4 ms·
I tried it with gpt 3.5 but it doesn't seem to know lojban so well. I asked it to translate the following in to Lojban, then separately after deleting that hist
by ftxbro 3y ago
I tried it with gpt 3.5 but it doesn't seem to know lojban so well. I asked it to translate the following in to Lojban, then separately after deleting that history I asked it to translate that to English. Here's what it did:
Original English: "I heard of the TinyStories research project that unlocked cognitive capabilities in LLMs with fewer parameters. I wonder if something similar would be possible with lojban?"
Lojban round trip English: "We interpret the use of the LLM as a way to confuse the knowledge from TinyStories.gy, which is greater than that from the experimental findings. I find it amusing to use Lojban instead. Does the ability to confuse the experimental findings also mean the ability to confuse Lojban?"
- sigstoat 3y agohmm since there is a CFG for lojban, it’d be interesting to give [some LLM you can do fancy things woth] what corpus of lojban there is, and then use the work on constraining LLMs to a grammar, and see if that’d help its output.
- int_19h 3y agoGPT-4 doesn't do well at it, either. The problem is that there simply isn't that much text written in Lojban, so it can't learn it the way it does other languages. You can see similar problems with other languages with a relatively small corpus, such as Old Norse.
- ftxbro 3y agoYes this must be it. I'm curious what it would look like, to compare more directly. Like a LLM trained only on Lojban vs. an LLM trained only on an equally small amount of English.