Debunking the Mass AI Training Myth: Insight Is What Actually Matters

Opinion | September 24, 2026

Author: David Tang, Managing Director, Standard Kepler

Originally published in HKET (香港經濟日報)

Executive Summary

The Hong Kong government plans to spend $50 million on AI training for the general public. A noble intention — but one built on a fundamental misunderstanding of how AI skills actually develop. YouTube already has more AI tutorials than anyone could watch in a lifetime. Government-led courses, by their nature, cannot keep pace with AI’s weekly evolution. And the real competitive advantage in the AI era is not technical proficiency — it is insight (洞察力). This report traces the rapid evolution of AI interaction paradigms, from vibe coding to graph engineering, and argues that Hong Kong should invest in cultivating judgment, not teaching buttons.

The $50 Million Problem

Government training programs operate on an industrial-era model: standardize curriculum, certify completion, deploy workforce. This model fails for AI because:

YouTube already contains everything a government course could teach — and it updates daily. The marginal value of a government-certified AI course is near zero for anyone with internet access and motivation.

The Six Stages of AI Interaction

AI interaction paradigms have evolved faster than any educational curriculum could track. Here is the progression:

Stage 1: Vibe Coding (Early 2025)

“Tell the AI what you want, let it write the code.” Pure natural language programming. No technical knowledge required. The output: functional but fragile code that falls apart in production.

Stage 2: Prompt Engineering (Mid 2025)

The realization that vague instructions produce vague outputs. Practitioners learned to craft detailed prompts with context, examples, and constraints. The “prompt engineer” job title briefly existed — then disappeared as models became better at understanding intent.

Stage 3: Context Engineering (Late 2025)

The insight that what matters is not the prompt, but the context window. The most productive AI users became experts at feeding the right documents, codebases, and background information into the model. Prompts became shorter; context became richer.

Stage 4: Harness Engineering (Early 2026)

The recognition that single-model interaction is limiting. “Harness engineering” means orchestrating multiple AI systems — routing tasks to specialized models, chaining outputs, building feedback loops. The skill shifts from talking to AI to designing AI systems.

Stage 5: Loop Engineering (Mid 2026)

The current frontier: building self-improving loops where AI systems evaluate their own outputs, identify errors, and iterate. This requires not just technical skill but meta-cognitive judgment — knowing what “good” looks like and how to evaluate it.

Stage 6: Graph Engineering (Emerging)

The next evolution: knowledge graphs as the interface layer. Instead of linear conversations, AI interactions become network traversals — moving through connected concepts, data sources, and reasoning chains. This requires spatial reasoning and systems thinking, not typing speed.

By the time a government curriculum reaches Stage 3, the industry will be at Stage 6.

What Actually Matters: Insight

Technical skills in AI have a half-life of months. What does not decay is insight — the ability to:

1. Frame the Right Problem

AI can solve problems brilliantly. But if you ask it to solve the wrong problem, you get a brilliant wrong answer. The scarcest skill is knowing what question to ask.

2. Evaluate Output Quality

AI generates plausible-sounding nonsense with confidence. The ability to distinguish genuine insight from sophisticated garbage is judgment, not technique.

3. Understand System Boundaries

Every AI system has limits — what it can do, what it cannot, where it fails. The productive user knows these boundaries and designs around them. The naive user crashes into them.

4. Integrate Across Domains

The most valuable AI applications happen at the intersection of fields: law + AI, medicine + AI, finance + AI. Deep domain expertise + AI literacy is worth infinitely more than AI literacy alone.

The Right Investment

If Hong Kong wants to prepare its workforce for the AI era, the investment should target:

What to Fund Instead

1. Cross-domain AI labs: Pairing domain experts (doctors, lawyers, engineers) with AI tools to solve real problems
2. Output evaluation competitions: Training people to spot AI hallucinations and errors
3. Problem-framing workshops: Teaching the skill of asking the right question
4. Open challenge platforms: Government-sponsored AI application contests with real-world problems

The Global Context

Other jurisdictions are making similar mistakes:

No country has cracked the code. The fundamental challenge is that AI capability grows exponentially while educational institutions operate linearly.

Conclusion: Teach Fishing, Not Fish

The AI training debate reveals a deeper misconception: the belief that AI is a tool to be learned, like Excel or Photoshop. It is not. AI is a cognitive amplifier — it magnifies whatever the user brings to it. A person with insight becomes superhuman. A person without insight produces sophisticated nonsense faster.

Hong Kong’s $50 million would be better spent cultivating the underlying capability that AI amplifies: critical thinking, domain expertise, and judgment.

In the AI era, technique is cheap. Insight is scarce. And scarcity is what commands value.

The government cannot teach AI — it changes too fast. But it can teach the timeless skill of thinking clearly. That is the only investment with a positive return in an exponential world.

Standard Kepler Research | standardkepler.com