Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

Prologue — "Will AI Take My Job?"

April 2026, and this is the question developers ask most.

The short answer: somewhere between "it takes it" and "it expands it." When the ATM arrived in 1950, bank tellers did not disappear — they changed. Simple deposit and withdrawal work went away, but the job moved to consulting, sales, and product. The number of tellers per capita actually went up.

But there is an important difference:

ATM: automation of a specific task
AI coding: automation of generally intelligent work

In other words, the future splits sharply depending on which scenario plays out.

This post is the finale of the Season 3 series:

The present — what AI coding tools actually are in 2025
A ten-year forecast — three scenarios
How the value of junior and senior changes
What AI cannot do — the territory that remains
Ethics, safety, and governance
Personal strategy — starting today
The position of Korean developers
A preview of Season 4

Chapter 1: The Present in 2025 — The AI Coding Tool Landscape

1.1 The Tool Spectrum

Generation 1, autocomplete:

IntelliSense, Eclipse
Rule-based

Generation 2, pattern prediction:

GitHub Copilot (2021)
Prediction at the token and line level
Built on GPT-3 → Codex → GPT-4

Generation 3, conversational:

Cursor, Windsurf
Whole-file edits
Codebase understanding

Generation 4, agentic:

Claude Code, Cursor Agent, Devin
Carries out multi-step work
Explore files → edit → test → repeat
"Give it a ticket and it opens a PR"

1.2 Adoption Statistics

GitHub 2024 Survey:

92% of developers are using or have tried AI tools
Copilot paid subscribers: 1.8 million+
Cursor 2025 ARR estimated at $100M+

1.3 The Productivity Effect

GitHub research: Copilot users develop 55% faster
Internal at Microsoft: PR review time down 20%
Caveat: the effect is large on simple work; on complex architecture work it is small or even negative

1.4 Problems

Hallucination: calls to APIs that do not exist
Silent bugs: code that looks right but is wrong
Security: wrong cryptography, injected vulnerabilities
License: GPL code can be suggested
Overconfidence: copy-paste without review

1.5 The New Stage of Agentic Coding

Agents like Claude Code, Cursor Agent, and Devin do:

Multi-step planning
File manipulation
Command execution
Test execution
Error recovery

A new role: the developer moves from "typist" to "orchestrator."

Chapter 2: Three Ten-Year Scenarios

2.1 The Optimistic Scenario (2035)

AI reaches the level of a competent junior:

Every developer gets 10x productivity
The volume of software grows 10x
An explosion of new businesses
Demand for developers actually rises

Why:

Jevons paradox: when efficiency rises, demand explodes
Software still covers only 1% of the world
AI expands the coverable territory to 9%

Implications:

Staff+ engineers wield principal-level influence
Even juniors implement complex features
Developer salaries hold or rise

2.2 The Neutral Scenario (2035)

AI automates up to mid-level:

AI takes over simple CRUD and repetitive work
The entry barrier for juniors rises
Senior+ roles expand
The total number of developers stays at current levels

Implications:

The definition of "developer" changes
A first-year junior becomes an "AI orchestrator"
Senior+ = architecture + team management + AI verification
The middle disappears (U-shaped employment)

2.3 The Pessimistic Scenario (2035)

AI reaches senior level:

Most coding is automated
Demand for developers collapses
Only AI specialists survive
Only a handful of principal engineers remain

Implications:

Career changes at scale
Developer salaries fall
A move into new industries becomes necessary

2.4 The Official Positions of Anthropic and OpenAI

Dario Amodei (Anthropic CEO, 2024):

"By 2026-2027, AI could be better than humans at almost every cognitive task."

Sam Altman (OpenAI, 2024):

"AGI is a gradual evolution; jobs will actually increase."

Demis Hassabis (DeepMind, 2024):

"AGI is 5-10 years out. Jobs will change."

Caution: these leaders have a forecasting bias (fundraising, vision).

2.5 Probabilities

An informal developer survey (2024):

Optimistic: 20-30%
Neutral: 50-60%
Pessimistic: 10-20%

My own view: the neutral scenario is the most likely. A big hit to juniors, stability for seniors, prosperity for staff+.

Chapter 3: How the Junior Role Changes

3.1 Traditional Junior Work

Boilerplate code
Simple bug fixes
CRUD APIs
Writing unit tests
Data transformation scripts

AI is good at all of this.

3.2 The New Junior Role

AI Orchestrator:

Relaying requirements to the AI
Verifying the AI output
Integration testing
Production deployment

Critical review of AI output:

Detecting hallucinations
Security review
Performance evaluation
Code quality

Understanding users and the business:

Requirements analysis
Collaboration with PMs and designers
Business logic modeling

3.3 What Juniors Need to Learn

Debugging: understanding why the AI code is wrong
System design fundamentals: judging AI output in the full context
Business understanding: technology < product
Communication: the language of PMs and designers
Prompt Engineering: collaborating with AI

3.4 Changes in Junior Hiring

Decreasing:

Large-scale junior hiring
Automatic employment out of bootcamps
"Run a new grad for three years and turn them into a senior"

Increasing:

High-caliber juniors (a CS degree + AI fluency)
Direct hiring at mid-level
Rising senior salaries

3.5 Responses

Universities and bootcamps:

CS fundamentals matter more (in order to verify AI output)
Introduce curricula that use AI
Project-centered

Individuals:

OSS and a portfolio over degrees and certificates
Experience solving real problems
Fluency with AI tools

Chapter 4: How the Value of Senior+ Changes

4.1 The Senior's New Territory

The senior in the AI era = "conductor":

Coordinating several AI coding agents the way an orchestra conductor does
Architecture decisions
Guaranteeing code quality
Measuring business impact

4.2 The Amplification of Staff+ Engineers

Before: the "best person" on the team After: "the person with 10x productivity" (amplified by AI)

One staff+ engineer + AI = what used to be a team of 5-10.

4.3 What Becomes More Important

Taste: judgment about what good code is
System design: the whole picture
Security intuition: sensing vulnerabilities
A feel for performance: intuition about bottlenecks
Product thinking: user value
Communication: the language of PMs, designers, and the CEO

4.4 What Becomes Less Important

Syntactic perfection
APIs to memorize
Implementing simple algorithms
Repeating fixed patterns

Just as the compiler checks the grammar, AI fills in the syntax.

4.5 The Principal's New Role

Designing the AI systems themselves
Company-wide AI adoption strategy
Governance, safety, ethics
Industry leadership

The number of principals increases. Juniors can grow fast thanks to AI.

Chapter 5: What AI Cannot Do

5.1 Long-range Reasoning

Architecture decisions for a six-month project
Complex interactions across dozens of files
Setting a long-term vision

AI is strong on short-range work; the long range is still human.

5.2 Novel Problem

A business problem no one has seen before
A new domain
Creative resolution

AI works within the patterns in its training data. New patterns are human.

5.3 Aligning Stakeholders

Trade-offs with PMs and designers
Briefing executives
Customer interviews

Requires social context.

5.4 Responsibility and Judgment

A bug causes a $1M revenue loss → who is responsible?
The deployment decision
Security and privacy judgment

Legal and ethical judgment is human.

5.5 Product Sense

What users actually want
The Steve Jobs-style "no, this button is gone" decision
Aesthetics, taste

Requires an understanding of culture and context.

5.6 On-call and Incident Response

The 3AM alarm
Fast judgment
Coordinating the team
Communication (customers, press)

Requires real-time social coordination.

5.7 A Summary by Area

What AI is strong at:

Boilerplate, CRUD
Existing patterns
Writing documentation
Generating tests

What humans are strong at:

Long-range design
New problems
Social coordination
Responsibility and ethics

Chapter 6: Ethics, Safety, and Governance

6.1 Concerns About Code Quality

The 2024 GitClear study:

Code churn rose after AI adoption
A suspected increase in defects
Worries about "copy-paste engineering"

6.2 Licensing and Copyright

Cases of Copilot suggesting GPL code
Litigation: GitHub Copilot vs developers (2022)
Conclusion: undecided

Safety strategy:

License scanning tools
A license policy
Documented code origin

6.3 Security

Code suggested by AI:

SQL injection risk
Wrong crypto
Hardcoded secrets
Vulnerable dependencies

→ A security review is mandatory.

6.4 Bias

Training data bias
Weak on minority languages and frameworks
An overall US and English bias

6.5 AGI Safety

The alignment problem: does the AI behave the way humans intend
Corporate responsibility: OpenAI, Anthropic, Google
Government regulation: the EU AI Act, Korea's AI Framework Act

6.6 The Developer's Role

Refusing blind adoption
Holding the quality bar
Building verification processes
Spreading ethical use

Chapter 7: Personal Strategy — Starting Today

7.1 Fluency With AI Tools

Mandatory:

Daily use of Copilot or a similar tool
Mastery of one of Cursor, Claude Code, or Windsurf
Experiments with agent mode
Prompt engineering

Thirty minutes a day of experimentation.

7.2 From T-Shaped to T+ Shaped

One depth (as before)
One breadth (AI)
The result: T+ (expertise × AI)

7.3 Leading AI Adoption at Your Company

Evaluating AI tools
Writing team guidelines
Sharing success stories
Security and quality standards

A promotion opportunity.

7.4 Entering a New Domain

Accelerating learning with AI:

Pairing with Claude or ChatGPT
A new language in 30 days
A new domain in 6 months

A pace that was hard until 2025, now possible.

7.5 Side Projects

A fast MVP with AI tools
Solving a real problem
Income or a portfolio

7.6 Writing and Speaking

In the AI era the value of clear thinking rises
Blogs, talks
Presenting the "why" that AI cannot

7.7 Strengthening the Fundamentals

CS fundamentals (algorithms, OS, networking)
System design
Debugging

In the AI era the fundamentals matter more.

Chapter 8: The Position of Korean Developers

8.1 Korea's Strengths

Fast adaptation (quick to adopt new technology)
A high level of education
English ability on the rise
Active government investment in AI

8.2 Korea's Weaknesses

A small share of global AI research (compared with the US and China)
Korean-language LLMs lag behind GPT and Claude
A risk-averse culture

8.3 Opportunities

A power in AI applications: inference, fine-tuning, deployment
Application in fintech and commerce
AI for K-culture content: generating drama and K-pop
Combining with hardware: Samsung and SK Hynix + AI chips

8.4 Threats

Dependence on the global AI giants (OpenAI, Anthropic, Google)
Data sovereignty
An outflow of AI talent

8.5 Personal Strategy

Developers living in Korea:

Contribute to global open source
Write in English
Look for remote positions
Build AI expertise

Considering a move abroad:

US: OpenAI, Anthropic, Google, Meta
Europe: DeepMind, Mistral
Southeast Asia: the Singapore AI hub

Chapter 9: A Season 3 Retrospective

9.1 What Season 3 Covered

Ep 1: Architecture case studies (Netflix, Stripe, and others)
Ep 2: Anatomy of famous postmortems
Ep 3: Scaling inflection points
Ep 4: Big tech culture
Ep 5: The staff/principal path
Ep 6: OSS maintainers
Ep 7: Writing for developers
Ep 8: Finances
Ep 9: Time management
Ep 10: Mental health
Ep 11: Learning
Ep 12: Networking and branding
Ep 13 (this post): AI and the next ten years

9.2 Common Themes

Sustainability: long-term, not short-term
Authenticity: what is real, not what is inflated
Human: technology + people
Compounding: the accumulation of small things
Community: together, not alone

9.3 The Message of Season 3

Development is not only technology. It is a comprehensive profession that handles people, culture, career, mind, money, time, learning, networks, and the future all at once.

In the AI era, what is human becomes more important.

Chapter 10: A Preview of Season 4

10.1 Candidate Topics

AI applications in practice: RAG, fine-tuning, agent systems
The complete security guide: red team, blue team, zero trust
Data Engineering: ETL, data lake, warehouse 2.0
Observability in practice: OpenTelemetry, traces, cost
The latest in web standards: WebGPU, WebAssembly 2.0, Privacy API
Cross-platform: Tauri, the evolution of Electron, Flutter
Developer Tools: language servers, LSP, Tree-sitter
Compiler stories: LLVM, the internals of the Rust compiler

10.2 Format Experiments

Shorter, more focused posts
Experiments with an interview format
In-depth visits to major Korean companies

Chapter 11: A 12-Item Checklist for the AI Era

Chapter 12: Ten Anti-Patterns of the AI Era

1) Avoiding AI Entirely

"I do not use AI, only pure skill." → Your competitors have 10x productivity. Obsolete within a year.

2) Depending on AI Entirely

All the code is handed to the AI. No understanding of your own. → When a bug appears, you cannot debug it.

3) Adopting Without Verification

AI output goes straight to production. → Serious bugs plus security problems.

4) Abandoning the Fundamentals

"The AI does the algorithms, so I do not need to know." → You cannot evaluate the AI output.

5) Abandoning Writing and Thinking

Nothing but "AI, write it for me," over and over. → Your own capacity to think disappears.

6) No Long-Term Plan

"In a year the AI will do it, so I am not interested." → You meet the change unprepared.

7) Chasing FOMO

Trying every new AI tool. → You master none of them.

8) Doomerism

"AI is going to take all the jobs anyway." → A self-fulfilling prophecy. Paralysis of action.

9) Believing the Hype

"AGI will solve everything next year." → Inflated expectations that ignore reality.

10) Isolation From the Community

Depending only on AI, cut off from people. → Loss of your network and support system.

Closing — The Developer Ten Years From Now

Principle 1: AI Is a Tool

Just as the hammer did not replace the carpenter. How you use it decides your fate.

Principle 2: What Is Human Rises in Value

Creativity, judgment, communication, ethics. They become more precious in the AI era.

Principle 3: The Fundamentals Are the Leverage

To evaluate AI output you have to know the basics. The fundamentals matter more.

Principle 4: Adaptability Is Survival

Do not cling to the old languages and frameworks. Relearn every year.

Principle 5: Community

Absorbing the change alone is hard. Learn and adapt together.

Principle 6: Optimism + Preparation

Pessimism is paralysis. Optimism alone is dangerous. Optimistic realism.

Principle 7: Read the Originals

Machines of Loving Grace - Dario Amodei (Anthropic, 2024)
Situational Awareness - Leopold Aschenbrenner
The Intelligence Age - Sam Altman
AI Snake Oil - Arvind Narayanan
The Coming Wave - Mustafa Suleyman
Scaling Hypothesis - Gwern Branwen
[Korean-language sources: AI Times, ETRI reports, the Ministry of Science and ICT AI strategy]

Wrapping Up Season 3

Across these 13 posts I covered the whole of a developer's life:

Technology: architecture, postmortems, scaling
Organization: big tech culture, the staff+ path
Outside: OSS, writing, networking
Self-management: finances, time, mental health, learning
The future: AI and the next ten years

I hoped it would become a comprehensive guide to what kind of developer to be.

Thanks to everyone who read it, and I will be back with Season 4. Until then, may you build both your technology and your life well.

— April 2026, Youngju Kim

Next Post Preview — Season 4 Ep 1: "The Complete Practical RAG Guide: Retrieval, Embeddings, Vector DBs, and the Boundary of Fine-tuning"

Season 4 Ep 1 will cover:

The basic principles of RAG and how it evolved
Choosing an embedding model (OpenAI, Cohere, Korean models)
Comparing vector DBs (pgvector, Pinecone, Weaviate, Qdrant, ChromaDB)
Chunking strategies
Rerank, Hybrid Search
Fine-tuning vs RAG vs Prompt Engineering
The impact of long context (100K+ tokens)
Real-world RAG architectures (Notion AI, Claude)
The particulars of Korean-language RAG
Cost and latency optimization

See you in the next post.