so many replies i probably said something 4 other people already said lol
I’m in no way an expert but this Veritaserum video explains one way to go about predictive text (the whole video is interesting but it’s around the 25’ mark). If you think Google auto complete somehow understands your behavior, somehow GETS what you’re interested in, then you’d be wrong.
Plus this.
It’s all math, statistics and algebra.
Say you’re playing shiritori. An AI will probably never fail to provide the most correct answer. A human, capable of intuition and creativity, may find new ways to keep the game amusing. I’d rather play shiritori with a human for the laughs.
LLMs may show signs of understanding but they do not operate on genuine comprehension, unlike humans. The fact some people are not aware of this, and somehow buy all the Tech giants BS is worrying to say the least.
11 Likes
ooo veritasium, i like these types of youtubers.
yea idk who would ever enjoy playing a game of shiritori with ai lmao, it sounds too boring.
yea i know LLMs do not know everything or are 100% right but some are still good enough to practice with.
2 Likes
That’s the problem with AI, people are very polarized about it.
AI enjoyer and users wants more of it everywhere and AI hater will push against any mention of it (even algorythm and machine learning.)
I am still waiting for the mythical LLMs used to find then treat cancers and other medical advances that we were promised 4 years ago…
PS: do that bingo card, another AI thread will probably be posted this year so it will be handy.
6 Likes
I wonder if it’s fine for n5 or n4 content, probably has enough ressource to not get those wrong but error will pile up later on.
Plus you might not understand when ChatGPT is wrong at an early level.
Is it the worst thing ever? No as you will probably learn that it’s a mistake later on but it’s still annoying and I would personnally only trust certified source as bunpro and textbook.
But if works for you and fell like you improve quicker, then that’s great.
Use whatever tool you like but don’t force it on others 
3 Likes
Went to bed, woke up to 33 new notifications from this chat 
1 Like
I know this is not the point of this thread but thank you for saying this lmao the cram UI is a bit arcane
1 Like
Yea, I hope it gets beginner grammar right 
idk about more advanced grammar though
1 Like
I KNOW, THATS WHY ITS SO UNUSED. It’s soooo annoying to select what you want and choose specific tabs
1 Like
Can we talk about the gen-AI in the B-Point shop though : / Not a big fan of that either.
2 Likes
Not sure if you are referencing someone else mentioning using it in the shop, or that you are saying you think the shop stuff uses it.
If the latter, all of the graphics/assets on the site are human made, including everything in the B-Point shop.
14 Likes
Actually, it is really bad to learn something wrong. That’s why we have tons of material on breaking bad habits!
If you learn something incorrectly, at the bare minimum, you need to relearn it. That’s double the effort.
If AI hallucinates at 20% then over 1000 interactions you would see the following results.
Have 800 items learned correctly the first time
180 items learned correctly the second time (360)
18 items learned correctly the third time (54)
2 items learned correctly the fourth time (8)
1222 / 1000 or 1.222 interactions per item learned. A 22.22% efficiency loss.
Keeping in mind that free AI models are not particularly powerful and that Japanese training data is far lower volume than English, I think a 20% hallucination rate is very very generous.
7 Likes
This is probably a bad time to admit I am actually generative AI and not a human. hello its cafelattebot here
(this is a joke please dont ban me)
6 Likes
a frontier model for sure!
1 Like
I’m sorry, all the pieces by Darts in the shop give all the hallmarks of a photo put together using AI generated assets then edited.
TBH they just look like graphic design pieces to me
Maybe the worst thing AI has done is sow this kind of discord where we accuse people of using AI haha
7 Likes
The point of an llm is predicting the most likely sequence of words that would follow an input sequence of words, based on the statistical makeup of the data it’s been trained on. It has nothing to do with correctness, accuracy or information gathering.
You don’t need an llm for this. Also, algorithms are not ai.
9 Likes
I agree but most AI user will take that chance.
What I meant is that sometimes you will learn things wrong (especially at a beginner level) so it’s not too problematic as you will quickly unlearned it or improve on it.
I still make grammar mistakes writing in my mother tongue after 30+ years.
So we always make mistake and keep improving.
Main problem with AI if it makes mistakes but doesn’t improve so you have to hope the next update will fix it or catch the mistake.
I tell everyone I know that want to use AI is taking risk of learning wrong grammar or words and better ressource are available but I don’t want to police them.
I get the allure of AI, it’s easy and very convincing and it probably get it correct 80 to 90% of the time, people using AI will accept it.
20% sounds really scary, but if the prompt is correct and doesn’t demeand to much it can be reduced soo soo much.
We have a telegram bot that makes a ditailed explanation for words, but it’s within 200 tokens, uses gemini mini 3. It’s build so well that it really seems to now make any mistakes. Of course, sometimes it can be broken by sending it non existant word, but it wont make a mistake on just explaining a simple word that is just somewhat rere.
I woulnd’t go as far as to learn a sentences it give, as it has no soul to it, but the amount of mistakes good llms make with a good prompt is less then a human being.
Of course not saying that we should add some llm to bunpro, I woundn’t like it because I become kind of dizzy as my brain has to assess if what llm is saying is wright or wrong, natural or not, 100% of the time I use it. Plus it seems to not have any usecase for now, as a lot of folks has already mentiond.
As for me, the problem with llms was not to lear the wrong thing, but to activelly doubt 100% of what I’ve learned from llm. That is why it just woundn’t stick with me, anything that is the product of ai stays only as a passive knoladge that I have to confirm, like a theory I’ve heared one time from not very trustworthy friend and I just won’t say it to no one before I’m sure it wasn’t just his mad imagination.
For me it’s ok to have some of it, but I cannot imagine learning the whole language like that, it’s just tiering, even if 99.5% of it is write, my brain won’t stop looking for those 0.5% in all 100% on the background.
1 Like