Practical example of LLM’s hallucinating !

When I started working with large language models (LLMs), the concept of “hallucination” wasn’t entirely clear. I would explicitly mention in the prompt something like, “Please don’t hallucinate; if you don’t know, say you cannot answer.” Despite this, LLM’s can give answers without verifying if the information is correct. Understanding why this happens and the internals behind it might require another post, as I still need to dig deeper.

Recently, I was exploring the internals of Log-Structured Merge (LSM) Trees. Typically, each write is first stored in an in-memory data structure and at the same time written to a Write-Ahead Log (WAL) file.

I asked ChatGPT to dwell upon the semantics of writes in LevelDB/Cassandra.

question to chatgpt

Here’s the response I received:

“But from LevelDB docs here, LevelDB by default is async when it comes to writes.”

This response contradicted the docs. So, I asked ChatGPT to recheck the information.

Upon rechecking, ChatGPT admitted it had made a mistake. I then asked how it missed this detail initially.

This experience highlighted a crucial point: you must be careful and provide better prompts when drawing conclusions from LLMs. Be as specific as you can.