The Mystery Behind why LLMs Use the Word "Delve" So Often

section-content

The Mystery Behind why LLMs Use the Word "Delve" So Often

💡 Remember this graph of the surge in medical articles that use the word "delve", a favorite word of ChatGPT?

Well, that word has long been accepted as a sign that the text is generated. But people are still perplexed as to why "delve"? What is the mystery surrounding this word?

I think I've figured out what it's all about.

Yes, in the US the word delve is rarely used, but in Nigeria it is very common: for example, it is often pronounced by teachers in elementary school.

And GPT has RLHF, where the quality of the model's answers is evaluated by people. The RLHF is often worked on by people whose labor is cheaper than that of the average US resident.

See where I'm going with this? For RLHF, they hire people from Nigeria for whom the word "delve" is a classic, they describe the model's responses using that word ("delve deeper"), and then their feedback is fed back to the model for pre-training. There llm memorizes that word as "good".