, and human judgment. Political goals collide with the goals of people who see the same situation differently.
An agent aligned perfectly with one person can still make the larger system worse.
Agents need to understand what their user wants. They also need to see what those goals depend on, whom they affect, and where several people’s interests meet.
This does not require universal consensus. It requires a shared structure in which disagreement remains visible and cooperation is still possible.
Superintelligence changes execution. It does not solve intention.
Meta’s vision gives everyone a more powerful engine. I am interested in the map it follows.
Who sets the direction? What counts as arrival? Who owns the agent’s representation of what matters to us?
We should be able to inspect that map, change it, move it to another agent, and share parts of it without exposing everything else. We should also be able to compare different maps and find points of common progress without forcing everyone into one worldview.
That, to me, is the alignment problem.
If you want to test this idea, bring one real goal, unresolved problem, or difficult decision to matters.global. Turn it into matters, resolution conditions, and dependencies. Then ask your AI what can move now.