Yet Another LLM Rant
15 days ago
- #AI-limitations
- #LLM
- #ChatGPT
- ChatGPT 5 provided incorrect information about ZSTD compression support in Swift on iPhone, claiming it works on iOS 16+ when it has never been supported by Apple.
- LLMs like ChatGPT generate responses based on statistical likelihood, not factual accuracy or logical reasoning, leading to confident but incorrect answers.
- The author emphasizes that LLMs cannot think, learn, or reason; they only produce text that seems plausible based on their training data.
- A human analogy illustrates how humans can question, test, and update their beliefs based on new evidence, unlike LLMs which stick to statistically likely responses.
- The post advises against relying on LLMs for critical tasks and encourages using human judgment, creativity, and expertise instead.