The morning started brightly.
Each prompt to the AI agent gave you information, so you seemed to be making progress towards your goal.
But, frustration was mounting. Some part of the information was off; sometimes the mistakes were obvious, with other errors arriving subtly.
[Five hours later]
Frustration has turned to anger, both at yourself and the AI agent. What am I doing wrong? Why is this taking so long? Why won’t the AI agent give me the right information?
Though tactics on setup and prompting can help, the deeper issue is the way we think about the AI agent. We often think about the AI agent as this all-knowing, mystical other; as many have said before, a better frame is to think about the AI agent as a knowledgeable, overeager junior employee.
The reason for the better frame is deeply rooted in the work by Karl Popper, then extended by David Hestenes. Their ideas – extended again to include an AI agent – are part of the philosophical grounding when using AI agents.
By understanding the philosophy, your tactical work with AI agents will be more efficient and effective – helping you make a bigger difference for yourself and those you lead.
Note: I’m using the term “AI agent” to include all versions of the machine – from prompting into a chat window to building fully-meshed groups with different roles.
Popper – Three Worlds as a Philosophy of Knowledge
Karl Popper was a philosopher of science working in the mid-twentieth century. Popper’s most cited claim is falsifiability; however, there is a more important idea when working with AI.
In Objective Knowledge, Popper asked this question: If a person's mind holds a piece of knowledge, what happens to the knowledge when the person dies?
The obvious answer is that the knowledge dies too – unless the person recorded the knowledge, taught another person, or embedded the knowledge in an institution. Popper pushed further, arguing that externalized knowledge exists as a genuine object in the world. The externalized knowledge has structure and causal force, independent of any person with the knowledge.
To name this claim precisely, Popper proposed three worlds:
- The Physical World 1 – Physical objects and processes.
- The Personal World 2 – An individual’s subjective experience, beliefs, and knowledge and skills.
- The Conceptual World 3 – Humanity’s objective knowledge and skills that have been collectively discovered, extended, and applied.
Popper’s radical claim is that the ideas in the Conceptual World 3 are genuinely real. The structures in the Conceptual World 3 have causal power, acting on individuals – whether the individual does or does not know about the concept.
Consider how a framework like monthly recurring revenue (MRR) enters the personal world of a startup founder. The founder does not invent MRR; the relationship between MRR and business health exists independently – in the Conceptual World 3 – before the founder ever sees the framework. The framework in the Conceptual World 3 then acts on the founder’s Personal World 2: The founder encounters the framework and brings an understanding of MRR into their Personal World 2. By updating their Personal World 2 with the framework from the Conceptual World 3, the founder reorganizes their knowledge of business.
A “light gray font” metaphor makes Popper's radical claim tangible: Ideas in the Conceptual World 3 exist before anyone finds them, but the font is nearly invisible. When someone begins working on the structure seriously, the font darkens; when more people engage, the font darkens further. Eventually, the concepts have reorganized the Personal World 2’s of everyone in the domain – and the concepts that were once imperceptible are now the shared conceptual ground throughout the domain.
Popper's Three Worlds framework answers the question of how humans pass along knowledge and skills through concepts. However, there is an important question left unaddressed: How does a Personal World 2 actually make contact with ideas in the Conceptual World 3?
This is the question a physicist and physics educator named David Hestenes spent his career answering – and the answer turns out to matter as much for using an AI agent as for teaching a physics student.
Hestenes – Learning through the Modeling Cycle
David Hestenes was a physicist and physics educator who spent his career asking a specific question: Why do students who complete a full physics course still hold inconsistent beliefs about forces and motion?
Students in the physics course were doing experiments, memorizing definitions, and solving problems – all parts of a “good” physics course. To measure students’ understanding of forces and motion, Hestenes and his colleagues created a diagnostic. The expectation was that students would do well on the diagnostic, but the results were shockingly bad.
Many students had deep conceptual misunderstandings about forces and motion, such as “heavier objects fall faster” or “force is related to velocity.” The ideas for forces and motion in many students’ Personal World 2 were inconsistent with the scientific version in the Conceptual World 3 and the results from experiments in the Physical World 1.
To help students build consistent ideas for forces and motion, Hestenes and his colleagues created the modeling cycle. This approach used experiments in the Physical World 1 to create questions and initial ideas, which gave students a way to understand the scientific ideas in the Conceptual World 3. After working with the ideas in Conceptual World 3, students did more experiments to modify and extend the ideas in their Personal World 2 – continuing the cycle to build consistent ideas.
As an example of the modeling cycle, think about a small cart rolling on the floor (Physical World 1). Students measured positions and times, then graphed those data points. Students created a personal definition and equation for velocity (Personal World 2), comparing their definition and equation against the scientific version (Conceptual World 3). After practicing with the definition and equation, students used velocity to predict a specific position and time when rolling the small cart again – comparing their calculations (Personal World 2 and Conceptual World 3) with the results (Physical World 1).
By teaching with the modeling cycle, results from the diagnostic completely changed. Before, many students only recognized the symbols in the Conceptual World 3; the symbols did not connect to a deeper meaning, causing students to have deep conceptual misunderstandings. After using the modeling cycle, almost all students showed a strong understanding of the ideas in forces and motion. The interplay between the three worlds helped students build well-connected and organized conceptual models – sets of knowledge and skills that are explanatory and predictive.
Hestenes's modeling cycle answers the mechanism question Popper left open: A Personal World 2 makes contact with the Conceptual World 3 through repeated cycles of experience, grounded in the Physical World 1.
The combined ideas from Popper and Hestenes work for a single person learning a topic. However, we can extend the ideas into leadership: What if there is an Other?
The Four Worlds Model – Leadership for Humans and AI Agents
A leader is anyone responsible for an Other – whether that other is a plant, animal, person, or machine. The leader’s responsibility includes creating and maintaining a system for the other, helping the other flourish.
When an other is included, the three worlds extend through a second Personal World 2 – creating the Four Worlds Model:
- The Physical World 1 – Physical objects and processes.
- The first World 2 is the Leader’s Personal World 2 – The leader’s subjective experience, beliefs, and knowledge and skills.
- The Conceptual World 3 – Humanity’s objective knowledge and skills that have been collectively discovered, extended, and applied.
The second World 2 is the Other’s Personal World 2 – The other’s subjective experience, beliefs, and knowledge and skills.
Leadership is hard because the leader has to manage all the arrows!
When the other is a person, the leader must understand the other’s connections to the Physical World 1 and the Conceptual World 3 – plus correctly interpret any signals that come from the Personal World 2-to-Personal World 2 connection. Through the many different forms of leadership training, leaders can get better at managing the arrows to become expert leaders.
(This idea is the foundation of Belcher’s Dual-Layer Leadership Theory, which will be discussed in future essays.)
The image changes when the other is a machine – specifically an LLM-based AI agent. Assuming that the AI agent does not have a continuous stream of data, the connection between the AI agent and the Physical World 1 is broken. In addition, the over-the-top arrow that connects the leader’s Personal World 2 with the other’s Personal World 2 is broken. When the other is a person, the leader can sense verbal and non-verbal cues – such as when the other is nervous or confident. With an AI agent, there are no cues; the leader can only see the stated steps and output from the AI agent.
The single connection from the leader to the AI agent through the Conceptual World 3 creates an enormous challenge: An AI agent can only access the slice of the Conceptual World 3 that the leader’s Personal World 2 points towards.
Principles – Working with AI Agents
This challenge creates a few different principles for using AI agents.
First, the leader's Personal World 2 is constructed into conceptual models through cycles with the Physical World 1 and Personal World 3. Although the AI agent was trained on the written record of the Conceptual World 3 (symbols, equations, definitions, and explanations) the AI agent does not have access to the full set of ideas in the Conceptual World 3. The leader’s Personal World 2 has to constantly point the AI agent towards a slice of the Conceptual World 3, making the conceptual models in the leader’s Personal World 2 the chokepoint for the AI agent’s Personal World 2. The better the conceptual models from the leader, the better the conceptual models for the AI agent.
Second, a leader has to be part of the system. The technical version is that an AI agent’s Personal World 2 comes from a set of probability distributions in a slice of the Conceptual World 3, with no grounding in World 1. Because the AI agents do not have access to the Physical World 1, these agents cannot use the same modeling cycle as a human to develop conceptual models; there is no small cart in the Physical World 1 that the AI agent can use to predict and check an outcome. Even with many agents working together on different slices of the Conceptual World 3, a human leader still must manage each AI agent and the structure for coordinating the entire set of AI agents.
Third, AI agents can hold large slices of the Conceptual World 3 – giving the AI agents an ability to find interesting patterns, especially across massive data sets. However, the disconnect between the AI agent’s Personal World 2 and the Physical World 1 means that the AI agent cannot tell which patterns have explanatory and predictive power in the Physical World 1. The leader’s role is to connect the patterns to the Physical World 1, updating the slice of the Conceptual World 3 for the AI agent.
Popper’s three worlds, Hestenes’ application in teaching, and my extension are all part of the philosophical grounding when using AI agents. The challenge for using AI agents points back to the conceptual models in the leader’s Personal World 2: How strongly built are the conceptual models?
A leader with weak conceptual models points the AI agent toward a generic slice of the Conceptual World 3; a leader with robust conceptual models points the AI agent towards a precise and rich slice of the Conceptual World 3.
By pointing at a precise and rich slice, the leader with robust conceptual models will amplify their work – making a bigger difference for themselves and their others.
