If you frequent Hacker News regurlarly, you have likely noticed the buzz around engineers using AI (specifically Large Language Models, or LLMs) to tackle Computer Science problems. I want to be clear: I’m not against LLMs. LLMs are incredibly powerful tools, and can be a huge boon to engineers. They can automate repetitive tasks, generate code snippets, help with brainstorming, assist in debugging, … and this can frees up engineers’ time and mental energy, which could be channeled into more complex, creative problem-solving. But, like any tool, LLMs should be used wisely. LLMs can hallucinate, exhibit inconsistencies (especially with self-reflection models), and harbor biases. These limitations mean that LLM outputs require careful review before they can be trusted.
Micro-Macro Retrieval: Reducing Long-Form Hallucination in Large Language Models