Plenty of fun to be had with LLMs.
Copilot seemed to be a bit better tuned, but I’ve now confused it by misspelling strawberry. Such fun.
That’s one example when LLMs won’t work without some tuning. What it does is probably looking up information of how many Rs there are, instead of actually analyzing it.
It cannot “analyze” it. It’s fundamentally not how LLM’s work. The LLM has a finite set of “tokens”: words and word-pieces like “dog”, “house”, but also like “berry” and “straw” or “rasp”. When it reads the input it splits the words into the recognized tokens. It’s like a lookup table. The input becomes “token15, token20043, token1923, token984, token1234, …” and so on. The LLM “thinks” of these tokens as coordinates in a very high dimensional space. But it cannot go back and examine the actual contents (letters) in each token. It has to get the information about the number or “r” from somewhere else. So it has likely ingested some texts where the number of "r"s in strawberry is discussed. But it can never actually “test” it.
A completely new architecture or paradigm is needed to make these LLM’s capable of reading letter by letter and keep some kind of count-memory.
the sheer audacity to call this shit intelligence is making me angrier every day