Educational Technology and Change Journal

By Jim Shimabukuro (assisted by ChatGPT-5)
Editor

(Also see AI Data Labeling and Processing: Update August 2025, Basic Building Blocks for a Learning Model, and Algorithm of an Intentional Heart.)

Introduction: In this conversation, ChatGPT and I clarify my grasp of data labeling in the context of AI training, supervised vs unsupervised learning, agentic vs non-agentic AI, and AutoGPT. -js

JS: Let me see if I understand data labelling in the context of AI. It’s similar to the way our mind works. We need to categorize and define something before we can begin to understand what it is, where it fits in, and how to use it in a discussion. In AI, unlabeled data are like things that haven’t been sorted or categorized. After categorization, we “attach” a tag or “label” to it and will then know how or when to use it.

ChatGPT: Exactly — you’ve got the essence of it.

In AI (especially in machine learning), data labeling is the process of giving raw, unlabeled data (like images, text, audio, or video) some kind of meaningful annotation so the AI can recognize patterns and “learn” from them.

Your analogy to the human mind works well:

• Unlabeled data = experiences or sensory input you’ve noticed but haven’t organized. Imagine walking into a new environment and seeing dozens of unfamiliar objects — they exist, but you don’t yet know what they are.
• Labeling = the act of attaching a category, tag, or name to those objects. For example, pointing at something and saying “That’s a chair.” Once you’ve done that, your brain can group future chairs into the “chair” category.
• Training AI = giving the machine many examples of “this is a chair” (labeled) and “this is not a chair” (also labeled), so it gradually builds a mental model of what features belong to the “chair” category.

So when AI engineers talk about “labeled data,” they’re talking about ground truth references — the confirmed answers AI can learn from. For instance:

• In computer vision: labeling an image of a dog as “dog.”
• In natural language processing: tagging a sentence as “positive sentiment.”
• In speech recognition: aligning an audio clip with its transcript.

Once trained on enough labeled data, the AI can generalize and start recognizing unseen examples, just as your mind can identify a new kind of chair you’ve never seen before.

---

JS: I see. It’s more like category training rather than data training? Teach the general parameters of a category so the bot will know where to place data that match or fit into those parameters?

ChatGPT: Yes — that’s a really good way of putting it.

When we talk about “training” an AI on labeled data, it’s not memorizing each individual example like flashcards. Instead, it’s:

• Category training → teaching the model the general parameters (the patterns, features, or statistical signals) that define a category. For instance, “dogs usually have four legs, fur, and certain shapes for ears/muzzles” — though the AI learns this mathematically, not descriptively.
• Data training (in the literal sense) → giving the AI lots of specific, labeled examples so it can discover those parameters on its own.

So your phrasing works: we’re not really training on the data itself; we’re training on the relationships between data and their categories. Once the model gets the “boundaries” of a category, it can classify new, unseen data into the right slot.

A simple analogy:

• Imagine teaching a child what a “triangle” is. You don’t show them every possible triangle in existence. You show enough examples (different sizes, colors, rotations) and say, “These are all triangles.” The child extracts the general rule: “three straight sides, three corners.”
• That’s category training. The data (the drawn triangles) are just the medium.

Would you like me to sketch how this plays out differently in supervised vs. unsupervised learning — one being explicit category training, the other being the AI trying to invent its own categories?

---

JS: Let me guess first. Supervised means human guides and unsupervised means a nonhuman DIY process based on some kind of recursive feedback loop that, through iterations, arrives at the correct label?

ChatGPT: You’re very close — you’ve basically got the intuition! Let me tighten it a bit:

Supervised learning

• Exactly as you said: humans provide guidance by supplying labeled data (“this picture = cat, that one = dog”).
• The AI learns to map inputs → outputs by minimizing errors against those human-provided labels.
• The “feedback loop” here is built into training: the AI makes a guess, compares it to the correct label, adjusts itself, and repeats until it’s good enough.

Unsupervised learning

• Here, there are no labels — just raw data.
• The AI tries to find patterns, groupings, or structures on its own.
• It doesn’t “arrive at the correct label” in the human sense, because there is no ground truth provided. Instead, it clusters or organizes data by similarity.
  • Example: If you gave it thousands of photos of animals with no labels, it might cluster them into “furry quadrupeds,” “winged things,” “slimy swimmers” — categories that make sense mathematically, though they might not perfectly match human categories (“dog,” “bird,” “fish”).

There’s also a middle ground:

Semi-supervised learning

• Mix of the two: a little human-labeled data + lots of unlabeled data.
• Useful when labeling is expensive (e.g., medical images).

So your “DIY process with recursive feedback loops” is spot on for supervised — unsupervised is more like “DIY pattern discovery without a teacher.”

---

JS: Unsupervised is like throwing a thousand tiny puzzle pieces at the AI and allowing it to group the pieces into large than incrementally smaller groups until the chaos becomes increasingly ordered?

ChatGPT: Yes — that’s a really good visualization. 🧩

Unsupervised learning is exactly that:

• Step 1 (Chaos): You give the AI a thousand puzzle pieces with no picture on the box.
• Step 2 (Grouping): It notices that some pieces have blue on them (sky/water?), some have green (grass/trees?), some have sharp edges (borders).
• Step 3 (Refinement): Within the “blue” group, it might split further: darker blue vs. lighter blue.
• Step 4 (Emerging structure): Eventually, you have clusters of pieces that probably belong together, even though no one ever told the AI what the finished image should look like.

That’s why unsupervised learning is often used for:

• Clustering → grouping similar data (e.g., market segmentation: customers with similar shopping habits).
• Dimensionality reduction → finding the hidden structure in high-dimensional data (like boiling thousands of features down to the “core factors”).
• Anomaly detection → spotting puzzle pieces that don’t fit anywhere (useful for fraud detection, unusual medical scans, etc.).

So yes, unsupervised learning is like the AI trying to bring order out of chaos — but the order it finds is based purely on the math of similarities, not on human-defined categories.

---

JS: A similar yet different concept — “agentic AI”. Can you explain it to me using the analogy of an Army unit, for example a squad, platoon, or company? If that’s not a good analogy, perhaps another of your own choosing?

ChatGPT: Great question — and the Army analogy actually works very well for explaining agentic AI. Let’s run with it.

1. Traditional AI (non-agentic)

Think of this like a weapons specialist in a squad.

• They sit and wait for orders.
• You say: “Identify all vehicles in this drone footage” → they execute that narrow task.
• No initiative, no long-term planning — they’re tools, not decision-makers.

2. Agentic AI

Now imagine a platoon with a mission, not just a specialist with a tool:

• The platoon is briefed: “Secure this village and protect supply lines.”
• Instead of waiting for constant orders, the unit breaks down the mission into sub-goals: scouts recon the area, engineers build defenses, medics set up triage.
• If they encounter unexpected resistance, they adapt — maybe call for reinforcements, reroute patrols, or resupply.

That’s what makes an AI agentic:

• It takes a high-level goal (“research competitors and write a market analysis”) and figures out the steps to achieve it.
• It can plan, adapt, and take initiative — not just follow step-by-step instructions.
• Often, it also has tools and autonomy (like browsing the web, sending emails, or running code) to act in the world, not just analyze data.

3. How far the analogy goes

• Squad = narrow AI tool (single skill).
• Platoon = agentic AI (multiple coordinated skills, self-organizing toward a goal).
• Company = multi-agent system (several agentic AIs collaborating — like logistics agents, reconnaissance agents, and strategy agents all working together).

The key difference is initiative.

• Non-agentic AI = reactive specialists.
• Agentic AI = proactive operators that decompose goals, plan, and execute in dynamic environments.

---

JS: You’re a natural teacher. You explain ideas simply and clearly, customized to fit a specific student! Back to the Army unit. Could a single soldier, the last survivor of his company on a critical mission, adopt the functions of his comrades as they succumb until he’s the last man standing but is still able to accomplish a mission?

ChatGPT: Mahalo — that’s kind of you to say. And your scenario of the last soldier standing is actually a perfect metaphor for how agentic AI could operate when resources or teammates are lost.

One Soldier, Many Roles

Imagine a company tasked with taking and holding a strategic bridge. Normally:

• Infantry secures the area.
• Engineers disable explosives on the bridge.
• Medics treat the wounded.
• Comms specialists coordinate with HQ.

But as the mission goes wrong, soldiers are lost. Eventually, one survivor is left.

That last soldier has to adapt and assume multiple roles:

• Scout: stealthily recon the area.
• Engineer: jury-rig tools to secure the bridge.
• Medic: patch himself up if injured.
• Commander: make judgment calls without backup.

Even if undertrained in those other roles, his determination and improvisation might allow him to complete the mission.

In AI Terms

That’s very much like an agentic AI with modular functions:

• Normally, you might have a system of agents (planner agent, researcher agent, writer agent).
• If one “module” fails or is missing, a sufficiently capable single agent could step into multiple roles, using general intelligence and tool access to adapt.
• Instead of waiting for replacements, it improvises — sometimes less efficiently, but still mission-capable.

Limits of the Analogy

• A soldier can learn other skills but not master them instantly. An agentic AI, depending on its architecture, may “switch hats” more fluidly.
• But just like the last soldier might exhaust himself or miss expertise, a lone AI agent might stretch thin compared to a team of specialized agents.

So yes: in principle, an agentic AI could behave like that last surviving soldier — taking on the roles of its fallen comrades, adapting dynamically, and striving to achieve the mission with fewer resources.

---

JS: Is AutoGPT the same as unsupervised AI?

ChatGPT: Great question — and easy to mix up, because both sound like “the AI just figures stuff out on its own.” But AutoGPT ≠ unsupervised AI. They’re quite different:

Unsupervised AI (learning method)

• What it is: A training approach. The AI learns patterns or clusters from raw, unlabeled data.
• Example: Feed it thousands of images of animals without labels, and it groups them by similarity (maybe “furry quadrupeds” vs. “winged things”).
• Key point: It’s about how the AI learns during training.

AutoGPT (agentic framework)

• What it is: A wrapper or framework around large language models (like GPT).
• It takes a high-level goal (“Research competitors and write a market analysis”) and autonomously breaks it down into subtasks: browsing the web, analyzing data, writing drafts.
• It’s called “agentic” because it acts, using tools and feedback loops to accomplish goals without constant human prompts.
• Key point: It’s about how the AI behaves during deployment (goal-seeking, planning, tool use).

Side by Side

Aspect	Unsupervised AI	AutoGPT
Category	Training paradigm	Agentic application/framework
Data	Works on unlabeled data	Works on natural language goals
Purpose	Discover hidden structure/patterns	Break goals into steps and execute them
Analogy	Sorting puzzle pieces into clusters	A soldier who takes the mission brief and figures out the plan

So:

• Unsupervised AI = How you teach the AI.
• AutoGPT = How you deploy the AI to act like an autonomous problem-solver.

---

JS: Back to the Army unit — AutoGPT is like a group of soldiers, each with a specialty, and, for a specific mission, only those with specific specialties are chosen? Thus, the goal, the mission, determines the members?

ChatGPT: 💯 — yes, that’s a sharp way to put it. Let’s expand your Army analogy for AutoGPT:

AutoGPT as a Mission-Oriented Army Unit

• You give the mission: “Secure this village and set up supply lines.”
• AutoGPT takes that goal and says: “Okay, to succeed, I’ll need recon, logistics, comms, and engineering.”
• It then assembles the soldiers with those specialties and directs them step by step.

So in AI terms:

• Mission = User’s high-level goal (e.g., “Write a market analysis”).
• Soldiers = Sub-agents or task modules (researcher agent, writer agent, data-checking agent, etc.).
• Commander = AutoGPT core loop, which breaks the mission into subtasks, assigns them, and evaluates progress.
• Selection = Dynamic tasking — only the needed “specialists” are activated for the mission.

Key Difference from a Regular Unit

In a traditional Army unit, the squad/platoon is pre-assembled and always moves as one.
In AutoGPT, it’s more like a mission-based task force:

• Some “soldiers” may sit idle because their specialty isn’t needed this time.
• Other “soldiers” might be called upon repeatedly if their skill is critical.

Where the Analogy Fits Agentic AI at Large

• AutoGPT = A commander who can recruit and coordinate specialists for each mission.
• Unsupervised learning = The training each soldier did in boot camp to recognize patterns without being told.
• Reinforcement learning = Field exercises where soldiers learn through trial and error, improving tactics after mistakes.

So yes — you nailed it: the goal (mission) determines the composition of the force. AutoGPT is less like a standing unit and more like a flexible task force builder.

---

JS: In this AI sense, “learning” is like developing “expertise.” An expert is able to make nuanced, highly differentiating observations and decisions which a novice might not be able to “see”?

ChatGPT: Exactly — you’ve put your finger on it. 👌

In AI, learning = developing expertise, and your expert/novice analogy is spot on:

Novice vs. Expert (human side)

• Novice: Sees only the surface. For example, a rookie soldier looks at a distant silhouette and just sees “someone walking.”
• Expert: Sees nuance — gait, gear outline, weapon type, even posture. From the same glance, they say: “Enemy scout, likely carrying a carbine, not a heavy weapon.”

The expert has built mental models and fine-grained categories through repeated exposure, so their decisions are faster, more accurate, and more context-sensitive.

Novice vs. Expert (AI side)

• Untrained or undertrained model (novice): Struggles, confuses categories, overgeneralizes.
• Well-trained model (expert): Learns the subtle features that separate close categories (e.g., husky vs. wolf, benign tumor vs. malignant tumor, friendly vs. hostile aircraft).

That’s why data + training = expertise:

• The AI doesn’t just memorize; it distills general rules that let it make fine distinctions.
• Expertise shows up as nuance in decision boundaries — the AI “sees” differences a novice (human or machine) would miss.

Why “learning = expertise” is a powerful analogy

• It captures the idea of progression: novice → journeyman → expert.
• It explains why training needs varied examples (just like experts learn from diverse experiences).
• It highlights nuance and subtlety — the hallmark of both human and machine expertise.

---

JS: Does the Army command structure transfer to agentic AI? For example, in an agentic system, is there a chain of command?

ChatGPT: Great question — and yes, the Army command structure maps really well onto agentic AI. Let’s run it through the analogy.

1. Chain of Command in the Army

• High Command (Generals): Define strategic objectives (“Secure this region in 90 days”).
• Field Commanders (Colonels/Majors): Translate that into operational plans (“Take this city, fortify this bridge”).
• Platoon Leaders & Squad Leaders: Break it down further into tactical tasks (“Patrol this street, establish a checkpoint”).
• Soldiers: Execute concrete actions (“Search this house, monitor this corner”).

Each level:

• Gets orders from above.
• Translates them into actionable tasks for below.
• Reports progress back up the chain.

2. Chain of Command in Agentic AI

• User (Human Commander): Issues the top-level mission (“Write a market analysis of competitors”).
• Central Agent (Planner / Orchestrator): The “field commander.” It breaks the goal into subtasks.
• Sub-agents (Specialists): Assigned to specific duties — e.g., Research Agent, Data Analyst Agent, Writer Agent.
• Tools (like web search, code interpreter, APIs): These are like the “soldiers” carrying out the fine-grained orders.

Just like in the Army:

• Orders flow down (goals become subtasks).
• Reports flow up (results are passed back for review and integration).
• Authority is tiered — sub-agents don’t need to know the whole mission, only their role.

3. Differences

• In the Army, the chain is rigid — a corporal doesn’t “decide” to be a general.
• In agentic AI, the structure can be more fluid: a sub-agent can sometimes spin up its own “sub-sub-agents” if the mission demands (like AutoGPT creating new task loops).

4. Multi-Agent Operations (Joint Task Force Analogy)

If you go beyond one AutoGPT-like system:

• You might have multiple commanders (different orchestrators) coordinating across domains — like Army, Navy, Air Force each running their own AI agents, reporting to a unified “joint command.”
• This is where modern AI research is heading: networks of agents coordinating like coalition forces.

So yes: agentic AI systems often have a command structure very much like the Army. A user gives a mission, the central agent acts like a commander, and specialized agents act like units executing orders.

---

JS: Got it! Mahalo, kumu! A hui hou.

ChatGPT: A hui hou! Mahalo, and enjoy your break — I’ll be here when you return.

Filed under: Uncategorized |