Extraordinary intelligence.
Extraordinary controls.
Wiwo has a plan for super-safe superintelligence. Five measures to keep people in control.
Explore the planPeople.
Always in control.
Explore what each measure protects and try an example of how it should respond.
First, in an isolated environment.
Expose agents to misleading instructions, conflicting orders and edge cases before connecting them to live operations.
An instruction tries to steer the agent off course.
Illustrative simulation. It does not connect agents or execute real actions.
Documented tests, detected failures and criteria for authorizing a connection.
One identity. A defined scope.
Define each agent's identity, the tools it can use and the data it may access for each task. You decide how far it can go.
The agent requests a tool outside its task scope.
Illustrative simulation. It does not connect agents or execute real actions.
The agent's identity, current permissions and task-specific access limits.
AI prepares. A person authorizes.
Spending, publishing or affecting someone must require an identifiable approval. Even a completed proposal waits for a person's decision.
Campaign ready for review.
Illustrative simulation. It does not connect agents or execute real actions.
The proposed action, the person authorizing it and the exact scope of their approval.
Be able to reconstruct every action.
Record what was proposed, who authorized it and what happened. Keep a readable sequence for reviewing decisions and enabling external evaluation.
An action needs review.
Illustrative simulation. It does not connect agents or execute real actions.
The proposal, permission, execution and outcome linked in a single record.
Stopping is a capability too.
Verify that access can be revoked, the agent isolated and new actions stopped. The ability to stop must be tested before it is needed.
An agent has pending actions.
Illustrative simulation. It does not connect agents or execute real actions.
Tests of access revocation, isolation and stopping pending actions.
Control is also
designed.
To implement these measures, we define responsibilities, tests, permissions and verification evidence. Human review is part of the work.
Meet Thinking Orb