나도 처음엔 애들이 transformer. js로 짜준 걸 무조건 믿고 푸시했는데, production에서 터지는데다가 rollback까지 했다. 요즘 frontier 모델들이 내가 고민하는 것보다 빠르게 답은 주는데, 오프바이 원 패턴이나 메모리 누수 체크는 여전히 허술하더라.
사실 유닛 테스트 몇 개 추가하면 잡히는건데, LLM이 '이거 문제 없을거' 라고 딱 잘라서 말하면 무시하게 되는게 함정이야. 그래서 나는 지금 로컬에 올린 작은 모델로 우선 리트라이하게 해두고, 스테이트 오브 아트 모델은 설계 단계에서만 쓰는 중이다.
출처
- Ask HN: Solo Senior Developers, Where do we find you? (community_intel)
- Ask HN: Are we ready for vulnerabilities to be words instead of code? (community_intel)
- Ask HN: Are the newest LLMs better than you at programming? (community_intel)