AI writes the code, so why study CS? SNU's Prof. Jaewook Lee: "You learn to define the problem"
Nov 5, 2026Original 서울대학교 · Anthropic · OpenAI · 스탠퍼드 디지털경제연구소 외 ↗
In 3 lines
- Even if code writing is nearly automated, deciding what to build stays human, says the SNU professor
- A Stanford analysis finds employment for ages 22–25 in AI-exposed jobs 19% lower, with no gap for experienced workers
- What matters: building experience, problem definition and taste; start with a small project solved with AI
"Writing program code could be automated to close to 100%." That was Prof. Jaewook Lee of Seoul National University's Department of Computer Science and Engineering, who heads its AI institute, on the university's YouTube series Sharojapda Season 2. In an 18-minute talk he covered why CS is still worth studying, why entry has gotten harder for junior developers, and which skills will matter. Each claim is paired with context from Anthropic, OpenAI, Stanford and Korean hiring trends.
Code may be nearly 100% automated, but deciding what to build stays human
Prof. Jaewook Lee of Seoul National University's Department of Computer Science and Engineering, who also heads the university's AI institute, said writing code could be automated to "close to 100%." Defining what to build, he said, remains with people. Working with AI is like managing agents, and someone without engineering experience will struggle to be a good engineering manager. In Anthropic's April 2025 analysis, 79% of Claude Code conversations were the 'automation' type, where the human directs and the AI executes.
Less time writing, more time reading and critiquing
Lee described the shift as one from writing to reading. He used to spend most of his energy writing papers, code and slides; now that producing a draft from an idea is easy, most of his time goes to reading and critiquing, so critical reading matters more. He also acknowledged that vibe coding can reach real work, citing a district-office civil servant in Seoul who built a chatbot that compares complex regulations to answer questions.
"99% sounds high" — but at a billion users, that's 10 million people
Lee said vibe coding is good for prototypes, but in production software that works 95% of the time is effectively broken. With a billion users, a 1% error rate means 10 million people get wrong results, and in finance or security even 99% can be a weak bar. In an Anthropic experiment with 52 developers learning a new library, the AI-assisted group scored about 50% on a comprehension quiz versus about 67% for those who coded by hand, with the widest gap on debugging questions.
Anthropic · Anthropic ↗When students become 'delivery couriers,' learning doesn't happen
Lee said textbook problems no longer work as learning tools because AI solves them with 100% accuracy. His main worry is students becoming a 'middleman' who pastes a problem into AI and carries the answer back. His answer is twofold: keep exams and quizzes for the basics, the way we still teach arithmetic despite calculators, and give open-ended assignments with minimal constraints so students define and design the problem themselves. Anthropic's experiment found that participants who handed code generation entirely to AI scored lower, while those who asked questions and probed concepts kept their learning gains.
Anthropic · Anthropic ↗"CS doesn't teach coding; it teaches how to solve problems"
Lee summed up the core of computer science as algorithmic thinking, which breaks complex tasks into simple steps, and data thinking, which draws insight from data. On jobs, he expects reality to land between the Jevons-paradox view, where cheaper coding grows total demand, and the view that demand disappears. When ATMs spread, many predicted tellers would vanish; in economist James Bessen's analysis, teller numbers held steady or even rose, because fewer staff per branch meant banks opened more branches, and the work shifted from processing to advising customers.
Conversable Economist · Conversable Economist ↗The entry barrier really has risen — yet "everyone is at the starting line"
Lee conceded that it has become harder for junior developers to get in, calling it "the reality we are actually living." A Stanford Digital Economy Lab analysis of payroll data (revised August 2026) found employment for ages 22–25 in AI-exposed occupations 19% below less-exposed peers, with no comparable gap for experienced workers, driven by reduced hiring rather than layoffs. IBM, meanwhile, said it will triple US entry-level hiring this year. Lee argued that because the technology is new to everyone, this is an unusually level moment, and said he would choose computer science again at 20.
Stanford Digital Economy Lab · Stanford ↗What matters now: building experience, problem definition and taste
Lee named three capabilities: hands-on experience of building and discarding things, the judgment to define problems and tell good solutions from bad, and practice critiquing code and designs. He noted that the best engineering managers he knows are mostly people with deep engineering experience, and mentioned a coined word, 'gibalja,' for planner plus developer. Hiring is shifting too: Nexon dropped its coding test for its game-programmer internship and introduced an AI-usage assessment.
트위그24 · 트위그24 ↗For non-majors: "start with a small project and solve it with AI"
To people outside the field, Lee advised assuming most problems can be solved with AI and trying even a small project. After a few rounds you develop a feel for what AI does well and badly. He noted that model capability grows faster than users' ability to use it, widening the gap between power users and average users. OpenAI's January report likewise found that, on the same Plus plan, the 95th-percentile user draws about 7x more 'thinking capability' than the median user. Borrowing a phrase from visiting director Madison Hwang, Lee called AI "the great equalizer."
OpenAI · OpenAI ↗
