7. Dissonance causes LLM confusion
Good prompt writers know to avoid certain combinations of words to avoid unintended connotations. For example, if you write that a ball flies through the air, it is structurally no different than saying that a fruit flies through the air. However, the compound noun “fruit fly” then causes AI confusion: is it an insect or fruit? Such linguistic dissonances can be particularly dangerous for prompt engineers who do not feed the AI with instructions in their native language.
8. Typography is a technique
A prompt engineer from a major AI player once explained to me why it makes a difference to his employer’s model whether there is a space after a period or two. This was because the developers had not normalized the training data corpus, which is why some sentences had two spaces and others had one space after the period at the end of the sentence. In general, texts written by older people often had a double space – as was common with typewriters in the past. As a result, the Large Language Model increasingly spat out results with double spaces that were based on older training materials. A subtle difference with a big impact.
9. Machines just regurgitate
The poet Ezra Pund once described the essential task of poets with the words “make it new”. Unfortunately, something new is one of the few things that large language models cannot deliver. They may be able to surprise us with obscure fun facts that they scrape from the back cracks of their training data sets. But essentially, LLMs using neural networks do nothing more than spit out a mathematical average of their input. Large language models, on the other hand, do not think outside the box.
10. Prompt ROI doesn’t always exist
Some prompt engineers sweat, tinker and fine-tune for days on the right AI instruction. A really well-crafted prompt can consist of several thousand words. In the worst case, the resulting output can only contain a few hundred words, only a few of which are actually useful. If you now have the impression that time expenditure and utility value are drifting apart here, you are right. (fm)
This article originally appeared at our sister publication Infoworld.com.
