출처가 달린 자료 위에서만 답하면 환각이 사라집니다 —
그럴듯한 거짓말 대신, 검증된 소스를 인용한 답이 나옵니다.
Answer only on top of cited sources and hallucination disappears — instead of a plausible lie, you get an answer that quotes verified material.
ai-toolingmemory-systemscontext-beats-prompt
출처가 달린 자료 위에서만 답하면 환각이 사라집니다 —
그럴듯한 거짓말 대신, 검증된 소스를 인용한 답이 나옵니다.
LLM의 큰 약점은 환각입니다 — 정보가 확실하지 않을 때 그럴듯한 거짓말을 합니다. 해법은 모델에게 자유 작문을 시키지 않는 것입니다. 내가 올린 소스 안에서만 답하게 하고, 문장마다 출처를 인용하게 하면 환각이 거의 사라집니다. Perplexity가 뜬 이유도, NotebookLM이 리서치 도구로 자리 잡은 이유도 같습니다 — 둘 다 근거 자료를 인용해 답을 만듭니다. 정보를 잘 찾고 검증하는 일이, 이제 모델 자체를 쓰는 일보다 중요해졌습니다.Also in EnglishAnswer only on top of cited sources and hallucination disappears — instead of a plausible lie, you get an answer that quotes verified material.
The big weakness of an LLM is hallucination — when it isn't sure, it tells a plausible lie. The fix is to stop letting the model free-write: make it answer only from sources you've supplied, and cite a source on every line, and the hallucination nearly vanishes. It's why Perplexity took off and why NotebookLM became a research tool — both build answers by quoting grounded material. Finding and validating the right information now matters more than working the model itself.