
2023λ
3μμ μμ±λ The End of Front-End Developmentλ AIλ‘ μΈν΄ FEκ° μνλ°λλ€λ 견ν΄μ λ°λνλ€λ λ΄μ©μ΄λ€.
μ μμ μκ°μ΄ λΉμ μ ν¨κ» μ μ€λͺ
λμ΄ μμ΄μ μΈμκΉμ λΆλΆλ§ μμν΄μ μ 리νλ€.
νμνΈμΈ The Post-Developer Eraλ κ³§ μμ±ν μμ μ΄λ€.
π‘TL;DR
- AIκ° λ°μ νκ³ μμ§λ§, νκ°μ΄ μμ μ μκ³ λκ·λͺ¨ μλΉμ€ κ°λ°μ λ―Έν‘νκΈ° λλ¬Έμ κ°λ°μκ° μ¬λΌμ§μ§ μμ κ²μ΄λ€.
- AIμ λ΅λ³μ λ§Ήμ νμ§ μκ³ λΉνμ μΌλ‘ μκ°νλ©° νμ λκ΅¬λ‘ νμ©νλ€λ©΄ κ°λ° μμ°μ±μ΄ ν₯μλ κ²μ΄λ€.
- κ°λ° μμ°μ±μ΄ ν₯μλλ©΄ λ€μν κΈ°λ₯μ κ·Έλ§νΌ λΉ λ₯΄κ² λ§λ€μ΄λΌ μ μκ³ λ€μν λ²κ·Έλ€μ λ§λ μ μκΈ° λλ¬Έμ κ°λ°μμ μμλ κ³μ μμ κ²μ΄λ€.
This sentiment is all over social media right now:
μ΄λ¬ν μ μ(AIλ FEκ°λ°μ λ₯μν΄μ§κ±°λ€)λ νμ¬ μμ λ―Έλμ΄ μ λ°μ κ±Έμ³ μμ΅λλ€:

I couldn't disagree more. I don't think front-end developer jobs are going anywhere. And I'm getting pretty sick of the FUD(Fear, Uncertainty, and Doubt) being spread online.
So, in this blog post, I'm going to share my hypothesis for what will happen. Things are going to change, but not in the scary way people are saying.
μ λμν μ μμ΅λλ€. νλ‘ νΈμλ κ°λ°μμ μΌμ리λ μ¬λΌμ§ κ² κ°μ§ μμμ. κ·Έλ¦¬κ³ μ¨λΌμΈμμ νΌμ§κ³ μλ 곡ν¬, λΆνμ€μ±, μμ¬μλ κ½€ μ§μ³κ°κ³ μμ΅λλ€.
κ·Έλμ μ΄ λΈλ‘κ·Έ κ²μλ¬Όμμλ μμΌλ‘ μΌμ΄λ μΌμ λν μ κ°μ€μ 곡μ νκ³ μ ν©λλ€. μν©μ΄ λ°λ κ²μ΄μ§λ§ μ¬λλ€μ΄ λ§νλ 무μμ΄ λ°©μμ μλ κ²μ λλ€.
λλ μλ
μ "FEκ°λ°λ§ νλ©΄ AIλ‘ λ체λ κ±°λ€"λΌλ λ§μ μ¬λλ€ μ¬μ΄μμ, SNSμμ κ½€ λ΄€μλ€.
κ·Έλ° κΈ νΉμ λ§μ λ³Ό λλ§λ€ μκ·Ό μ€νΈλ μ€ λ°μμλλ°, μ΄ λΆμ κΈμ μ‘°κΈ λ μΌμ° λ΄€μΌλ©΄ μ’μμ κ±°λΌλ μκ°μ΄ λ λ€.
Pretty much since the beginning, there's been a concern that web developers would be made redundant by some new technology. In the 2000s, it was WordPress. In the 2010s, it was Webflow. In the early 2020s, it was "no code" tools.
(μ€λ΅)
And yet, web developers continue to exist.
μ²μλΆν° μΉ κ°λ°μλ€μ΄ μλ‘μ΄ κΈ°μ λ‘ μΈν΄ μΌμ리λ₯Ό μμ κ²μ΄λΌλ μ°λ €κ° μμμ΅λλ€. 2000λ λμλ μλνλ μ€κ° κ·Έλ¬κ³ μ. 2010λ λμλ Webflowμμ΅λλ€. 2020λ λ μ΄λ°μλ 'λ Έμ½λ' λꡬμμ΅λλ€.
(μ€λ΅)
κ·ΈλΌμλ λΆκ΅¬νκ³ μΉ κ°λ°μλ κ³μ μ‘΄μ¬ν©λλ€.
There is an enormous difference between this HTML document and the sorts of code front-end developers write today.
μ΄ HTML λ¬Έμμ μ€λλ νλ‘ νΈμλ κ°λ°μκ° μμ±νλ μ½λ μ¬μ΄μλ μμ²λ μ°¨μ΄κ° μμ΅λλ€.
μ¦, κ°λ¨ν μΉμ¬μ΄νΈλ₯Ό λ§λλ κ²κ³Ό μ λ¬Έκ°(μ¬κΈ°μμλ νλ‘ νΈμλ κ°λ°μ)κ° λ§λλ κ²μ λΆλͺ ν μ°¨μ΄κ° μκΈ° λλ¬Έμ κ°λ°μκ° AIλ‘ λ체λκ±°λ μ¬λΌμ§μ§ μμ κ±°λΌκ³ μκΈ°νλ€.
LLMs aren't able to validate their assumptions, or test their hypotheses. They can't confirm whether what they're saying is true or not. (μ€λ΅) Sometimes, parts of that response are nonsensical. The OpenAI team refers to these as βhallucinationsβ.
LLMμ μμ μ κ°μ μ κ²μ¦νκ±°λ κ°μ€μ ν μ€νΈν μ μμ΅λλ€. κ·Έλ€μ μμ μ΄ λ§νλ κ²μ΄ μ¬μ€μΈμ§ μλμ§ νμΈν μ μμ΅λλ€. (μ€λ΅) OpenAI νμ μ΄λ₯Ό "νκ°"μ΄λΌκ³ λΆλ¦ λλ€.
And so, the accuracy will improve, but it'll never be perfect. (μ€λ΅) If you're not a programmer, you won't be able to tell which parts are accurate or not. You won't be able to spot the hallucinations.
λ°λΌμ μ νλλ ν₯μλκ² μ§λ§ μλ²½νμ§λ μμ κ²μ λλ€. (μ€λ΅) νλ‘κ·Έλλ¨Έκ° μλλΌλ©΄ μ΄λ€ λΆλΆμ΄ μ ννμ§ μλμ§ κ΅¬λΆν μ μμ΅λλ€. νκ°μ λ°κ²¬ν μ μμ κ²μ λλ€.
LLMμ νκ° νμμ μλ²½νκ² ν΄κ²°ν μ μμ κ²μ΄λ―λ‘ μ λ¬Έκ°μΈ κ°λ°μκ° μ§μ μ¬μ€μ¬λΆλ₯Ό νμΈν΄μΌ νλ€.
AIκ° λ§λ€μ΄λΈ μλͺ»λ μ 보(νκ°)λ₯Ό κ²μ¦νμ§ μκ³ κ·Έλλ‘ μΌμ λ λκ° μ±
μμ§ κ²μΈκ°?
AIκ° μ±
μμ§μ§ μμ κ²μ΄λ―λ‘ κ°λ°μκ° νμ νμν κ±°κ³ μ¬λΌμ§μ§ μμ κ²μ΄λ€.
AI isn't magic. It's only as good as its training data. Code snippets are all over the internet, and are often generic. By contrast, every codebase is unique. There are very few large open-source codebases. How's the AI supposed to learn how to build big real-world projects?
AIλ λ§λ²μ΄ μλλλ€. AIλ νμ΅ λ°μ΄ν°λ§νΌλ§ μ°μν λΏμ λλ€. μ½λ μ€λν«μ μΈν°λ·μ λλ € μμΌλ©° μΌλ°μ μΈ κ²½μ°κ° λ§μ΅λλ€. λ°λ©΄μ λͺ¨λ μ½λλ² μ΄μ€λ κ³ μ ν©λλ€. λκ·λͺ¨ μ€ν μμ€ μ½λλ² μ΄μ€λ κ±°μ μμ΅λλ€. AIλ μ΄λ»κ² μ€μ λκ·λͺ¨ νλ‘μ νΈλ₯Ό ꡬμΆνλ λ°©λ²μ νμ΅ν μ μμκΉμ?
But I think we're still a very long way from major tech companies letting go of their developer staff and replacing them with prompt engineers. It seems to me like there are several potentially-unsolveable problems that stand in the way of this becoming a reality.
νμ§λ§ μ£Όμ κΈ°μ κΈ°μ λ€μ΄ κ°λ°μλ₯Ό ν΄κ³ νκ³ μ΄λ€μ μ¦κ°μ μΈ μμ§λμ΄λ‘ λ체νκΈ°μλ μμ§ κ° κΈΈμ΄ λ©λ€κ³ μκ°ν©λλ€. μ΄λ₯Ό μ€ννλ λ°λ λͺ κ°μ§ ν΄κ²°νμ§ λͺ»ν μ μ¬μ λ¬Έμ κ° μλ κ² κ°μ΅λλ€.
μ½λ μ€λν«μ μ μ μνλ κ²λ§ λ³΄κ³ κ°λ°μλ AIλ‘ λ체νλ κ±°λ 2023λ μλ, μ§κΈλ μ£λΆλ₯Έ νλ¨μΈ κ±° κ°λ€.
You might not know it from this article, but I'm actually pretty optimistic about AI. π
I think the most likely scenario is that tools like GPT-4 become integrated into developer tooling, and used to amplify what a skilled developer can do.
μ΄ κΈμμλ μ λͺ¨λ₯΄μκ² μ§λ§ μ λ μ¬μ€ AIμ λν΄ κ½€ λκ΄μ μΌλ‘ μκ°ν©λλ€. κ°μ₯ κ°λ₯μ±μ΄ λμ μλ리μ€λ GPT-4μ κ°μ λκ΅¬κ° κ°λ°μ λꡬμ ν΅ν©λμ΄ μλ ¨λ κ°λ°μκ° ν μ μλ μΌμ μ¦νμν€λ λ° μ¬μ©λλ κ²μ΄λΌκ³ μκ°ν©λλ€.
Carpenters weren't replaced by power tools, accountants weren't replaced by spreadsheets, photographers weren't replaced by digital cameras / smartphones, and I don't think that developers will be replaced by LLMs.
λͺ©μλ μ λ 곡ꡬλ‘, νκ³μ¬λ μ€νλ λμνΈλ‘, μ¬μ§μκ°λ λμ§νΈ μΉ΄λ©λΌ/μ€λ§νΈν°μΌλ‘ λ체λμ§ μμλ―μ΄, κ°λ°μκ° LLMμΌλ‘ λ체λμ§λ μμ κ²μ΄λΌκ³ μκ°ν©λλ€.
κ°λ°μ λν μμλ μ¦κ°νκ³ μλ μνμ΄λ©°, κ°λ°μ μμ°μ±μ΄ LLM λμ ν₯μλλ€λ©΄ κ·Έλ§νΌ λ§μ κΈ°λ₯μ λ§λ€κ³ λ§μ λ²κ·Έλ€μ΄ λ°μν κ²μ΄κΈ° λλ¬Έμ κ°λ°μκ° μ¬λΌμ§μ§ μμ κ²μ΄λ€.
λ² ν λ μ λλ©μ΄ν° κ²Έ μΌλ¬μ€νΈλ μ΄ν°μΈ Aaron Blaiseμ μ¬λ‘λ€.
he doesn't see these tools as a threat, he sees them as something that will increase animator productivity, and lead to more animator jobs.
Artists and knowledge workers across dozens of industries are having the same conversation right now. People are worried that their job is about to be gobbled up by AIs like GPT-4, DALL-E 2, and Midjourney.
κ·Έλ μ΄λ¬ν ν΄μ μνμΌλ‘ λ³΄μ§ μκ³ μ λλ©μ΄ν°μ μμ°μ±μ λμ΄κ³ λ λ§μ μ λλ©μ΄ν° μΌμ리λ₯Ό μ°½μΆν μ μλ κΈ°νλ‘ λ³΄κ³ μμ΅λλ€.
μμ κ°μ μ°μ λΆμΌμ μν°μ€νΈμ μ§μ κ·Όλ‘μλ€μ΄ μ§κΈ κ°μ λνλ₯Ό λλκ³ μμ΅λλ€. μ¬λλ€μ GPT-4, DALL-E 2, Midjourneyμ κ°μ μΈκ³΅μ§λ₯μ΄ μμ λ€μ μΌμ리λ₯Ό μ μν κΉ λ΄ κ±±μ νκ³ μμ΅λλ€.
κ°λ° λΏλ§ μλλΌ μ¬λ¬ μ°μ
λΆμΌμμλ AI λꡬμ μνμ λλΌκ³ μλ€.
νμ§λ§ AI λꡬλ₯Ό μΌμ리λ₯Ό μννλ κ΄μ λμ νμ
μ ν΅ν μΌμ리 μ°½μΆμ κ΄μ μΌλ‘ λ°λΌλ³΄λ κ²λ νμν κ±° κ°λ€.
The wrong way would be to treat it like GPS navigation. When I have to drive somewhere, I pop the address into my GPS, and follow its instructions indiscriminately. I do generally wind up where I need to go, but it requires 0 mental effort on my part. As a result, my sense of direction has totally atrophied. I can't go anywhere now without a synthesized voice telling me where to turn. π¬
μλͺ»λ λ°©λ²μ GPS λ΄λΉκ²μ΄μ μ²λΌ μ·¨κΈνλ κ²μ λλ€. μ λ μ΄λκ°λ‘ μ΄μ ν΄μΌ ν λ GPSμ μ£Όμλ₯Ό μ λ ₯νκ³ λ¬΄μ°¨λ³μ μΌλ‘ μ§μλ₯Ό λ°λ¦ λλ€. 보ν΅μ μ κ° κ°μΌ ν κ³³μ λμ°©νκΈ΄ νμ§λ§ μ μ μ μΈ λ Έλ ₯μ΄ μ ν νμνμ§ μμ΅λλ€. κ·Έ κ²°κ³Ό λ°©ν₯ κ°κ°μ΄ μμ ν μμΆλμμ΅λλ€. μ΄μ λ μ΄λλ‘ κ°μΌ νλμ§ μλ €μ£Όλ ν©μ±λ μμ± μμ΄λ μ무 λ°λ κ° μ μμ΅λλ€.
AIμ λ΅λ³μ μ λμ μΌλ‘ λ―Ώλ μ¬λμ΄ μ’ μ’ μλ κ² κ°λλ°, μ μ ν λΉμ μΈ κ±° κ°λ€.
Instead of treating it like a GPS, I'd suggest treating it like you're a member of a jury, and the LLM is the defendant, taking the stand.
You'll listen to what they have to say, but you won't accept it as fact. You'll be skeptical, and think critically about every word.
Instead of blindly copy/pasting the code that ChatGPT generates, go through it line-by-line, and make sure you understand. Ask it for clarification. And double-check things that seem suspicious with an authoritative source (eg. the official documentation). Keep in mind that LLMs are 100% confident, but not 100% accurate.
GPSμ²λΌ μ·¨κΈνλ λμ λ°°μ¬μλ¨μ μΌμμΌλ‘, LLMμ νΌκ³ μΈμΌλ‘μ μ¦μΈμμ μ μλ€κ³ μκ°νλ©΄ μ’μ κ² κ°μ΅λλ€.
κ·Έλ€μ΄ νλ λ§μ λ£κΈ°λ νκ² μ§λ§ κ·Έκ²μ μ¬μ€λ‘ λ°μλ€μ΄μ§λ μμ κ²μ λλ€. νμμ μΈ νλλ₯Ό μ·¨νκ³ λͺ¨λ λ§μ λν΄ λΉνμ μΌλ‘ μκ°νκ² λ κ²μ λλ€.
ChatGPTκ° μμ±νλ μ½λλ₯Ό 무ν±λκ³ λ³΅μ¬/λΆμ¬λ£κΈ°νμ§ λ§κ³ ν μ€ ν μ€ μ΄ν΄λ³΄κ³ μ΄ν΄νλμ§ νμΈνμΈμ. μ€λͺ μ μμ²νμΈμ. κ·Έλ¦¬κ³ μμ¬μ€λ¬μ΄ λΆλΆμ 곡μ λ ₯ μλ μΆμ²(μ: 곡μ λ¬Έμ)λ₯Ό ν΅ν΄ λ€μ νμΈνμΈμ. LLMμ 100% νμ ν μλ μμ§λ§ 100% μ ννμ§λ μλ€λ μ μ λͺ μ¬νμΈμ.
If you follow this strategy, I think LLMs can offer a lot of value.
μ΄ μ λ΅μ λ°λ₯Έλ€λ©΄ LLMμ λ§μ κ°μΉλ₯Ό μ 곡ν μ μλ€κ³ μκ°ν©λλ€.