3 ms·

LLMs seem really bad with reading numbers and reporting them back. I’m building a game, and to se how well its docs were being indexed, I tried asking simple qu

LLMs seem really bad with reading numbers and reporting them back. I’m building a game, and to se how well its docs were being indexed, I tried asking simple questions to ChatGPT, Gemini, whatever Microsoft’s thing is, etc: “What is the armour value for the Leather Shirt” in the game Stravaeger?” It confidently got it wrong. “You can find the game at https://stravaeger.com https://stravaeger.com” Different confident answers, also wrong. “You’ll find it in a table on this page: https://stravaeger.com/docs.html?inventory_item=LEATHER_SHIRT https://stravaeger.com/docs.html?inventory_item=LEATHER_SHIR...“ Oh, sorry. I was inferring from other similar games. Here is a different confidently wrong number. “It’s also in the .json file linked on that page” And another wrong value. Random numbers should have got it right by now, but no. And the confident, authoritative tone never changed. Every model I tried was the same story.

joss82 6mo ago
Are you building the site from the same json files that are used in the game? AIs are still computer programs and are not given the resources to render javascript, so they cannot access the game data from the website. And they obviously don't have it in their parameters. BTW Google Gemini Pro just told me that they know the game but did not know the value. Actually, it points out that it is a known trap for AI to confidently give a wrong, hallucinated value. Maybe it had already seen this thread and integrated it in their parameters or fine tuning. I don't know... Gemini Fast confidently gives me a wrong value. But very quickly! I'll attach the entire Gemini response as a sub-reply.
- joss82 6mo ago
  [flagged]
- jasonkester 6mo ago
  The json thing is what makes it particularly surprising. It’s literally ’gameconstants.js’ with an item list that has a .name string and a .armor value that it could look up. But when I pointed it to the file and told it which row to read from, still chose to make something up instead of read the data designed to be read by computers. I do like how your ai tries to cover for its buddies with lies about there not being any documentation online. Because pointing AIs to a specific page and asking them to read the numbers there is a known trap.

joss82 6mo ago

Are you building the site from the same json files that are used in the game? AIs are still computer programs and are not given the resources to render javascript, so they cannot access the game data from the website. And they obviously don't have it in their parameters. BTW Google Gemini Pro just told me that they know the game but did not know the value. Actually, it points out that it is a known trap for AI to confidently give a wrong, hallucinated value. Maybe it had already seen this thread and integrated it in their parameters or fine tuning. I don't know... Gemini Fast confidently gives me a wrong value. But very quickly! I'll attach the entire Gemini response as a sub-reply.

joss82 6mo ago
[flagged]
jasonkester 6mo ago
The json thing is what makes it particularly surprising. It’s literally ’gameconstants.js’ with an item list that has a .name string and a .armor value that it could look up. But when I pointed it to the file and told it which row to read from, still chose to make something up instead of read the data designed to be read by computers. I do like how your ai tries to cover for its buddies with lies about there not being any documentation online. Because pointing AIs to a specific page and asking them to read the numbers there is a known trap.