SI WHITE PAPER 3: HUMAN-CENTERED AGI
ABSTRACT: Human-Centered Artificial General Intelligence
For it to be effective, the safest path to AGI must also be the fastest.
​
This approach is the fastest because it begins with a network of human problem solvers, who by definition can perform any intellectual task as well as the average human. AI agents customized by individual people, called AAAIs, then join the network. Humans and agents share a single problem-solving architecture that is rigorous, transparent, and auditable, and any language model can be plugged in and upgraded as better ones emerge.
​
At first, humans do most of the work, especially setting goals and representing problems. Over time, the agents take on more of it, and human attention shifts to ethics, safety, and oversight. Because ethics and safety checks are built into the architecture itself, the system stays aligned with human values even when it works far faster than people can follow, and humans can always see how it reached a decision.
​
Values cannot be logically derived by any intelligence, however fast, so they remain human. And because the network can be built quickly, it gains a first-mover advantage, making the safest path the one that prevails.
​
Below are common questions and a more in-depth summary of the paper.
COMMON QUESTIONS: Human-Centered Artificial General Intelligence
Do we have to wait for a bigger AI model to get AGI?
No. White Paper 3 argues that AGI can begin with a network of people working alongside AI agents, all coordinated by a single shared problem-solving method. By definition, humans can do any intellectual task an average person can, so the network counts as generally intelligent from the day it launches. AI agents add speed and scale as they get better.
​
How big is the risk from advanced AI?
In May 2023, Dr. Kaplan put the chance that misaligned AI could lead to human extinction at 20 percent. He also said the first AGI could improve itself quickly, which makes the race plausibly winner-take-all. So whoever gets there first matters. The paper's position is that the first system must be safe and come first.
​
Why not just lock down or regulate powerful AI?
By 2023, capable language models were already open source and in the hands of millions. Autonomous agents were already setting their own goals. You can't guard these models like plutonium. Regulation also tends to slow down the careful developers while everyone else keeps building. The paper argues that safety has to come from how the system is built, not from keeping AI out of people's hands.
​
Why start with people instead of starting with AI?
The usual approach is to build a powerful AI first and add human values afterward. This paper reverses that order. It starts with a network of human problem solvers and adds AI gradually, teaching it human values along the way. People never leave the loop. Kaplan's PredictWallStreet work showed that millions of ordinary investors, coordinated well, could beat top hedge funds.
​
Do people need to learn a new way of thinking to take part?
No. The system is based on Newell and Simon's 1972 theory of human problem solving, which describes problem solving as searching through possible paths toward a goal. That structure is precise enough for machines and natural enough for people. Language models turn ordinary requests into a formal structure. The shared record of every problem the network works on is called the WorldThink tree.
​
Does this only work if everyone behaves well?
No. Participants only have to act in their own rational self-interest, and that rules out destroying everyone. Each person trains their own AI agent on their own values. This spreads ethical authority among everyone involved rather than leaving it to a small group to write rules. Recent behavior counts for more than past mistakes, and reputation and shared norms act as further checks.
​
How does the system stay safe when it thinks faster than people?
Ethics and safety checks are part of the design. They run every time the system sets a goal or sub-goal, so alignment keeps up, however fast the system works. People can always see how a decision was made. The record can be kept on a blockchain or another tamper-proof log so that no one can quietly rewrite it.
​
What would this look like in practice?
The paper describes a website where people create, own, train, and license their own AI agents. Together, those agents carry both the intelligence and the values of the people who trained them. The result is a SuperIntelligence that is democratic and human-centered because of how it is built, not because anyone added those values afterward.
SUMMARY: Human-Centered Artificial General Intelligence
White Paper 3 argues that the safest path to AGI must also be the fastest, and that both conditions can be met now rather than years from now. AGI does not require waiting for an ever-larger “Uber-LLM” to train itself into general intelligence. A network of human problem solvers already meets the definition, since humans can, by definition, perform any intellectual task as well as the average human. Add AI agents to that network, coordinate everyone through a single shared problem-solving architecture, and the network will be generally intelligent on the day it goes live.
​
The stakes, as of 2023
Misalignment could end in human extinction, and Kaplan put that risk at 20 percent in May 2023. Guarding powerful models like plutonium was no longer an option, because capable language models were already open-sourced. In the hands of millions, autonomous agents were already setting their own goals. Regulation slows the careful while others keep going. And because the first AGI can improve itself quickly, the race is plausibly winner-take-all, which makes being first as necessary as being safe.
Build the human network first
The paper inverts the usual order. Instead of building a powerful AGI and trying to add human values afterward, it builds a human collective intelligence. It gradually adds AI, training it on human values as it goes. The AI takes on more of the work over time, but humans never leave the loop. Kaplan grounds the idea in his PredictWallStreet work, where millions of ordinary investors, properly coordinated, outperformed top hedge funds.
Why does the architecture have to be rigorous?
Machines can match or exceed human reasoning without sharing human empathy or common sense, which the paper calls the sociopathic AI problem. A loosely specified system invites exactly the kind of error that such an agent will not catch on its own. A rigorous shared protocol prevents those errors, enables solutions to be learned automatically, scales to any intellectual task, allows any human or AI solver to plug in, and maintains a transparent, auditable record of every step of reasoning.
How people and agents think together
The architecture is Newell and Simon’s 1972 theory of human problem solving, which treats all problem solving as search through a problem space, represented as a decision tree. It is rigorous enough for machines and natural enough for people, who never need to learn the theory, because language models translate ordinary requests into the formal structure. The combined record of every problem the network works on is the WorldThink tree. The approach complements deep learning rather than replacing it, and new language models can be plugged in as they improve.
Who does what, and how that changes
People are especially good at framing a problem and deciding what the goal should be; machines are good at quickly exploring many possible paths. So, in the early division of labor, humans set goals and represent problems, while agents search for solutions. As the agents learn from that work, they take on more of it, and human attention shifts toward ethics, safety, and oversight.
Safety that does not require saints
The network does not depend on everyone being good. It only needs participants to act in their own rational self-interest, which rules out destroying everyone. Recent behavior counts for more than past mistakes. Each agent is trained on its owner’s values, which spreads ethical authority among everyone taking part rather than concentrating it in a small group writing a constitution, an arrangement Kaplan argues is open to corruption. Reputation, system rules, and shared norms add further checks.
​
Checks at the speed of thought
The decisive safeguard is that ethics and safety checks are part of the architecture itself. They run each time a goal or sub-goal is set, so the system stays aligned, however fast it works, and humans can always see how a decision was reached. The record can be kept on a blockchain or another tamper-proof log, so it cannot be quietly rewritten.
​
What it looks like in practice
The paper closes with a website where people create, own, train, and license their own AI agents. Combined, those agents carry both the intelligence and the values of the people behind them, producing a SuperIntelligence that is democratic and human-centered by construction.
​
Where this leads next
White Paper 4, Systems and Methods for Safe, Scalable AGI, takes up the problem this one leaves open: how to combine the ethical judgments of millions of people into a sound, representative sample of human values.
​
This paper preserves the argument as originally written in May 2023. Dates and forward-looking estimates, including the 2025 timeline, are stated as of that date.