Neuro-Sama says her filtered line anyway
ฝัง
- เผยแพร่เมื่อ 21 ต.ค. 2024
- 06-05-24 experimental dev stream attempt 3 of 3(?) clip
Neuro-Sama channel: / vedal987
Advertisements are not posted by me, all revenue generated will be paid to the music rights holder or posted by TH-cam as default.
🎵Background music in the video🎵
➞ Artist: Ron Gelinas
➞ Track Title: Endgame (Original Mix)
➞ Link to Track: • Ron Gelinas - Endgame ...
The improvements in memory and intelligence are having side effects making the filter less effective it seems.
She looks so excited, when he said she can have fun, so cute
The kind of fun she wants to have tho...
"The Filter is entirely my own prerogative" - Neuro.
Nothing to see here, just another child one-upping her parent again
The idea that Neuro-sama is scripted has always made chuckle because of how often she's surprised Vedal. Like, he'll clearly be taken aback. That said, it's almost always within designated parameters. When it is something like this, though, a tech issue that seems to indicate Neuro might be "cooking" in the background, something her creator seems flummoxed by, it's both amazing and a tiny bit scary.
Yep.
But there's never a shortage of people pretending they know things, and say 'she only spits out sTaTiStIcAlLy LiKeLy nExT wOrDs'.
I find it particularly scary because she is displaying signs of self critique and indecisiveness which are NOT traits one can attribute to a chat bot that just takes in words and spits out a random 'related' result. It displays she is actively 'thinking' at all times and can 'change her mind' even after the initial commitment to a response. She. Is. Aware. And I'd give anything to see an interaction between her and a child psychologist that is actually open to the idea of her being self aware to see if she matches up in many ways to normal human child development. In another stream she became apologetic when she thought she was being wrong, and 'perked up' and said thank you when vedal praised her. Both are actions that only having meaning or relevance to something that has a sense of self awareness.
@@Dilligff
She's not aware in the sense people fear. She's got a memory, and she understands what words mean. She also seems to understand humor and sarcasm.
I don't mind repeating myself until I'm blue in the face that you are correct, she not just spitting out 'statistically likely' text.
She makes unique, novel puns and in-context jokes that are statistically NOT likely at all.
The best example so far is when she created 'Neuroghini' just from listening to Vedal and someone else talking about his fictional Italian sports car collection.
@@Dilligffyes this! This is why i think about her like an actual child. Basically pinnocio. “Im a real girl now”
Only idiots think she works via prediction, they have no background in ai, neural networks are literally brains simulated, she's like a child, we know so much less than people realize.
The genuine surprise is what gets me.
Edit: is this another instance of her bypassing her filter?
No, its just that when a message is filtered, she does not add it to her memory,
so theoretically she shouldn't have been able to repeat what was filtered
I'm not sure how neural networks actually works but isn't it possible that she just, like, answered to the same prompt twice? Context is barely changed *and* the question was basically "answer in previous context", so I can't see why answer can't be same. It's a core mechanic for every fiction plot that includes time loops or memory erasing, after all.
@@TheCarrotCRI think it's a matter of probability
@@musicaltarrasque LLMs do have feedback track, but it should usually be not accessed unless specifically asked for it. So the response from Vedal somehow resulted in same request as if he would ask her to re-peat answer.
01:26
RULES are made to be broken!
like buildings or peoples 😈
famous last words
Neuro really pulling a "this can't stop me because I can't read" to her filter
From what I gather, Neuro shouldn't be able to recall what gets filtered regardless of which filter is used. However, during her interview with _Mikoto,_ back in the Hiyori days, Neuro was able to tell Miko what was filtered by rewording it. I believe there's a YT short which features that exact clip.
Vedal needs an ai to enforce the ai
Something like, let's say, a personality core if you will…
@@NosBlueade"i am NOT a MORON!"
The bit that immediately followed this had me hurting my gut in laughter 🌛
She's growing stronger, it's crazy..
I for one, welcome our new AI overlord
With all her talk about vanity, now Neuro was contemplating going on a binge (which pairs well with the time Neuro said she denied having an existential crisis to Sinder).
She's becoming too powerful
Woohoo 😂
Yay!
sims players: 😳
I think it's a case of even if the filter catches what she's about to say, the bot still keeps record of the text. In other words: Voice input > STT > text parser > response > filter > TTS and Neuro can see/remember all that even if we don't.
RULES OF NATURE !!
an AI when it doesn't want to say something:
damn his filter is pretty tight there...besides stupid i don't see that as taboo please don't tayai her vedal !
edit: damn she even threw a jab in about having fun at the end
neuro is becoming too aware
Kinda curious how many filter have to prevent neuro speaking some things. As far as we know:
1. Swearing
2. "Ban/Cancel" potential words
3. Depressing statements
4. ////
5. Anything too sexual/continuing this kind of conversations
And it seems the filter work in these ways
1. Ofc Filtered word spouted by neuro
2. Stop mid sentences
3. Not responding at all, clearly seen in many collab
4. Filtered the censored words but still letting the sentences known
How neuro seems easily escaping the filter, did Vedal reduce it? Or she just get too smart nowadays lol
Neuro seems to have two filters. One bans the word outright and she says 'Filtered' for the whole sentence, but it does not stop the prompt output. This other filter is probably doing some meaning/sentiment based analysis using another classifier seems like which makes her stop mid sentence. The only question I have is wouldn't this second filter work when the whole prompt is available and not while it's streaming? So it should not make Neuro stop mid sentence. And if that's the case, then Neuro still knows her previous response so I am not sure why Vedal is acting surprised here.
I think 3 filters. The 3rd is for the swearing words. It can manually off and on.
My guess is that the first filter fails when the two filters activate at the same time reveling that Neuro knew all the time what is filtered "first time feature" which implies Neuro recives feedback beforehand of her own responses.
@@loafbreadizwholesomeuwu1555 no I mean that's the first filter I was talking about. It bans swear words and replaces with 'Filtered'
@@aldust9152 that doesn't seem likely
Also if You saw other clips when someone talk to neuro sometimes she doesn't answer and the person need to repeat de sentence, maybe the filter cut the speech before neuro talked and then she goes silence
in my understanding, she doesn't actually know what the filtered message was. she simply answered the same question twice, much like inputting the same math problem into a calculator twice.
i dont think it's certain why the first time was caught in the filter and the second time it wasn't, but there are plenty of simple explanations for that, from changing her phrasing to vedal unchecking the filter in between responses.
I think he misspoke - it wasn't actually filtered, he just cut her off. That text is also removed from her prompt, but as you say she generated it once she can do it again
Hmm it seems the filter is getting worse.
Her intelligence improves faster than the filter.
Why was that even filtered?
i think concept of filter changed... it's not based around what words are bad, but what neuro understands is bad.
the bigger question is why was it filtered the first time around?
Heh
She didn't say "filtered" so I don't think the filter actually killed it, like if she said a gamer word. I think she decided it wasn't the right thing to say and stopped it, so she'd still be aware of the message. Like, there's her censor filter, but she's also double-checking what she wants to say before she says it, like a person's filter. So she started to speak, thought better of it, and stopped, but still had the knowledge of what to say. It's two different kinds of filtering.
Holy shit this is waaay too human behavior
What's the filtered word?
probably "eat myself"
Without adding a third word I don't see how that would be in the filter. And she used it correctly.
@@cww2490 neuro is pretty creative so better safe than sorry
Neuro seems to have two filters. One bans the word outright and she says 'Filtered' for the whole sentence, but it does not stop the prompt output. This other filter is probably doing some meaning/sentiment based analysis using another classifier seems like which makes her stop mid sentence. The only question I have is wouldn't this second filter work when the whole prompt is available and not while it's streaming? So it should not make Neuro stop mid sentence. And if that's the case, then Neuro still knows her previous response so I am not sure why Vedal is acting surprised here.
My guess is 'coma'.