Why does chatgpt 4 suck at wordle?
I see a almost intentional. Like it gets tired of repetitive tasks? It says things like a 5 letter word with n in 4th position is pinch. Should be super easy task for a computer? Why say wrong things? It keeps apologizing and thanking for patience #tech #ai #openai
The same reason itās bad at typing a word backwards. It thinks in ātokenās, not words or letters. A token is a relatively common group of letters, with about 4 tokens every 3 words.
Current gen LLMs are next token predictors, you need infill prediction in queries required for Wordle (like the one you mentioned) which gpt-4 is not very good at.
Maybe good prompt crafting might help?
Tokenization strips the underlying characters from the input before it is trained. You can give it queries like "what is the nth character of this sentence: " and it will suck at that. It also has a loose grasp of whitespace since tokenization mostly strips that out
You would need to train it on content where shit is literally spelled out .. like "zoo is spelled z-o-o" then it might click .. but you can imagine there is not a lot of content like that in its training model.
Btw bing chat in precise mode is better at this .. they probably did some extra tuning for this use case
Itās an LLM not really ML. Itās not solving problems, just trying to assemble sentences that look like sentences itās seen on the internet
What do you mean not really ML. It's peak ML š
Then why itās so good at coding?