0:15 Ah! I know this one. No だ after an い adjective.

Yuta, but that's the problem. GenAI conflates and confuses enough things that overall diminishes and dumbs down humans. Teaching and learning should always be clear and make distinctions useful or valuable

9:49 ChatGPT's response isn't even consistent with itself

It says 港区 is (0) (heiban/flat), yet it says the pitch drops on a following particle. If that's the case, it isn't heiban: it should instead say (4) (odaka)


It's still wrong, but just inconsistently so

9:48 so chatgpt used a discussion board for pitch accent (because of course there is one), and one of the comments literally said "さらに言うと、「港区」は「＿￣￣＿」と表されるのが一般的なのですが". Maybe it found the website but couldn't figure out what _ and ￣￣ meant?

Thank you so much for the video Yuta! Have a great rest of your day!

ChatGPT has an issue even in English of sounding unnatural and forced. It has a tone that can easily be heard and repeats the same sentence structure over and over to a ridiculous degree. So I'm not surprised that it can't naturally correct Japanese 😅

ありがとうございます！

Sorry if this is a bit off topic, but ChatGPT says something like it's not very natural to say "Thank you for the stream!" in chat at the end of a livestream while "Thanks for the stream!" and "Thank you for streaming!" are very natural. I asked it about this several times, but I got a bit confused. I'd appreciate native speakers' opinions. Also, I feel that GPT-5.6 Luna isn't really good at logical reasoning compared to GPT-5.5

I asked it questions about archaic (warring states era, etc.,) Japanese and it did seem to understand that there are meaningful differences with modern Japanese.

do grok expert mode next!

I just tried the prompts in GPT Sol with intelligence cranked up to "pro" and the answers are worlds apart, in the first one it also starts lecturing about the weirdness of repeated "ndesu" sentences and makes suggentions to merge it all together with a "n desu kedo/ga" clause etc. It seems if you use maximum intelligence it seems you can really get far. BUT it took like 30-60s for one answer, which is kinda silly.

I tried it with Gemini Flash 3.6 and had much better results. It would catch most issues and give you good natural alternatives - and as the name suggests, it does it in a "flash"/instantly.

What it also didnt really catch was the "omitting the ga". But to be fair, while your previous video was super interesting and made a lot of sense to me and I surely learned a lot, 99% of all Japanese teaching channels, which are thought by natives and Japanese study books, made by natives, do mostly simply just say "dropping particles = casual speech" - so the LLMs can only know what is most commonly out there. If there are a few rare videos discussing this and a few recent ish papers explaining it, it will not suffice to go against the big "consensus", even if its kinda wrong. Had they invented LLMs in the year 1200, they would confidently tell us that the earth is flat.

i think something japan different word order in AI in transliteration. but, it kind of a level of speeches for example like "very casual", "casual", "neutral", "polite", "very polite", "formal", and "very formal" expression with different dialogue subtitles. but, there's one problem that especially for those japan honorifics like full name and nickname. and especially not english.

Do you have any thoughts regarding the use of ChatGPT for speaking practice?

I tried using it a few weeks ago for speaking practice since speaking is still my worst skill. I was trying to tell it a story. When I stopped at a sentence fragment expecting some kind of 相槌, but based on that little fragment I got out, it went on its very (much longer) rant.

Maybe I need to be able to speak more quickly, but this issue with ChatGPT actually pushed me to get over my social anxiety and speak to Japanese natives on VRChat instead. It's a bit of an effort to find people to talk to, but it was more productive.

and what should you use if you have no native speakers?

On rewriting perfectly natural part as well (写真を撮るのは禁止です to 写真撮影は禁止です). It constantly does this, even when I send it chats from a native Japanese person. It's good with obvious mistakes, but sometimes it's weird about what it chooses to focus on in my experience. Like 夕飯 being better than 夜ごはん...  I've even had my teacher correct me for using 夕飯 in casual conversation because it sounds a bit stiff lol

have you done a similar one with Claude? i have heard that Claude is a lot better at doing this kind of thing

10:17, haha.

こんにちは。綾瀬からみています！

Yuta, what dictionary do Japanese people use?

I don't like use ChatGPT  as written.