It’s not always wrong. What I’m saying is that it needs to be correct often enough, not always. And evaluating whether that’s the case or not is difficult and a bunch of work. Like I wrote before, just last year it was bad enough to be completely wrong about basic stuff. It has improved by a lot since then and it will probably improve more in the future.
Apart from corretness, one big problem with LLMs is that they always sound super authoritative and confident in their answer. A normal human does not, and drunk humans are also generally recognizable as drunk. But a hallucinating LLM is not recognizable as hallucinating unless you know the topic it’s hallucinating about.
For example, I previously mentioned that last year ChatGPT would claim that 日曜日 is read as hiyoubi. This was a common issue with Japanese because if you train the models only on normal text, then it will often not include the readings, and LLMs used to be completely oblivious to differences in readings. This has improved by a lot since then.
But something that still happens is that it will incorporate information from the question into its answers. If I ask it about my friend from Aomori claiming that their word for Sunday is hiyoubi it’ll write that it’s possible and that it’s how dialects work.
It’ll also still be confident even when it flip-flops during a conversation. For example, I had roughly this conversation with ChatGPT (I’m abbreviating):
Q: What was the Japanese word for Sunday 500 years ago?
A: 日曜日, same as today.
Q: I thought it only appeared less than 200 years ago.
A: Yes, Japan only introduced the Western calendar during the Meiji Restoration, and the concept of Sunday as a distinct day of the week did not exist before then.
Q: So, 500 years ago, it was not 日曜日 after all?
A: Correct.
It’s not only flip-flopping - all the answers are wrong, but all of them sounded very confident. And this can happen with any random topic. So to claim that ChatGPT is good for learners I would want to have a reasonable corpus of things that learners would actually ask it, and then have to evaluate the answers it gives, and whether it will stray from the truth in follow-up questions, and so on. That’s a ton of work.
Also, what do you make of this conversation (not abbreviated)?
Q: Is 御礼 read as おんれい in Japanese?
A: No, 御礼 is typically read as “おんれい” (onrei) in Japanese. It means “thanks” or “gratitude.”
Is it disagreeing or agreeing? Is it correct? I thought it was a pretty simple question…
Just in case anyone cares...
Days of the Week:
It’s true that the Western calendar with the current week days was only formally adopted during the Meiji restoration. But the word 日曜日 appeared before that, approximately 200 years ago. Before that time it would be written differently, but because the same week days were used in ancient China and in the Buddhist calendar, words for them did exist. Probably the oldest mention of the modern days of the week in surviving records is in the dictionary of Fujiwara no Michinaga, which is approximately 1000 years old.
Reading of 御礼:
While we probably all learned this as おれい, the reading おんれい exists in more formal contexts.