Comment by golfer
3 hours ago
There's lots of websites that list the spells. It's well documented. Could Claude simply be regurgitating knowledge from the web? Example:
3 hours ago
There's lots of websites that list the spells. It's well documented. Could Claude simply be regurgitating knowledge from the web? Example:
It didn't use web search. But for sure it has some internal knowledge already. It's not a perfect needle in the hay stack problem but gemini flash was much worse when I tested it last time.
If you want to really test this, search/replace the names with your own random ones and see if it lists those.
Otherwise, LLMs have most of the books memorised anyway: https://arstechnica.com/features/2025/06/study-metas-llama-3...
Couldn't you just ask the LLM which 50 (or 49) spells appear in the first four Harry Potter books without the data for comparison?
3 replies →
I think the OP was implying that it's probably already baked into its training data. No need to search the web for that.
This underestimates how much of the Internet is actually compressed into and is an integral part of the model's weights. Gemini 2.5 can recite the first Harry Potter book verbatim for over 75% of the book.
Do the same experiment in the Claude web UI. And explicitly turn web searches off. It got almost all of them for me over a couple of prompts. That stuff is already in its training data.
The only worthwhile version of this test involves previously unseen data that could not have been in the training set. Otherwise the results could be inaccurate to the point of harmful.
Honestly? My advice would be to cook something custom up! You don't need to do all the text yourself. Maybe have AI spew out a bunch of text, or take obscure existing text and insert hidden phrases here or there.
Shoot, I'd even go so far as to write a script that takes in a bunch of text, reorganizes sentences, and outputs them in a random order with the secrets. Kind of like a "Where's Waldo?", but for text
Just a few casual thoughts.
I'm actually thinking about coming up with some interesting coding exercises that I can run across all models. I know we already have benchmarks, however some of the recent work I've done has really shown huge weak points in every model I've run them on.
Having AI spew it might suffer from the fact that the spew itself is influenced by AI's weights. I think your best bet would be to use a new human-authored work that was released after the model's context cutoff.