Optimal Learning Strategies in the AI Era: Foundations First or Learning by Doing?

The Universal Learning Challenge in Modern Context

"How should we learn?"

This is the fundamental question I consistently return to after years of running educational programs and developing talent.

In technical domains, should a beginner who wants to create web applications with Next.js progress step-by-step through HTML, CSS, JavaScript, deployment, React, Next.js, and TypeScript? Or should they start directly with Next.js+TypeScript+Tailwind CSS, building something without fully understanding the components?

The same question applies to business skills. When learning data analysis, should one begin with statistical fundamentals, then progress to Excel, SQL, and visualization tools? Or tackle actual business challenges directly, researching necessary knowledge as needed?

In marketing too, should one study customer psychology and traditional frameworks before implementation, or learn through trial and error by running small-scale campaigns?

In our current era, this universal question has acquired a new dimension. We now operate under the assumption that we can ask ChatGPT for answers. When the environment changes, the optimal approach changes too. With this reality, how should we approach learning?

This question isn't limited to specific fields. It's a universal inquiry about the optimal sequence and method for learning in any specialized domain. Drawing from my experience running an AI education business and collaborating with AI systems, I want to share evidence-based insights to help us consider the future based on facts.

The Essence and Contrast of Two Learning Paradigms

Let's first clarify the characteristics of two primary learning approaches.

図表を生成中...

The step-by-step structured learning approach is a "bottom-up" method that progresses from fundamentals, understanding small components before moving to the whole. This traditional approach covers foundational knowledge comprehensively before advancing to applications, offering a systematic curriculum that allows beginners to build skills with certainty and minimal confusion.

• Example in technology: HTML/CSS → JavaScript → React → Next.js in sequence
• Example in business analysis: Statistical theory → Excel skills → SQL data extraction → Visualization → Analysis report creation
• Example in marketing: Marketing principles → Customer psychology → Advertising theory → Channel characteristics → Campaign design

On the other hand, the practice-first learning approach is a "top-down" method that begins with actual projects or business challenges, creating functional products without necessarily understanding all details. Learning while building maintains high motivation and achieves practical outcomes more quickly. In the AI era, one can resolve uncertainties by asking ChatGPT or similar tools.

• Example in technology: Starting directly with building a Next.js application while researching necessary knowledge as needed
• Example in business analysis: Creating analysis reports from actual business data while learning required analytical methods on the go
• Example in marketing: Running small-scale campaigns and improving subsequent initiatives based on result analysis

The fundamental difference lies in the sequence of learning and sources of motivation. The step-by-step approach prioritizes certainty of understanding, while the practice-first approach leverages proximity to outcomes as motivation.

This dichotomy is common in educational debates, but neither approach is absolutely superior. Each is based on different cognitive mechanisms, and appropriate application depends on the situation and purpose.

Insights from Learning Science: Knowledge from Educational Technology and Cognitive Psychology

To evaluate these approaches scientifically, insights from cognitive psychology and educational technology are valuable.

図表を生成中...

Guidance for Beginners and Cognitive Load Theory

From a cognitive psychology perspective, beginners benefit from certain levels of instruction and guidance. Research by Kirschner et al. (2006) suggests that "minimally guided instructional approaches (such as discovery learning and problem-based learning) are less effective and efficient than approaches with substantial guidance"¹.

This is based on cognitive load theory. When beginners tackle complex problems directly, excessive demands on working memory can impede efficient learning. Since working memory capacity is limited, beginners benefit from environments that allow them to focus on basic concepts and procedures.

Effectiveness of Project-Based Learning

However, educational technology also emphasizes learner activity and contextual importance. Meta-analyses of Project-Based Learning (PBL) effectiveness have shown that PBL significantly improves learning outcomes in academic achievement, motivation/attitudes, and thinking skills compared to traditional lecture formats².

PBL has been reported to contribute particularly to developing 21st-century skills (problem-solving and collaboration). Learning in practical contexts also enhances knowledge transferability (the ability to apply knowledge in different situations).

"Desirable Difficulties" and Long-Term Learning Effects

Another important concept in learning science is "desirable difficulties". This suggests that incorporating moderate challenges into the learning process may increase short-term struggle but enhances long-term memory retention and knowledge transfer³.

For instance, research indicates that learning is more effective in the long term when learners must organize and connect information themselves rather than receiving everything in a pre-organized state. In this respect, the trial and error inherent in practice-first approaches may function as "desirable difficulties."

Learning from the Knowledge Hierarchy Perspective

From the perspective of the DIKIW model (Data-Information-Knowledge-Intelligence-Wisdom) discussed in my previous article "Has AI Stolen Human Intelligence?", step-by-step learning primarily builds the "knowledge" layer, while practice-first learning directly approaches the "intelligence" layer (the ability to apply knowledge practically).

The concern, especially in the AI era, is the risk of degradation in human thinking abilities and problem-solving skills when outsourcing parts of "intelligence" to AI. Rather than merely creating a dependency where "AI provides answers," it's important to develop a relationship where AI helps enhance our own thinking.

Shocking Empirical Research: New Insights on Learning Effectiveness in the AI Era

In 2024, a team led by Bastani at the University of Pennsylvania's Wharton School published groundbreaking research on AI-assisted learning⁴. This study, which rigorously examined AI's impact on mathematics learning among approximately 1,000 high school students, provides crucial insights for considering learning methods in the AI era.

Experimental Design and Surprising Results

The experiment divided students into three groups with different AI usage conditions:

1. Unrestricted AI use group: Students who worked on tasks using a standard ChatGPT interface
2. Constrained AI use group: Students who used AI with guardrails that provided incremental hints as intended by teachers
3. Non-AI control group: Students who learned without using AI, following traditional methods

The results were striking. While performance improved during AI use (48% improvement in the unrestricted group, 127% in the constrained group), when AI access was removed for testing, the group that used AI freely performed 17% worse than the control group that never used AI.

However, the group that used AI with guardrails that minimized hints showed almost no such negative effect. This demonstrates that learning outcomes vary significantly depending on how AI is used.

Risks of AI Dependence and Countermeasures

This research suggests that unrestricted AI use may compromise long-term learning capabilities in exchange for short-term efficiency gains. When relying completely on AI for tasks that should involve personal thinking, one may "solve" tasks superficially without developing fundamental understanding or thinking skills.

Bastani's research suggests that the optimal learning method in the AI era is neither avoiding AI entirely nor using it without restrictions, but rather AI utilization with appropriate guardrails.

Implications for Programming and Business Skill Learning

These insights are deeply relevant to programming and business skill learning. Excessive dependence on AI in a practice-first approach may lead to the superficial appearance of completed applications or business reports without developing fundamental understanding or application capabilities.

For example, when encountering programming errors, immediately requesting AI correction without considering why the error occurred fails to develop debugging skills or fundamental language understanding. Similarly, in marketing strategy development, adopting AI-generated plans without examination doesn't cultivate market understanding or strategic thinking.

Effective AI-assisted learning requires elements such as:

1. Incremental hints: Providing hints gradually rather than complete answers
2. Requesting thought processes: Asking for explanation of reasoning rather than just answers
3. Iterative feedback: Providing adaptive support based on learner responses
4. Maintaining appropriate difficulty: Not removing all barriers but preserving moderate challenges

Historical Evolution of Learning Methodologies: Co-evolution of Technology and Paradigms

Learning methods have evolved with time. Approaches to acquiring specialized knowledge have transformed significantly with changes in technology and social environments. Reviewing the evolution over the past 30+ years provides important insights for considering learning in the AI era.

Changes Seen Through Vocational Training and Learning Effectiveness Data

With the evolution of learning methods, measurement of their effectiveness has also advanced. Recent data on programming school effectiveness is interesting. Early 2025 surveys show that 85% of programming school participants achieved at least half their goals, and approximately 70% reported income increases⁵.

However, I believe these are somewhat superficial success indicators. As I pointed out in "AI Collaboration Era Talent Investment Strategy", short-term indicators like "found a new job" or "increased income" don't necessarily represent essential success. What's important is whether one has built expertise that continues to create value in the long term.

The Reality of AI Tool Adoption

The spread of AI development support tools is advancing rapidly. 2024 surveys show 69% of developers have tried AI assistants like ChatGPT or Copilot, and about 49% use them daily⁶. Amid these changes, the eternal theme of balancing fundamentals and practical skills continues in transformed ways. In the AI era, judgments about "what to understand personally" versus "what to delegate to AI" have become increasingly important.

Optimal Learning Strategies from the Knowledge Hierarchy Perspective

Let's consider both learning approaches from the perspective of the DIKIW model (Data-Information-Knowledge-Intelligence-Wisdom) discussed in my previous article "Has AI Stolen Human Intelligence?". The following diagram summarizes the relationship between the DIKIW hierarchy and learning approaches.

図表を生成中...

The step-by-step learning approach primarily focuses on building the "knowledge" layer, while the practice-first approach directly addresses the "intelligence" layer (knowledge application, pattern recognition, problem-solving). Importantly, contemporary AI largely covers the "knowledge" layer and is beginning to extend into parts of "intelligence."

The optimal learning strategy in the AI era involves clearly distinguishing between parts that can be delegated to AI and parts that humans should actively manage. The "knowledge" layer can be efficiently acquired with AI support, while humans should deeply engage with the "intelligence" and "wisdom" layers.

Hybrid Iterative Learning Model

As Bastani's research suggests, effective learning comes not from complete dependence on AI but from AI utilization with appropriate guardrails. Based on this, I propose an effective learning strategy: the "Hybrid Iterative Learning Model."

図表を生成中...

This model repeats small cycles of basic knowledge and project implementation, using AI as a support tool. The key is rapidly cycling through "basics → practice → reflection." The optimal solution in the AI era lies in simultaneously developing fundamental and practical skills within these rapid cycles.

Building Expertise in 3000 Hours: Learning from Rule of Thumb

Deep expertise building requires certain time and immersion. From experience, I believe it takes about "3000 hours" of concentrated effort to become an expert in a field. This benchmark, often mentioned in research communities, can be achieved by working 10 hours daily, 25 days monthly for 250 hours, continued for 12 months.

This 3000-hour journey isn't smooth. While the first few hundred hours might feel productive, many people give up during the "plateau" between 500 and 1500 hours. However, expertise begins to take shape after 2000 hours, and by 3000 hours, tangible expertise has developed.

If you want to work professionally, you should first reconsider your approach and commit to at least one year of training without expecting immediate results. If you can immerse yourself in the same pursuit for 3000 hours, it will become your strong ally.

Optimal Strategies by Learner Type: Tailored Learning Approaches

To maximize learning effectiveness, strategy selection should align with learner characteristics and goals. Let's consider optimal approaches for different types.

For Complete Beginners: Step-by-Step Approach with Consideration for Cognitive Load

For absolute beginners, starting with a step-by-step approach that minimizes cognitive load is effective. In completely new fields, scaffolding is necessary to understand basic concepts and terminology.

Example in programming:

1. First learn HTML/CSS basic concepts and syntax (2-3 weeks)
2. Understand JavaScript fundamentals (1-2 months)
3. Work on small projects (e.g., simple TODO app)
4. Reflection and reinforcement of basics
5. Proceed to next step (React, etc.)

Example in business analysis:

1. Learn data analysis basic concepts and Excel fundamentals
2. Tackle simple business data analysis
3. Interpret and reflect on analysis results
4. Progress to more advanced tools (Tableau, etc.)

For Career Changers: Strategy Leveraging Transferable Skills

For those with expertise in other fields venturing into new areas, approaches that leverage existing skills and thinking methods are effective.

Identifying transferable skills: First identify skills and thinking methods from existing expertise transferable to new fields. For example, people with scientific backgrounds learning programming can leverage logical thinking and process decomposition abilities.

Practice-oriented approach: With existing thinking frameworks, one can engage in practical projects relatively early. However, field-specific fundamental concepts should still be thoroughly understood.

Using AI to complement knowledge gaps: Efficiently fill gaps in terminology and contextual understanding using AI while progressing.

For Practitioners: Rapid Iteration to Enhance Problem-Solving

For those with basic knowledge seeking advanced expertise, rapid iteration directly linked to problem-solving is effective.

• Selecting challenging projects: Work on projects slightly beyond current capabilities
• Comparing multiple solutions: Try multiple approaches to the same problem to understand respective advantages and disadvantages
• Expert feedback: Receive regular feedback from communities or mentors
• Systematic reflection: Systematize learnings after project completion to deepen conceptual understanding

For Leaders and Managers: Integration of Comprehensive Understanding and Practical Wisdom

For those in leadership positions, approaches that integrate technical details with broader perspectives are necessary.

• Grasping the big picture: First understand the overall map of the field and relationships between elements
• Analyzing diverse case studies: Learn from success and failure cases to recognize patterns
• Learning through teaching: Deepen understanding by teaching team members
• Improving decision-making processes: Apply specialized knowledge to actual decisions and learn from results

Concrete Learning Strategies to Maximize Results

Based on empirical research and theoretical considerations, I'd like to propose concrete approaches to maximize outcomes (income improvement and career development).

Implementing the Rapid Iteration Learning Model

The practical steps for implementing a learning model that rapidly cycles through "basics → practice → reflection" are as follows:

1. Setting small goals: Set clear learning goals achievable in 2-4 weeks
2. Minimal fundamental understanding: Understand the minimum basic concepts necessary for that goal
3. Small-scale project implementation: Apply learned concepts in actual projects
4. Reflection and knowledge systematization: Organize insights gained through practice to deepen systematic understanding
5. Expanding to next level: Set new goals and repeat the cycle

Key points for effectively using AI in this cycle include:

• Generate and compare multiple options rather than using answers directly
• Ask for explanations of reasons to deepen your understanding
• Organize your thoughts before seeking AI evaluation/feedback
• Develop habits of critically verifying AI outputs

Example of a Specific Next.js Learning Path

Here's a specific learning path example for web application development.

Iteration 1: Creating a Basic Web App (4 weeks)

• Minimum basics: HTML/CSS fundamentals and JavaScript basic syntax (1 week)
• Small-scale project: Creating a simple web page (2 weeks)
• Reflection: Organizing learned concepts and deepening understanding (1 week)

Iteration 2: Implementing Interactive Features (4 weeks)

• Minimum basics: JavaScript DOM manipulation and event handling (1 week)
• Small-scale project: Web app with interactive elements (2 weeks)
• Reflection: Problem analysis and solution consideration (1 week)

Iteration 3: Utilizing Next.js Framework (6 weeks)

• Minimum basics: React/Next.js basic concepts (1-2 weeks)
• Small-scale project: Simple web app using Next.js (3 weeks)
• Reflection: Understanding framework advantages and constraints (1 week)

Iteration 4: Full-Scale Web Application Development (8 weeks)

• Minimum basics: TypeScript and Tailwind CSS basics (2 weeks)
• Small-scale project: Developing a web app with complex features (4 weeks)
• Reflection: Project evaluation and improvement analysis (2 weeks)

By using AI in each iteration, you can quickly resolve obstacles while developing your own thinking and problem-solving abilities.

Example of Business Analysis Skill Acquisition

A similar approach is effective for acquiring business analysis skills.

Iteration 1: Basic Data Analysis (4 weeks)

• Minimum basics: Basic statistical concepts and Excel operations (1 week)
• Small-scale project: Basic analysis of actual business data (2 weeks)
• Reflection: Evaluation of analysis process and results (1 week)

Iteration 2: Advanced Analysis Methods (6 weeks)

• Minimum basics: Pivot tables and SQL basics (2 weeks)
• Small-scale project: Integrated analysis of multiple data sources (3 weeks)
• Reflection: Evaluation of analysis efficiency and accuracy (1 week)

Iteration 3: Visualization and Reporting (6 weeks)

• Minimum basics: Data visualization principles and tools (1-2 weeks)
• Small-scale project: Creating interactive dashboards (3 weeks)
• Reflection: Evaluation of storytelling and information communication (1 week)

Iteration 4: Business Insight Derivation (8 weeks)

• Minimum basics: Business frameworks and decision theory (2 weeks)
• Small-scale project: Analysis and proposal for actual business challenges (4 weeks)
• Reflection: Analysis of proposal impact and improvement points (2 weeks)

Building Networks for Learning and Growth with Others

Another important element for enhancing learning effectiveness is building networks for learning with others. Especially in the AI era, the value of human collaboration and mutual feedback has increased.

1. Joining learning communities: Sharing information and mutual growth with like-minded peers
2. Utilizing mentorship: Guidance and feedback from experienced practitioners
3. Habitualizing output: Publishing what you've learned on blogs or social media to receive feedback
4. Pair programming/pair learning: Working on challenges with others to gain diverse perspectives

While AI collaboration enhances efficiency, human collaboration expands creativity and perspective. Balancing both is essential for true growth.

The Essence and Future Outlook of Modern Learning

After examining learning approaches from various angles, what's ultimately important is a perspective that transcends the dichotomy between "learning step-by-step from basics" and "learning while building."

Becoming a Metacognitive Learner

What's required in the AI era is becoming a metacognitive learner who can objectively observe and adjust their own learning process. This means not just acquiring knowledge but understanding and optimizing how one learns.

Metacognitive learners have characteristics such as:

1. Clearly conscious of learning purpose: Always aware of why they're learning and what outcomes they aim for
2. Accurately evaluating understanding level: Precisely recognizing what they know and don't know
3. Flexibly adjusting learning strategies according to situations: Changing the balance between fundamental learning and practice based on circumstances
4. Establishing appropriate collaborative relationships with AI: Mastering AI as a tool without becoming dependent

The Role of Educators in the AI Era

I've advocated the mission of "liberating human potential through ideal education." In the AI era, the role of educators is changing significantly—from knowledge transmitters to learning companions.

What's important is discerning which parts AI can substitute and which parts only humans can perform, pursuing educational approaches that maximize human potential. This isn't just about efficiently acquiring knowledge but nurturing the ability to "continue thinking" in the true sense.

Next Steps: Designing Your Learning

As a concrete proposal for you reading this article, I encourage you to practice these steps:

1. Redefine your learning purpose: Clarify the essential purpose of why you want to acquire that technology or skill
2. Reflect on your learning style: Analyze approaches that work for you based on past learning experiences
3. Design small learning cycles: Plan small goals and iterations of 2-4 weeks
4. Build a healthy relationship with AI: Set your own rules for effective utilization without dependence
5. Find learning companions: Create an environment for mutual growth rather than learning alone

Choose the easy path or pursue the essence of learning? That choice will significantly influence your future. What ultimately matters isn't superficial knowledge but deep understanding and application capability. Even in the AI era, this essence remains unchanged.

References

Footnotes

1. Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work: An analysis of the failure of constructivist, discovery, problem-based, experiential, and inquiry-based teaching. Educational Psychologist, 41(2), 75-86. ↩

2. Chen, C. H., & Yang, Y. C. (2019). Revisiting the effects of project-based learning on students' academic achievement: A meta-analysis investigating moderators. Educational Research Review, 26, 71-81. ↩

3. Bjork, R. A., & Bjork, E. L. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. Psychology and the real world: Essays illustrating fundamental contributions to society, 2, 59-68. ↩

4. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2024). Generative AI Can Harm Learning. Wharton School Research Paper. ↩

5. TAG STUDIO (January 2025). Survey Results on Outcomes from Programming School Attendance. PR TIMES. ↩

6. JetBrains (2024). State of Developer Ecosystem Report. Software Developers Statistics 2024. ↩