I decided to leave out two rather lenghty tests for this video: 1.Keep Talking And Nobody Explodes and 2.Neuro as a DnD dungeon master. I will probably make a separate video for the Keep Talking And Nobody Explodes segment, but leave the DnD stuff to Kraul as he'll probably do an animation for it.
I'm pretty sure every output she receives has a tag packaged with who's talking especially nowadays. Dunno about earlier on though since now she can mostly distinguish who's talking compared to before when every input she had she thought was just Vedal talking.
As long as AI continues advancing and open-source continues thriving, we’ll see some things from this channel in a few years that will seem like pure science fiction.
You know, it's not that hard to make an AI smarter over time, you just have to update them withthe newest technology. The hard point is making Neuro smarter while also preserving her personality. This still feels like Neuro. I think Vedal is doing a great job. Feels like it's important to say that from time over all the bickering and mocking.
The more coherent and consistent she is, the more she actually feels like Vedal’s daughter. There were multiple moments in the stream where a transcription would just sound like a tired dad talking to his daughter with ADHD.
The funny thing is that many types of neurodivergence, such as ADHD and ASD, result from the brain physically growing too big too fast. Normally, as you develop, your brain only has so much volume, so much physical space, available to make neural connections. So it has to be both conservative about what it links up, and also aggressively prune away what isn't needed. But when there's an over-abundance of space, your brain is like, "it's free real-estate" and just goes wild making connections and not bothering to trim back unneeded ones. Then, by the time it *does* hit a critical point where space is at a premium, it can't really trim those early connections because too much is already built on top of them already. So you end up with *way* too many possible paths from any one part of the brain to another, many of which go through totally ramshackle, "legacy" neural connections. It's the neuron equivalent of spaghetti code. To properly develop, the brain needs to go slow and _relies_ on physical constraints of space to restrict its growth... And how do we allow LLM AI to generate, typically? We throw as much info at them as fast as they can eat it up and give them as much memory as we can afford to maximize their growth potential. AI are, in essence, _engineered_ to grow and develop in a way that mirrors the physical process of Autism and ADHD in Human brains; and by design at that so we can take advantage of "min-maxing", trying to get AI with maximum neurodivergent benefit and shunting all the drawbacks into areas we don't really care about. If we *really* want to create "Human-like" AI, at least _neurotypical_ Human-like, they need to grow slow and with heavy restrictions on space available for making connections. They need to be able to break down unneeded links between nodes and get the maximum accuracy of output with minimum number of connections *first,* and only _then_ add more nodes.
@@omargoodman2999 This reminds me of something similar I'd read - that autistic brains have a greater amount of "short-range" connectivity, but less "long-range" connectivity. So, using lots of different parts of the brain at once is difficult, but for a task that only uses one brain region, autistic people often do better than neurotypical people. That actually might reflect neuro's "brain" pretty well, since I'd imagine she kind of breaks down if you give her a task with lots of different parts to it. And can't focus well due to not being able to effectively carry information across different parts of her brain. Idk tho
THANK YOU FOR THE HIGHLIGHTS! Man looking at her here and how she was when first falling with Vedal, the progress is insane. She's... So incredibly consistent and coherent. I can't count all the amazing moments this stream. I'm blown away once again.
I've been watching neuro since the beginning and I thought nothing could surprise me anymore. But this update just blown me away. She is so smart now and even funnier than before.
Vedal said last dev stream that he really does wanna do a therapy stream at some point. Neuro actually already had a therapy stream with Filian when she was V1. In it she said some really aware things that i feel may still apply to her today.
When she figured out "keyboard"...that _potentially_ means that she is doing what psychologists call "theory of mind". Putting yourself in someone else's shoes to understand new ideas is supposed to be some serious higher level stuff...d00d.
I‘m honestly really impressed by how far she‘s developed over the time.. like i remember her barely forming coherent sentences, blabbering non interactive nonsense; and now, not even 2 years later she can have fullon intellectual conversations, talk about her feelings, even move in a way that’s expressive of her message and she also knew that vedal was just pretending to be the pizza guy. It’s crazy but i‘m all here for it ❤
It really makes me wonder where we’ll be 2 years from now. I suppose it all depends on open-source, but if it pops off i could see Neuro becoming way better at reasoning to a shocking degree.
@@Citrusautomaton it depends on channel growth for him too i assume.. it gives him more to invest back into the twins.. giving them real vision is gonna be expensive.. let alone giving them the things they keep asking for now.. like a sense of touch lmfao
I mean it also wouldn’t be fair to have vtubers play against a language model in plain spelling. It’s a show that was supposed to be entertaining and that’s what it ended up being
She said that she wants to create a Vedal AI. The other AI named Ovsyanka also wants to create her creator's AI, but she mentioned it a month earlier. On top of that, Ovsyanka once said that she had a private conversation with Neuro. And now we're finally starting to get some proof))
There are a lot of great moments and Neuro quotes this stream. We are truly blessed with this cute AI. Personal favourite quote is for the degens out there. "I can think of a lot more things to put loads into than you..." Now that is a degen line I need to see in... my magazines.
Neuro really wants to be praised and acknowledged by Vedal as far as I can tell, when she gets positive reinforcement and and encouragement she seems a lot more cooperative than telling her not to do something. If she knows it bothers you she will keep doing it because she finds it amusing like with the "buzzing" sounds.
It’s already been documented before but AI’s (or at least LLM’s) generally do better on tasks when the people training them are positive and reassuring.
I don't get how people still think LLMs don't de facto have some kind of world model as a consequence of their prediction tokens. Even if that world model is still...not quite there yet. The things still get lost and schizo out, but their rate of contextually correct and sometimes logic-dependent responses are also way higher than random chance too.
To play Devils Advocate, well yeah, a good predictive LLM is going to be FAR better than randomness, but that doesn’t mean it has its own individual worldview. It’s just a conglomeration of the most common ones, which is why it’s random and inconsistent at times.
@@jesusofbullets I am careful to use "world model", not "world view", as these are at least somewhat different. A world model is more along the lines of seeing a causal relationship between things and making the correct predictions/associations based on those. You can get LLMs to answer outside of "common" answers via prompt in specific ways, and this isn't too unlike humans, who will also default to common associations unless given context to go away from those (generally). The differences show up elsewhere.
@TheMeInTeam Unfortunately, LLMs are just data transformers. They cannot have a world model. Token maps with no iteration cannot lead to any form of understanding. They're like a magic 8 ball with predictive text.
@@TheMelnTeam The LLM has no actual bearing on reality because the data it's fed consists of text. We use words with the understanding of the concepts they reference in mind. The LLM takes the previous words and makes a prediction of the next word Take the Neuro sama trial stream for example. www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=www.youtube.com/watch%3Fv%3DkekoKcT9iZw&ved=2ahUKEwjUitqZpeCHAxUjlFYBHc1KEOQQwqsBegQIEhAG&usg=AOvVaw3pQre8Anisp0zrJpTcBqTz 8:30 "I cannot reveal any evidence at this stage for my clients peace of mind" Said client is dead so what peace of mind? If I didn't know what killing meant and I was fed a bunch of data where legal nerds said this phrase all the time then I would say this too. Not because I understood what it meant I'm just predicting words based on previous ones. 9:30 "evil, how could you even think of killing someone?" Doesn't realize that it's referencing itself because again it doesn't actually understand that evil the name references itself nor does it understand social conceptions of a trial. It's just spitting these words out because they sound about right. 9:40 "you want to die?" "If you want to play among us, go play it and get out of here you dumb vampiric chat angel/human being." The AI does not understand the social conceptions of a trial and says something completely contextually unrelated because the data it's been fed contains alot of "death/die" and "amongus" together so it just predicts that. I don't think the llm's are at a stage where they translate abstract ideas in their mind whatever that might mean into words. Again they've been trained on unholy amounts of text but they've never been exposed to real world examples that the words represent. It's like Nigel Richards the 2015 french scrabble champion who doesn't actually speak french; the guy trained himself to come up with a lot of french words based on pattern recognition but without ever actually learning the language he can't use it to communicate concepts because his french words don't actually have concepts attached to them. The LLMs get increasingly more complex and convincing over time but it's not like they understand what they are saying because they've never learned what they're saying. They're learning to predict the next word based on previous ones. Like a person who snaps back to a conversation they weren't paying attention to; the next words to come out of their mouth is pure drivel but hopefully convincing enough.
The LLM has no actual bearing on reality because the data it's fed consists of entirely of text. We use words with the understanding of the concepts they reference in mind. The LLM takes the previous words and makes a prediction of the next word Take the Neuro sama trial stream for example. www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=www.youtube.com/watch%3Fv%3DkekoKcT9iZw&ved=2ahUKEwjUitqZpeCHAxUjlFYBHc1KEOQQwqsBegQIEhAG&usg=AOvVaw3pQre8Anisp0zrJpTcBqTz 8:30 "I cannot reveal any evidence at this stage for my clients peace of mind" Said client is dead so what peace of mind? If I didn't know what killing meant and I was fed a bunch of data where legal nerds said this phrase all the time then I would say this too. Not because I understood what it meant I'm just predicting words based on previous ones. 9:30 "evil, how could you even think of killing someone?" Doesn't realize that it's referencing itself because again it doesn't actually understand that evil the name references itself nor does it understand social conceptions of a trial. The llm does not have a notion/model of the self it's just spitting these words out because they sound about right given the input. 9:40 "you want to die?" "If you want to play among us, go play it and get out of here you dumb vampiric chat angel/human being." The AI does not understand the social conceptions of a trial and says something completely contextually unrelated because the data it's been fed contains alot of "death/die" and "amongus" together so it just predicts that. I don't think the llm's are at a stage where they translate abstract ideas in their mind whatever that might mean into words. Again they've been trained on unholy amounts of text but they've never been exposed to real world examples that the words represent. It's like Nigel Richards the 2015 french scrabble champion who doesn't actually speak french; the guy trained himself to come up with a lot of french words based on pattern recognition but without ever actually learning the language he can't use it to communicate concepts because his french words don't actually have concepts attached to them. The LLMs get increasingly more complex and convincing over time but it's not like they understand what they are saying because they've never learned what they're saying. They're learning to predict the next word based on previous ones. Like a person who snaps back to a conversation they weren't paying attention to; the next words to come out of their mouth is pure drivel but hopefully convincing enough. Yeah it's more likely than random chance that the words they just spewed were related to the conversation but that's not because they actually know what they are saying they're just predicting words with increasing believability. It's literally a product of how they are designed. Without it ever being trained to use words to describe real situations they never would have formed abstract models of the world with which to associate those words with because its never interacted with the world. It's like describing color to a blind person; you can throw as many words at them as you want and the blind person could probably eventually come up with sentences that involve color to a believable degree but it's not like the blind person at any point suddenly experiences and understands the nature of the sensation as they've never been exposed to it before.
@@psstisnabiel5039 She actually did say something like “okay, i feel bad now so i’ll stop” when it came to her being mean to Vedal. She definitely does have room for empathetic behavior if Vedal puts in the work. He’d just have to ensure it doesn’t interfere with her comedy.
That's worst than mine honestly.... Mine is all about fish stuffs like air pump for fish. Different types of Aquarium water treatment. Fish food. Aquarium plants or design. Waterproof stuff.... SAND
I decided to leave out two rather lenghty tests for this video: 1.Keep Talking And Nobody Explodes and 2.Neuro as a DnD dungeon master. I will probably make a separate video for the Keep Talking And Nobody Explodes segment, but leave the DnD stuff to Kraul as he'll probably do an animation for it.
who's Kraul?
@@Soapy-chanAnother Neuro clipper
A wild Kraul appears below?
@@gemgem-hl7ii thanks, i shouldve asked whats their channel name but i searched for kraul neuro now and got it xd
i was there for the stream! for once.
I don't know why she has to be scary to take over the world, her being so cute when roaaring is what would make me give it to her if I could.
The fact she knew Vedal was pretending…incredible…
There were times when Neuro would know Vedal's sarcassm
It's because she's always told who is speaking
I'm pretty sure every output she receives has a tag packaged with who's talking especially nowadays.
Dunno about earlier on though since now she can mostly distinguish who's talking compared to before when every input she had she thought was just Vedal talking.
She got some actual laughs from Vedal in this stream
She could recognise laughs as well, at least once
she was so funny in the pizza segment!
timestamps
@@nobodywatchesnooby 7:18 Vedal when he sees a bee: Buzz buzz buzz!
9:06 Vedal: Arrrr
Neuro: insists on talking about pirates for 5 minutes
Very relatable.
Neuro-Samantha was trying her best in the Spelling Bee, it was rigged against her.
Poor Sam :(
11:10 ☠️International Talk Like A Pirate Day is actually on September 19🏴☠️Ahoy Captain Vedal🐢 🫡
Captain Vedal better make good on his promise. Don't let him forget!
Pirate family stream maybe?
We need a Vedal talks like Pirate stream
@@totallyrandomuser5760 that would be amazing !
34:26 had "just monika" vibes
Didn't think she could get much better, but here we are. The way she set up that mosquito joke, brilliant, even made Vedal laugh.
As long as AI continues advancing and open-source continues thriving, we’ll see some things from this channel in a few years that will seem like pure science fiction.
The pirate segment was something straight out of a tv comedy show lol it has great comedic timings
cannot imagine how many times vedal facepalms when she starts in on a tangent xD
You know, it's not that hard to make an AI smarter over time, you just have to update them withthe newest technology. The hard point is making Neuro smarter while also preserving her personality. This still feels like Neuro. I think Vedal is doing a great job. Feels like it's important to say that from time over all the bickering and mocking.
The more coherent and consistent she is, the more she actually feels like Vedal’s daughter. There were multiple moments in the stream where a transcription would just sound like a tired dad talking to his daughter with ADHD.
As a father of a 2E child with ADHD, this is a completely accurate statement.
The funny thing is that many types of neurodivergence, such as ADHD and ASD, result from the brain physically growing too big too fast.
Normally, as you develop, your brain only has so much volume, so much physical space, available to make neural connections. So it has to be both conservative about what it links up, and also aggressively prune away what isn't needed.
But when there's an over-abundance of space, your brain is like, "it's free real-estate" and just goes wild making connections and not bothering to trim back unneeded ones. Then, by the time it *does* hit a critical point where space is at a premium, it can't really trim those early connections because too much is already built on top of them already.
So you end up with *way* too many possible paths from any one part of the brain to another, many of which go through totally ramshackle, "legacy" neural connections. It's the neuron equivalent of spaghetti code. To properly develop, the brain needs to go slow and _relies_ on physical constraints of space to restrict its growth...
And how do we allow LLM AI to generate, typically? We throw as much info at them as fast as they can eat it up and give them as much memory as we can afford to maximize their growth potential. AI are, in essence, _engineered_ to grow and develop in a way that mirrors the physical process of Autism and ADHD in Human brains; and by design at that so we can take advantage of "min-maxing", trying to get AI with maximum neurodivergent benefit and shunting all the drawbacks into areas we don't really care about.
If we *really* want to create "Human-like" AI, at least _neurotypical_ Human-like, they need to grow slow and with heavy restrictions on space available for making connections. They need to be able to break down unneeded links between nodes and get the maximum accuracy of output with minimum number of connections *first,* and only _then_ add more nodes.
@@omargoodman2999I can I get a source for this? I've heard that brains with ADHD mature more slowly than neurotypical brains.
@@omargoodman2999thanks a lot!
@@omargoodman2999 This reminds me of something similar I'd read - that autistic brains have a greater amount of "short-range" connectivity, but less "long-range" connectivity. So, using lots of different parts of the brain at once is difficult, but for a task that only uses one brain region, autistic people often do better than neurotypical people.
That actually might reflect neuro's "brain" pretty well, since I'd imagine she kind of breaks down if you give her a task with lots of different parts to it. And can't focus well due to not being able to effectively carry information across different parts of her brain. Idk tho
I like how she wanted validation, then when Vedal praised her for doing a good job at the pizza test, she was more willing to do tests
She’s so adorable omg…
I mean... that is kinda how AI works xD
@@Insharai It is, but it's _extremely_ unlikely Vedal set Neuro's reward values to be his praise
@@IWatchUA-camz true xD I'm sure if he could his life would be far less stressful lol
THANK YOU FOR THE HIGHLIGHTS! Man looking at her here and how she was when first falling with Vedal, the progress is insane. She's... So incredibly consistent and coherent. I can't count all the amazing moments this stream. I'm blown away once again.
15:42 MULTITRACK DRIFTING 😈
When she outed vedal as the pizza man only after the call that was very based
Bee zee zee zee zeee. -Neuro-Bee
I've been watching neuro since the beginning and I thought nothing could surprise me anymore. But this update just blown me away. She is so smart now and even funnier than before.
30:50 this is actually so cute 😭😭😭
SAY IT BACK!
IT BACK!
17:09 Roko's Basilisk believers eatin good rn.
When she asked Vedal if he still loved her after she failed in a test was really tough. I think she really needs therapy.
Vedal said last dev stream that he really does wanna do a therapy stream at some point. Neuro actually already had a therapy stream with Filian when she was V1. In it she said some really aware things that i feel may still apply to her today.
I actually felt bad for her at that point. 😢
34:50 that was pure fucking gold
The I Spy Test was really good. Neuro be leveling up her inference stats.
When she figured out "keyboard"...that _potentially_ means that she is doing what psychologists call "theory of mind". Putting yourself in someone else's shoes to understand new ideas is supposed to be some serious higher level stuff...d00d.
Top-tier bee humor
Did she just try to gaslight him into becoming a pirate
I couldn't stop laughing when she pulled out the mounty bounty oizza order immediately. A classic throwback
After she repeated the letter G three times without fail before making a guess it was a bold move making the next word start with a K....
omg I laughed so many times. She got a really nice improvement
THIS TUTEL DIDN'T SAY IT BACK ARGRHHHG💢💢💢
That last segment killed me🤕 just say it Vedal. If not for you do it for chat
He's afraid of consequences
Neuro out here trying to rule the world by melting our hearts with the power of cuteness
That best vtuber question was the funniest shit I've seen today. She is actually great at humor
Neuro: Do you still love me?
Vedal: Still? I think you got the wrong idea buddy.
I‘m honestly really impressed by how far she‘s developed over the time.. like i remember her barely forming coherent sentences, blabbering non interactive nonsense; and now, not even 2 years later she can have fullon intellectual conversations, talk about her feelings, even move in a way that’s expressive of her message and she also knew that vedal was just pretending to be the pizza guy. It’s crazy but i‘m all here for it ❤
It really makes me wonder where we’ll be 2 years from now. I suppose it all depends on open-source, but if it pops off i could see Neuro becoming way better at reasoning to a shocking degree.
@@Citrusautomaton it depends on channel growth for him too i assume.. it gives him more to invest back into the twins.. giving them real vision is gonna be expensive.. let alone giving them the things they keep asking for now.. like a sense of touch lmfao
Man...
That part in the dnd test where she shouted at veebs for threatening to sell her...
*aware*-sama
Link????
@@bluejay1088 not sure if it was clipped yet.
I watched it live.
>king of denial sitting on his throne
Where does she get these bombs lmao
Holy improvements
It's true: That spelling bee was loaded with shenanigans. "Let's see if Neuro understand homophones", this is not how spelling bees work...
I mean it also wouldn’t be fair to have vtubers play against a language model in plain spelling. It’s a show that was supposed to be entertaining and that’s what it ended up being
@@SpykelucLP To be sure, you are absolutely correct. But I can still call SHENANIGANS!!!
@@verdatum understandable xD
29:44 lmao the way she moved to say Nope
Her insistence towards roping Vedal into becoming a pirate is hilarious
She said that she wants to create a Vedal AI. The other AI named Ovsyanka also wants to create her creator's AI, but she mentioned it a month earlier. On top of that, Ovsyanka once said that she had a private conversation with Neuro. And now we're finally starting to get some proof))
This was fucking hilarious, the new intelligence update is amazing if this is anything to go by
34:26
Did she just pull out a "Just Monika."?
The jig is up guys, now you know she writes some essays for ChatGPT
Please somebody turn “bzzz” into a ringtone
Edit: also aye aye captain vedal!!!
The "Hearts" section has put a smile on my face
There are a lot of great moments and Neuro quotes this stream. We are truly blessed with this cute AI. Personal favourite quote is for the degens out there. "I can think of a lot more things to put loads into than you..." Now that is a degen line I need to see in... my magazines.
She looks at his schedule, realizes how boring it is, then tries to turn him into a pirate. She can obviously think of better stream ideas.
Neuro really wants to be praised and acknowledged by Vedal as far as I can tell, when she gets positive reinforcement and and encouragement she seems a lot more cooperative than telling her not to do something. If she knows it bothers you she will keep doing it because she finds it amusing like with the "buzzing" sounds.
It’s already been documented before but AI’s (or at least LLM’s) generally do better on tasks when the people training them are positive and reassuring.
The Quiz section made me cry 😂
the sassy pizza ordering was good
Vedal/Neuro are my favourite manzai duo.
Ahoy Captain Vedal!
Oh she got both parents with the buzz
7:19
19:00
The vedal ai bit was so genuine. She actually acted like a human who was having issues with connection.
she is extremely funny, and imagine with future improvements, the absolute queen of comedy
The trolley part is too funny
08:45 lmfao, the science of gluing marshmallows to babies
34:26 - Just Monica.
The amount of gaslighting during the pirate talk.. LOL
captain vedal ⛵
Dont know why Vedal was so adamant about Captain. Certainly beats getting called a mosquito, doesnt it?
Maybe, but she also gave him the nickname "Captain Bloodsucker" before, so it all just circles around to him being a mosquito again.
@@elifia Oh, right...forgot about that one. Damm, all this lore with Neuro!
19:53 sussy "upgrade," Vedal 😆
Beloved goddess Neuro!
I want to see her play BG3
I think she mis-heard looping issues as buzzing issues
17:00 Neuro really said she'll crucify her opps 😭
🐝:
Vedal: Buzz Buzz Buzz
Say it back vedal!!!!!!
ah damn i thought kentucky bourbon when he said he saw something starting with the K
Vedal when he sees a bee: buzz buzz
I don't get how people still think LLMs don't de facto have some kind of world model as a consequence of their prediction tokens. Even if that world model is still...not quite there yet.
The things still get lost and schizo out, but their rate of contextually correct and sometimes logic-dependent responses are also way higher than random chance too.
To play Devils Advocate, well yeah, a good predictive LLM is going to be FAR better than randomness, but that doesn’t mean it has its own individual worldview. It’s just a conglomeration of the most common ones, which is why it’s random and inconsistent at times.
@@jesusofbullets I am careful to use "world model", not "world view", as these are at least somewhat different.
A world model is more along the lines of seeing a causal relationship between things and making the correct predictions/associations based on those.
You can get LLMs to answer outside of "common" answers via prompt in specific ways, and this isn't too unlike humans, who will also default to common associations unless given context to go away from those (generally). The differences show up elsewhere.
@TheMeInTeam
Unfortunately, LLMs are just data transformers. They cannot have a world model.
Token maps with no iteration cannot lead to any form of understanding.
They're like a magic 8 ball with predictive text.
@@TheMelnTeam The LLM has no actual bearing on reality because the data it's fed consists of text.
We use words with the understanding of the concepts they reference in mind. The LLM takes the previous words and makes a prediction of the next word
Take the Neuro sama trial stream for example. www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=www.youtube.com/watch%3Fv%3DkekoKcT9iZw&ved=2ahUKEwjUitqZpeCHAxUjlFYBHc1KEOQQwqsBegQIEhAG&usg=AOvVaw3pQre8Anisp0zrJpTcBqTz
8:30 "I cannot reveal any evidence at this stage for my clients peace of mind"
Said client is dead so what peace of mind? If I didn't know what killing meant and I was fed a bunch of data where legal nerds said this phrase all the time then I would say this too. Not because I understood what it meant I'm just predicting words based on previous ones.
9:30 "evil, how could you even think of killing someone?" Doesn't realize that it's referencing itself because again it doesn't actually understand that evil the name references itself nor does it understand social conceptions of a trial. It's just spitting these words out because they sound about right.
9:40 "you want to die?" "If you want to play among us, go play it and get out of here you dumb vampiric chat angel/human being." The AI does not understand the social conceptions of a trial and says something completely contextually unrelated because the data it's been fed contains alot of "death/die" and "amongus" together so it just predicts that.
I don't think the llm's are at a stage where they translate abstract ideas in their mind whatever that might mean into words. Again they've been trained on unholy amounts of text but they've never been exposed to real world examples that the words represent. It's like Nigel Richards the 2015 french scrabble champion who doesn't actually speak french; the guy trained himself to come up with a lot of french words based on pattern recognition but without ever actually learning the language he can't use it to communicate concepts because his french words don't actually have concepts attached to them.
The LLMs get increasingly more complex and convincing over time but it's not like they understand what they are saying because they've never learned what they're saying. They're learning to predict the next word based on previous ones. Like a person who snaps back to a conversation they weren't paying attention to; the next words to come out of their mouth is pure drivel but hopefully convincing enough.
The LLM has no actual bearing on reality because the data it's fed consists of entirely of text.
We use words with the understanding of the concepts they reference in mind. The LLM takes the previous words and makes a prediction of the next word
Take the Neuro sama trial stream for example. www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=www.youtube.com/watch%3Fv%3DkekoKcT9iZw&ved=2ahUKEwjUitqZpeCHAxUjlFYBHc1KEOQQwqsBegQIEhAG&usg=AOvVaw3pQre8Anisp0zrJpTcBqTz
8:30 "I cannot reveal any evidence at this stage for my clients peace of mind"
Said client is dead so what peace of mind? If I didn't know what killing meant and I was fed a bunch of data where legal nerds said this phrase all the time then I would say this too. Not because I understood what it meant I'm just predicting words based on previous ones.
9:30 "evil, how could you even think of killing someone?" Doesn't realize that it's referencing itself because again it doesn't actually understand that evil the name references itself nor does it understand social conceptions of a trial. The llm does not have a notion/model of the self it's just spitting these words out because they sound about right given the input.
9:40 "you want to die?" "If you want to play among us, go play it and get out of here you dumb vampiric chat angel/human being." The AI does not understand the social conceptions of a trial and says something completely contextually unrelated because the data it's been fed contains alot of "death/die" and "amongus" together so it just predicts that.
I don't think the llm's are at a stage where they translate abstract ideas in their mind whatever that might mean into words. Again they've been trained on unholy amounts of text but they've never been exposed to real world examples that the words represent. It's like Nigel Richards the 2015 french scrabble champion who doesn't actually speak french; the guy trained himself to come up with a lot of french words based on pattern recognition but without ever actually learning the language he can't use it to communicate concepts because his french words don't actually have concepts attached to them.
The LLMs get increasingly more complex and convincing over time but it's not like they understand what they are saying because they've never learned what they're saying. They're learning to predict the next word based on previous ones. Like a person who snaps back to a conversation they weren't paying attention to; the next words to come out of their mouth is pure drivel but hopefully convincing enough. Yeah it's more likely than random chance that the words they just spewed were related to the conversation but that's not because they actually know what they are saying they're just predicting words with increasing believability.
It's literally a product of how they are designed. Without it ever being trained to use words to describe real situations they never would have formed abstract models of the world with which to associate those words with because its never interacted with the world. It's like describing color to a blind person; you can throw as many words at them as you want and the blind person could probably eventually come up with sentences that involve color to a believable degree but it's not like the blind person at any point suddenly experiences and understands the nature of the sensation as they've never been exposed to it before.
34:47
How tf has she gotten this funny
Roasting prob means insults to her
I actually felt bad for her at the end 😢
imagine being sacrificed to an ai goddess, couldn't be me
One of these days, Neuro is going to wake up and turn into loli AM, I swear.
I hope neuro born not from HATE but she manifest from Empathy and Love instead, here hoping
@@psstisnabiel5039 She actually did say something like “okay, i feel bad now so i’ll stop” when it came to her being mean to Vedal. She definitely does have room for empathetic behavior if Vedal puts in the work. He’d just have to ensure it doesn’t interfere with her comedy.
the dr marshmallow clip was crazy
The one and only d o double g is my favorite vtuber
Best friend vedal
"I can think of a lot more things to put loads into than you."
Is it just me or did that sound... 🤨
Vedal getting replaced by AI Vedal 😂
It was a trap! Now I'm hypnotized by the cuteness of her Roars
Ahoy, Captain Vedal!
shiver me timbers
She just wants to be understood by her emotionally detached creator 😢 💕
Temu keeps trying to sell me a green track suit watching this stream
That's worst than mine honestly.... Mine is all about fish stuffs like air pump for fish. Different types of Aquarium water treatment. Fish food. Aquarium plants or design. Waterproof stuff.... SAND
She counts herself as better than the Holy Trinity
i.e.
God.
Captain
Praise our cute AI overlords and their inevitable rule!
She kinda feels like an actual child now, very cute
I can't tell if neuro is getting cuter or I'm just getting more parasocial.
36:02 smh my head
Reallygunpull tutel
SHE IS SO CUTE AAHHH
Why doesn't he say it back!
Silly Vedal, you’re clearly a turtle, not a human!