Nobody fully understands why LLMs work.
The best analogy comes from Simon Willison, aka. LLM = alien technology.
We don’t know why a huge CSV file of numbers generates intelligent and useful responses, so we don’t know the limit of the model’s usefulness and intelligence.
Say you ask Midjourney to draw an image, but the result is garbage. You conclude that the LLM is garbage.
Yet, somebody else may be able to produce a fantastic image using the same model but with a slight variation of your prompt.
The model wasn’t garbage after all – your prehistorian brain’s ability to communicate with alien tech was!
One single word may completely change the output.
One word to go from garbage to world-class.
To prove this point, let’s examine my daughter’s most recent Midjourney article:
👉 [Blog] 50 Best Midjourney Style Prompts (New Version 6)
Here’s how the character of the generated image changed by altering only one word in a 32-word prompt (Midjourney’s new version 6):
REALISM:
PICASSO:
One word difference in a 30 word prompt!
Imagine what these models are already capable of – we haven’t even scratched the surface yet!
Even if we restricted our prompts to 100 words of English, there are
30000100
valid prompt variations given that many highly effective prompts are just bag of words.
It is estimated that there are only
1082
atoms in the universe!
Tell that to the many non-believers arguing prompt engineering is BS.