Over the past month seeing several people commenting or posting an incorrect explanation for why LLM's give the wrong answer when asked: How many r's in the word "strawberry"
[#]LargeLanguageModels #LLM #GenerativeAI
=> More informations about this toot | More toots from bornach@masto.ai
The flawed explanations typically center around the initial tokenization stage masking the letters so that all the LLM sees is a vector of numbers
https://youtube.com/shorts/7pQrMAekdn4
=> More informations about this toot | More toots from bornach@masto.ai
If the tokenization prevents a LLM from seeing the individual letters that spell a word, then why are all LLMs from the big AI tech companies so good at creating acrostics?
In the same session that Gemini miscounted the r's it was able to write this:
=> More informations about this toot | More toots from bornach@masto.ai
The tokenization apparently didn't get in the way of Perplexity AI when asked to:
Write a sentence where the first letter of each word spells out "STRAWBERRY"
However change "first" to "second" and it completely loses the plot
=> View attached media | View attached media
=> More informations about this toot | More toots from bornach@masto.ai
The real explanation may be far more prosaic. There are plenty of forums where people challenge each other to create acrostics of different words. For example
https://coasterforce.com/forums/threads/acrostic-poems.37323/
These forums have most likely been scraped into The Pile or Common Crawl that is used to train Large Language Models. So more than enough examples for an AI to see past its tokenization.
However there are very few forums devoted to counting r's in words, or using the second letter of words to spell other words.
=> More informations about this toot | More toots from bornach@masto.ai This content has been proxied by September (3851b).Proxy Information
text/gemini