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.
​
The preferred implementation of AGI is the fastest method for achieving AGI because it begins with a network of human problem-solving agents, who, by definition, can perform any intellectual task as well or better than the average human. AI agents trained and customized by individual humans (AAAIs) are introduced to the network as AI problem-solving agents. The human and AI agents share a common problem-solving architecture that is rigorous, scalable, transparent, auditable, safe, and powerful.
​
This architecture supports automatic learning and self-improvement. It is compatible with LLMs, which can be “plugged in” to the network and upgraded as more powerful LLM models become available. The AGI network begins with humans doing most of the problem-solving work, especially the most important aspects, such as setting goals, and the most difficult, such as representing the problem. Over time, AAAIs do more and more of the actual work, more effectively and efficiently than humans could, while human attention is increasingly directed to issues of ethics, safety, and oversight.
​
Because ethics and safety checks are built into the architecture itself, as the speed of problems increases far beyond the capability of humans to “keep pace,” the system remains aligned with human values and ethics. At any time, humans can see exactly how the system is making decisions, including all ethical information.
The AGI network is highly scalable and will become more powerful over time, yet the fundamental values and ethics of the system, which cannot be logically derived by any intelligence, no matter how smart or fast, remain aligned with human values. Thus, this approach solves the alignment problem in a democratic and scalable manner.
​
The fact that the AGI network can be implemented rapidly, far faster than estimates for when AGI will develop from other approaches, ensures a first mover advantage that allows this safest path to AGI to dominate other approaches, thereby fulfilling the two key requirements for the approach, namely that it be not only the fastest path to AGI but also the safest.
SUMMARY: Human-Centered Artificial General Intelligence
This paper 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. Craig A. Kaplan's central claim is that 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 of AGI, since by definition, humans can perform any intellectual task as well as the average human. Add AI agents to that network, coordinate everyone through a shared problem-solving architecture, and you have AGI on day one.
​
The paper opens with the stakes. Misalignment could lead to human extinction, and Kaplan estimated the risk at 20% as of May 2023. The "guard it like Plutonium" reasoning fails because powerful LLMs are already open-sourced and in the hands of millions, and autonomous-agent systems are already setting their own goals. Regulation only slows the cautious, while others race ahead. Because the first AGI can improve itself exponentially, the situation is plausibly winner-take-all, which makes being first and being safe equally necessary.
​
The proposed solution inverts the usual order. Rather than build a powerful AGI and bolt on human values afterward, you build a human collective intelligence system and add AI to it incrementally, training the AI on human values as you go. Over time, AI does more of the work, and human attention shifts to ethics, safety, and oversight, but humans never leave the loop, and their values are built into the system from the start. Kaplan grounds this in his PredictWallStreet work, where millions of ordinary investors, properly coordinated, outperformed top hedge funds.
​
Section 1 explains why the shared architecture must be rigorous: machines lack human empathy and common sense (the "sociopathic AI" problem), so loose specification invites catastrophic error. A rigorous protocol also yields five benefits: it avoids unintentional error, enables automatic learning, enables scalability to any intellectual task, enables modularity and stability so any human or AI solver can plug in, and maximizes safety through a transparent, auditable record of every thought step.
​
Section 2 provides the technical foundation: Newell and Simon's 1972 theory of Human Problem Solving, which models all problem-solving as search through a problem space represented as a decision tree. This framework is rigorous enough for machines yet natural for humans, who never need to know the underlying theory because LLMs translate natural language into the formal specification. The full problem tree is called the WorldThink Tree. Humans excel at problem representation and goal-setting, so early division of labor has humans framing problems while AI explores solution paths. The approach is complementary to deep learning, not a replacement, and new LLMs can be plugged in as they improve.
​
Section 3 addresses how human and AI behavior shape safety. Safety depends partly on solver values. Still, the network does not require saints; it only requires solvers to act in rational self-interest, which excludes destroying everyone. Recent behavior matters more than past sins. AAAIs are trained on their owners' values, which democratizes ethics rather than concentrating it in a small group writing a "constitution," an approach Kaplan argues risks corruption. Reputation, system rules, and norms provide further checks. Critically, ethics and safety checks are built into the architecture itself and run at every goal and subgoal, so the system stays aligned no matter how fast it thinks. Blockchain logs can make the record tamper-proof. A closing implementation example describes a website where users create, own, train, and license customized AI agents, whose combined intelligence and values produce a democratic, human-centered SuperIntelligence.
​
The paper preserves the argument as filed in the May 2023 provisional patent application, with dates and forward-looking estimates (including the 2025 timeline) stated as of that filing. It closes by pointing to White Paper 4, which takes up the unsolved problem of combining ethical judgments from millions of people into a sound, representative sample of human values.