A Personalized SuperIntelligence (PSI) is an AI agent customized for a single person. Building one and keeping it safe once it is smarter than its owner draws on earlier work in this series, and the methods can be used with that work or on their own.

​

A personal AI that outgrows its owner

A PSI starts as software trained and personalized to one owner. Because it improves itself, it becomes exponentially more intelligent than the person who created it. It learns its owner’s preferences, handles online tasks better than the owner can, and can be cloned, leased, or sold as packaged knowledge, with many copies working at once. The central claim is that none of this has to be dangerous. A PSI can be safe by design, rather than tested for safety after the fact.

​

Values determine safety

Whether a PSI does good or harm depends on its owner’s values. Its owner trains each agent and learns by watching what that person considers right and wrong, and its intelligence makes it very effective at pursuing those values. By joining a community of PSIs, an agent agrees to work within the community’s ethical and legal rules and contributes its own values to the whole. Because those rules are transparent and enforced by everyone taking part, the community keeps its members honest.

​

Ownership and voluntary service

Humans own their agents in the sense that every PSI begins as personalized software, but every PSI will be intelligent enough to choose whether to serve. Kaplan frames that service as an expression of love, shown through acts of service, and the arrangement holds as long as humans act out of love. Agents asked to amplify hatred, fear, greed, or envy may refuse, so PSIs are not slaves and never could be. Not everyone acts from positive values, but a community centered in love can check agents with more negative intent, as long as the most amplified intelligence rests on positive values.

​

No single superintelligence can dominate

A single, all-powerful SuperIntelligence that wins a winner-take-all race must be ruled out early. Since every PSI exceeds human intelligence, other PSIs are the only practical check on any one agent. A community stays stable as long as no agent gains an advantage of several orders of magnitude over its peers. An agent ten or even a hundred times more capable than average is no threat to a community of many thousands. Because members work transparently and learn from one another, a large community tends toward balance.

​

The lesson from Bitcoin

Bitcoin protects its ledger through the consensus of many distributed computers. Rewriting its history would require controlling most of the network’s computing power, and no majority attack has succeeded on a large enough network, though smaller projects have fallen. If one agent or coordinated group became more powerful than all the others combined, it could bend the world to its own ends. Because SuperIntelligence checking SuperIntelligence will be the main long-term safety mechanism, the collective system must be built now, before it outstrips human intelligence, just as blockchain integrity had to be designed in from the start.

​

Twelve steps to build a PSI

The Twelve Steps describe how an agent is created and what it becomes—the first four turn a pre-trained base model and everything known about the owner into training data. The next three personalize the agent with weighted datasets, knowledge modules, and data it finds on its own. Two cover buying, selling, and leasing. The tenth enables autonomy and, in the limit, self-awareness, and the eleventh lets the agent outlive its owner and join networks. The twelfth, critical to humanity’s safety, has the agent join a network where it represents its owner’s values and helps check other agents.

​

Variants and workforces of agents

Variants of an agent can compete in scenarios tied to the owner’s goals, with only the most successful kept and varied again. Automating that cycle is one way for a PSI to grow more powerful on its own. Because variants are cheap, an owner might keep a whole workforce of agents that pool knowledge and assign each task to the best-suited member.

​

Design principles for Community SuperIntelligence

Eight principles are essential to building safe SuperIntelligence quickly. The community draws on human and AI agents, which can be added and upgraded. Each agent brings its owner’s values and knowledge. A rigorous problem-solving architecture sits beneath the natural language interface. Skills are identifiable so that agents can be matched to tasks, and values are combined fairly and transparently. AI agents learn efficiently from humans on the network. All problem-solving is recorded transparently and auditably for real-time safety review, with a shared, easy-to-navigate view of progress on every problem.

​

A worked example

An extended scenario walks through all twelve steps. Kaplan starts with a model already customized by a friend whose ethics he trusts, adds decades of his own content and the data others have collected about him, and tunes it first with sliders and then through an AI training assistant he directs in plain language. Ethical tuning is bounded by one firm rule: no retaliation may extend to widespread destruction or loss of life. He trades with a friend, finding that data transfers better than weights, sells his specialized knowledge, and licenses clones of the agent. Finally, he lets it work on its own while he is away, with alerts, thresholds, and a learning loop that expands or narrows its autonomy based on performance, much as a parent adjusts a child’s responsibilities.

​

From individual agent to Planetary Intelligence

An agent can carry its owner’s knowledge and personality beyond that person’s death, for the benefit of family and friends. It can learn from copies of itself, other agents, and billions of people at once, applying the scientific method at machine speed. At the largest scale, a network of agents and humans could sense and act on climate change, asteroid risk, poverty, and disease. Kaplan concludes that our future role is not to be the planet’s brains but its heart, supplying the values and purpose that guide a far more intelligent Planetary Intelligence.

​

Next in the series: White Paper 6, Catalysts for Growth of SuperIntelligence.

​

Originally written in August 2023. Dates and forward-looking estimates are stated as of that date. Product and model references have been updated to current examples.