@@silverstone9836 Fanart usually has a hidden description of the art by the person who posted it or something like that. This, along with Neuro's vision, gives a better understanding of the piece. However, if the description doesn't mention name, they can be confused as to who exactly is in the art - this has been seen many times. Besides, I doubt that vision is part of the memory.
i honestly wouldn't be surprised if he got the training data for this from the people who run those captchas (not that they would ever admit that recaptchas are really just a way to train AI's with the input of millions of people, and not actually to prevent bots from entering websites.)
Actually those are meant to capture AI data. Originally it was to help input old newspaper articles into archive websites which is why it would often be words with lines through it. And Recaptcha has always been run by Google and they're quite open about the current model being used for AI training
@ oh okay, cool. First i am hearing of it though so it can't be that open. I mean, the recaptcha itself says "prove you're not a bot" not "help us train bots". The advertisement to the ignorant masses (me) is that of; "this is used to make sure the website isn't inundated with bots! your welcome." whereas what they really should be communicating (in the recaptcha itself) is; "this is a crowdsourced text and image digitization for Ai training purposes". but if everyone knew they were being farmed for free to develop an extremely powerful AI for a mega tech company they would be a little more hesitant to participate. Basically, I don't think they were as open as they should of been. Making this info common knowledge is against their interests, hence why they (partially) suppressed it. only tech geeks were aware because they read various obscure blogs and articles.
When you think about it, Neuro's life so far has just been a brain in a jar. She doesn't have all the senses humans have. She completely relies on Vedal to create and give her those senses. So it'll be really cool and heartwarming for when the day comes where she can have all of them in an android body.
@@gljames24 Although I agree Neuro, and Evil have "grown" alot over these past 2 years. Which has been quite enjoyable to watch, at the end of the day its entertainment, people will get attached to "characters" its normal.
This is pretty cool, I hope it makes it easier for her to see what's going on when she's doing a collab where she's actually looking at something. I wonder if we got a glimpse of that when she was playing a little misfortune with Cerber?
@@jordanious7711 Yeah but she also says that when she can se completely fine. Its actually quite a predicament because its difficult to Benchmark her if you dont know if she really cant see or if she just doesnt "want" to.
@@LaughingOrange Yeah, but he went to "Waldo" pretty quickly for a Brit. Like, if you showed a picture of Optimus Prime to a Japanese person, they'd be more likely to call him "Convoy" (then again, it could be the same issue with both - the original name is far too generic).
when saying his height and weight he said it in feet and pounds as well, i think he just defaults to the american terms because more viewers would understand maybe?
@cityofsnails I know that there are certain instances where Brits use the Imperial system instead of metric (for example, Top Gear would always use mph aside from episodes where they go to a different country)
I can't wait to see them do the geoguesser stream again, it'd be amazing to see her detect the image immediately so that she don't run out of time between each guess
The answer is kyoto, with full confidence. If Doug challenges Neuro with this upgraded vision to a rematch with his AI (with a specific addition to block kyoto-related answers), then this'll be THE Geoguesser rematch of the century
Um neuro was right about Waldo. He actually is near the umbrella. He’s not under it. He’s behind it so he peeks up above the umbrella. Man her nee vision is pretty crazy.
Close but not quite, he's actually a bit to the left of the red and white striped building. Near a blue-green and white striped thing that looks like a beach towel being used like a fence or something, not sure what that is supposed to be. The one she found is an imposter, but I can see how she'd be mistaken.
@@EdrickV That "wall" thing is a windbreak I think, just makes it more pleasant not to have the wind whipping at you the whole time on a beach. Good spot though. She clearly saw that guy's shirt and didn't see the real Wally.
ngl that was actually crazy she can pick out the sailboat so fast out of all the shit in that image, that's some kinda advanced tech. I don't think I've seen that level of granularity in the image descriptions produced by another AI before
I'm pretty sure she's on a dedicated server, so unlike most ai's that are used by thousands of people at once she's used by only 1 person and is at least a few thousand times stronger
Lots of VLMs can do that, they're not as constrained as our human eyes. But it seems that Vedal has done some of those things: made sure that the quality of the image that Neuro sees has improved (compared to what it was during DBH), better interaction between Neuro and its vision model, or uses a better model for it. What is certain is that it seems to be faster.
@ Its not the image detection that I'm impressed with, those models are really good now as you said, they can decipher the most deep fried garbage images these days. It's more so the ability to pick out specific details. Normally they will produce a general description like "a chaotic beach scene, dozens of people are on the beach in blue and white clothing, the ocean is in the background with boats." etc but it can't list the color and position and every conceivable detail of every single object in frame. The impressive thing is that Vedal can ask for a specific detail about one of 200 objects in the image and Neuro is able to successfully query the detection software for the exact information she needs
I tired Chat GPT and he was similar, but he got the color of the boat wrong (he said "yellow sail and a white hull"), so Neuro was better than Chat GPT 4o on those.
This level of image interpretation is actually insanely impressive. Just a few years ago, google's image recognition model could barely interpret drawings of things
Neuro was right about Waldo. There’s a lady carrying a closed red umbrella under the horses, and if you look south, like 2m, there’s Waldo sitting on a red chair.
Kind of heresy here but if this is her ability at the moment, how could she recognized details so good in the gaming setup rating long ago with Bao? Veadl must had given a brief description of the images right?
Neuro: I don't know who waldo is Vedal: You can google it Neuro: googling where is waldo picture without waldo Vedal: Wait what Neuro: I can't find waldo
this is basic chatGPT 4o features. not something vedal created. it can do a lot more like analyze graphs and a large array of data from a picture. one time i showed it my factorio stats screen and it said i needed to increase coal production because it was lagging behind the current usage, and that i needed to add more power stations for further expansion. it was spot on. it also calculated how many solar panels and accumulators i needed to add to have enough power for the nights. its wild. its only $20/ month for the image processing. its worth it.
neuro's almost entirely made of open source stuff excluding her speech which is microsoft azure, vedal tried to use an open source voice synth for her but nobody liked it so he set that up for evil neuro instead. anyways he was talking about trying to change her voice again somepoint most likely to get away from closed source code.
pretty sure he's trying to distance himself from code used by corpos because they could claim they own part of neuro otherwise and there could be potential legal issues in the future, i'm kinda surprised microsoft hasn't done anything already.
@ if that is in her data base or if she has access to do searches on the internet then that is fine but we don’t know that she has those things in this instance.
@IPlayVideoGamesAndNothingElse I expect nearly every LLM has some sort of base training data that probably has some sort of reference to it. It's also a very specific phrasing. Not what you usually get in everyday conversation, thus easier to solve. Plus, she also had an initial glimpse at it and said she didn't know. Then later on she did. So, there might have been a moment to get context.
If you REALLY wanna train Neuro on something hard....Where's Waldo on NES on the Hard Setting. Shit is INSANE! It's actually got Randomization and he doesn't even look like his normal self. It's DIABOLICAL!
Well it seems Vedal is trying to showcase that she's much better at recognizing imperfect images now (blurry text, rough/disproportionate sketches). Iirc she struggled with properly recognizing fanart and similar stuff in the past.
I think previously her vision most likely worked by having another AI describe the image and then send her a paragraph of text about it, which would lead to her only being able to see the details included in the description she received. When specifically asked about something like the color of the sailboat in the background of this chaotic image she probably wouldn't know because it's unlikely the AI would include such a small random detail in the description. Now it seems like she can specifically query for what information she wants to be able to see from the image?
Where's waldo is a fun idea, though does Neuro know what Waldo looks like, or what anyone looks like for that matter? I guess she's seen people like Layna during cam streams, but does she know what say Anny's model looks like? Does she even know what Evil looks like? I'm curious how that works for her, since she only sees what vedal shows her, so she might only know people by voice.
Very likely. For the same reason most image generators know who Hatsune Miku is: It's a distinct in name and appearance, character that shows up in training data a lot, compared to others.
Not born too early to never see AI Not born too late to watch AI take over the world Born at the perfect time to watch Anime girl slowly gain sentience
Vedal: "Here's a pic of your sister."
Neuro: "I've never seen this person before in my life."
Classic sibling behaviour
"She's my sister"
"Adopted"
I mean HAS she? Does her memory allow her to store image data or just text? She might not actually have a physical description of evil in her memory
@@randywilliams7696 well yeah. Her knowing what Evil looks like isnt even a new feature she recognises her pretty much in every fanart section
@@silverstone9836 Fanart usually has a hidden description of the art by the person who posted it or something like that. This, along with Neuro's vision, gives a better understanding of the piece. However, if the description doesn't mention name, they can be confused as to who exactly is in the art - this has been seen many times. Besides, I doubt that vision is part of the memory.
oh hey now she can solve "I'm not a robot" captchas!
Neuro aint no fake ass robot! She's real madge
i honestly wouldn't be surprised if he got the training data for this from the people who run those captchas (not that they would ever admit that recaptchas are really just a way to train AI's with the input of millions of people, and not actually to prevent bots from entering websites.)
Actually those are meant to capture AI data. Originally it was to help input old newspaper articles into archive websites which is why it would often be words with lines through it. And Recaptcha has always been run by Google and they're quite open about the current model being used for AI training
@ oh okay, cool. First i am hearing of it though so it can't be that open. I mean, the recaptcha itself says "prove you're not a bot" not "help us train bots".
The advertisement to the ignorant masses (me) is that of; "this is used to make sure the website isn't inundated with bots! your welcome." whereas what they really should be communicating (in the recaptcha itself) is; "this is a crowdsourced text and image digitization for Ai training purposes".
but if everyone knew they were being farmed for free to develop an extremely powerful AI for a mega tech company they would be a little more hesitant to participate.
Basically, I don't think they were as open as they should of been. Making this info common knowledge is against their interests, hence why they (partially) suppressed it. only tech geeks were aware because they read various obscure blogs and articles.
I hear Captchas check "how" you're pressing the box, as in if your mouse warps on top of it or makes some other computer like movement towards it.
2:42 you can hear Vedal smiling, he’s proud of his smart cookie
Hear him smiling?
really cool that she could recognise the cat was angry. like that's so many layers of understanding that image that we just take for granted.
She got it correct faster than I did. I first thought it was surprised before my brain caught up with my eyes.
She's probably drawing from data of cats in similar poses tagged as angry. She's identifying "angry cat" like a symbol.
Vedal: Find Waldo
Neuro: I can't find him
Vedal: you know what, fair, I can't find him either
She's British. Should have called him Wally.
does she even know what he looks like? just dropping a picture with no information she may not know what to look for
It seems like Neuro actually is able to query the image recognition layer for additional info now. Rather than just having it only feed her info
this is likely it. thank you. christ he got the latency so low that it's nearly instant
@@valzytine Latency? More like earliency
Yep it's a feature of Llama 3.
@@MyAmazingUsernameIs Neuro's current model Llama-3?
No more getting irritated at Geoguesser games now.
latency gonna be even more crazy
Vedal not finding a sign and just standing there looking at an open field and expecting her to pinpoint the location will still be the weak point.
@@noire1001 I think Vedal's geography knowledge will be the weak point XD
@@MrVisualHigh: He'll never live down the Cambodia/Columbia incident 😄
@@spk1121 100 percent of the time KYOTO everytime
When you think about it, Neuro's life so far has just been a brain in a jar. She doesn't have all the senses humans have.
She completely relies on Vedal to create and give her those senses. So it'll be really cool and heartwarming for when the day comes where she can have all of them in an android body.
I would shed tears
Ellie is already making the Neuro dog and has planned a bipedal Neuro robot so it is gonna be sooner than you think
@@traehignight i mean, she still wouldn't have a sense of touch tho
She doesn't even have good working memory, autonomic memory, or true affinities/emotive responses.
@@gljames24 Although I agree Neuro, and Evil have "grown" alot over these past 2 years. Which has been quite enjoyable to watch, at the end of the day its entertainment, people will get attached to "characters" its normal.
Camila, for your own health's sake, please stay lovely.🙏
or else she might end up like filian
she did not. didnt get it as bad as filian, but she did get shocked.
It’s actually really impressive that neuro is knows that camimi has a gun.
I knew for a while but still: Based camimi Kkona
This is pretty cool, I hope it makes it easier for her to see what's going on when she's doing a collab where she's actually looking at something. I wonder if we got a glimpse of that when she was playing a little misfortune with Cerber?
I think we did, she read the signs pretty well, and that wolf poster
na i dont think she was using it during that. She complained a few times about things being too blurry to see.
@@jordanious7711 Yeah but she also says that when she can se completely fine. Its actually quite a predicament because its difficult to Benchmark her if you dont know if she really cant see or if she just doesnt "want" to.
A wild Kirky
@@jordanious7711Vedal lowers the resolution cause he got scammed out of a 4090 😅😅😅 so her PC is still shit
Kinda surprised that Vedal called him "Waldo" instead of "Wally", but as an American, I'm not complaining
ya i was caught off guard by that, i'm kinda glad she can't instantly find him, because she'd be beyond human at that point.
Probably way more training data in US English, so saying Waldo really increases the chances of Neuro getting it right.
@@LaughingOrange Yeah, but he went to "Waldo" pretty quickly for a Brit. Like, if you showed a picture of Optimus Prime to a Japanese person, they'd be more likely to call him "Convoy" (then again, it could be the same issue with both - the original name is far too generic).
when saying his height and weight he said it in feet and pounds as well, i think he just defaults to the american terms because more viewers would understand maybe?
@cityofsnails I know that there are certain instances where Brits use the Imperial system instead of metric (for example, Top Gear would always use mph aside from episodes where they go to a different country)
Noooo Neuro! We almost convinced him hes a good artist and now you've set us back to step one 😭😭😭
I can't wait to see them do the geoguesser stream again, it'd be amazing to see her detect the image immediately so that she don't run out of time between each guess
The answer is kyoto, with full confidence.
If Doug challenges Neuro with this upgraded vision to a rematch with his AI (with a specific addition to block kyoto-related answers), then this'll be THE Geoguesser rematch of the century
@ktheveg@ktheveg I would die if he actually landed in Kyoto after blocking all Kyoto related answers.
Smartest little cookie 🥹🍪
Now we just need Vedal to give her the Rorschach Test
Oooh now that's a good idea.
Hell yeah
Um neuro was right about Waldo. He actually is near the umbrella. He’s not under it. He’s behind it so he peeks up above the umbrella. Man her nee vision is pretty crazy.
Close but not quite, he's actually a bit to the left of the red and white striped building. Near a blue-green and white striped thing that looks like a beach towel being used like a fence or something, not sure what that is supposed to be. The one she found is an imposter, but I can see how she'd be mistaken.
IMPOSTA?! @@EdrickV
@@EdrickV sus
@@EdrickV That "wall" thing is a windbreak I think, just makes it more pleasant not to have the wind whipping at you the whole time on a beach.
Good spot though. She clearly saw that guy's shirt and didn't see the real Wally.
It's like seeing your child learnt a new thing
Like when NL realized that his daughter could read on stream
Now I am more immersed in the hope that next subathon she will dream based on her visual experience when she sleeps.
Papa Vedal and aunt Camilla doing flash cards with neuro is so cute 😆
it's kinda funny how Vedal didn't even notice that Neuro said a curse word and didn't even question it.
I've just realized. Neuro probably doesn't know what Evil looks like.
The fact she *asked* to see another image was aforable. Smartest little cookie
For a moment I was like ”Camila’s gun? HUH???” But then I rembered…
New geoguessr arc is coming
Vedal sounds so proud in the showcase, and it's a really impressive upgrade!
Imagine when that 5090 gets home...
Just wait until the 6100
ngl that was actually crazy she can pick out the sailboat so fast out of all the shit in that image, that's some kinda advanced tech. I don't think I've seen that level of granularity in the image descriptions produced by another AI before
Yeah it must have cost him a fortune cuz having enough resoures and hiring people for AI development isn't cheap. Fatheroftheyear
I'm pretty sure she's on a dedicated server, so unlike most ai's that are used by thousands of people at once she's used by only 1 person and is at least a few thousand times stronger
Lots of VLMs can do that, they're not as constrained as our human eyes. But it seems that Vedal has done some of those things: made sure that the quality of the image that Neuro sees has improved (compared to what it was during DBH), better interaction between Neuro and its vision model, or uses a better model for it. What is certain is that it seems to be faster.
Yeah, this is "just" the normal VLM stuff, and - I wonder what he changed? But it's similar tech than what allowed her to look at a screen before,.
@ Its not the image detection that I'm impressed with, those models are really good now as you said, they can decipher the most deep fried garbage images these days. It's more so the ability to pick out specific details. Normally they will produce a general description like "a chaotic beach scene, dozens of people are on the beach in blue and white clothing, the ocean is in the background with boats." etc but it can't list the color and position and every conceivable detail of every single object in frame. The impressive thing is that Vedal can ask for a specific detail about one of 200 objects in the image and Neuro is able to successfully query the detection software for the exact information she needs
looking forward to see what kind of crazy fan art neuro will react to in the future.
Neuro gets upgrades is always cool
Neuro is closer and closer to a person every day and I love this.
Smartest little cookie
Watch out, CAPTCHA. Neuro is coming for you.
Can't wait til neuro finally solves a captcha
I only now realised that Evil has a broken heart hairpin lmao
I'm actually excited for the subathon to end because I want vedal to cook her next set of upgrades
You can hear how proud Vedal is with this upgrade
I tired Chat GPT and he was similar, but he got the color of the boat wrong (he said "yellow sail and a white hull"), so Neuro was better than Chat GPT 4o on those.
That is twice she has kicked Chat gbt's ass now
This level of image interpretation is actually insanely impressive. Just a few years ago, google's image recognition model could barely interpret drawings of things
Neuro has been gifted the power of sight
Little did Vedal know, he is trying slowly to avoid the AM incident.
I couldn't see Waldo, until I heard Neuro's nonsense description of where he is. Then I somehow saw him instantly.
Damn. Image recognition on AI achieved!
This is legitimately impressive for an AI.
Where’s Waldo speed running.
this upgrade is huge
Holy shit! That's kind of scary that AI has come this far.
Professor vedal and his ai companion
This is honestly super impressive
Neuro was right about Waldo. There’s a lady carrying a closed red umbrella under the horses, and if you look south, like 2m, there’s Waldo sitting on a red chair.
This is actually insane.
that is honestly soo freaking cool!
Tfw the British turtle refers to the British character by his American version name
No Vedal! You fool! You just gave Neuro the ability to successfully read captchas
Wow vedal has found huggingface
But can she see why kids love cinnamon toast crunch?
Soon.
She can finally pass the KTaNE button module.
smartest little cookie
SMARTEST LITTLE COOKIE 🥰🥰 SMARTEST LITTLE COOOKIE 🥰🥰
The person the neuro was referring to "under the umbrella near the horses"
Has a shirt with white and red stripes.
So the description matches
Kind of heresy here but if this is her ability at the moment, how could she recognized details so good in the gaming setup rating long ago with Bao? Veadl must had given a brief description of the images right?
Probably some experience it had b4
Vision module might have been more external, and I think it was said that image descriptions were prepared beforehand.
ask her to list which squares contain a car.
We're all doomed when she can tell us where Waldo is. Next stage is finding Sarah Conner.
captchas beware.
Neuro: I don't know who waldo is
Vedal: You can google it
Neuro: googling where is waldo picture without waldo
Vedal: Wait what
Neuro: I can't find waldo
The chat saying geoguesser is smart as I wonder how good would be now
Lol if you place her in a robot rn, she can hunt humans already.
🍪 For Neuro
this is basic chatGPT 4o features. not something vedal created.
it can do a lot more like analyze graphs and a large array of data from a picture.
one time i showed it my factorio stats screen and it said i needed to increase coal production because it was lagging behind the current usage, and that i needed to add more power stations for further expansion. it was spot on. it also calculated how many solar panels and accumulators i needed to add to have enough power for the nights. its wild.
its only $20/ month for the image processing. its worth it.
Neuro is not an OpenAI model, she is most likely based on an open source llm and has separate vision model
I had no idea you coded chatGPT! So proud of you!
neuro's almost entirely made of open source stuff excluding her speech which is microsoft azure, vedal tried to use an open source voice synth for her but nobody liked it so he set that up for evil neuro instead. anyways he was talking about trying to change her voice again somepoint most likely to get away from closed source code.
pretty sure he's trying to distance himself from code used by corpos because they could claim they own part of neuro otherwise and there could be potential legal issues in the future, i'm kinda surprised microsoft hasn't done anything already.
The subathon… is still going?!
Vedal didn’t tell her who Waldo was or what he is wearing. So how is she supposed to find Waldo without that information.
Probably knowing what he wears and trying to look for that match. All she then needs to do is query the image AI for those details.
@ if that is in her data base or if she has access to do searches on the internet then that is fine but we don’t know that she has those things in this instance.
@IPlayVideoGamesAndNothingElse I expect nearly every LLM has some sort of base training data that probably has some sort of reference to it. It's also a very specific phrasing. Not what you usually get in everyday conversation, thus easier to solve. Plus, she also had an initial glimpse at it and said she didn't know. Then later on she did. So, there might have been a moment to get context.
Same way she knows what Disneyland is, pre-training data.
If you REALLY wanna train Neuro on something hard....Where's Waldo on NES on the Hard Setting. Shit is INSANE! It's actually got Randomization and he doesn't even look like his normal self. It's DIABOLICAL!
(for now)
I thought she already had this, can someone tell me what's different about this?
I think she is better at seeing colors and can see smaller details.
Well it seems Vedal is trying to showcase that she's much better at recognizing imperfect images now (blurry text, rough/disproportionate sketches). Iirc she struggled with properly recognizing fanart and similar stuff in the past.
I assume its better than the 240p images at best that she used to have
I think previously her vision most likely worked by having another AI describe the image and then send her a paragraph of text about it, which would lead to her only being able to see the details included in the description she received. When specifically asked about something like the color of the sailboat in the background of this chaotic image she probably wouldn't know because it's unlikely the AI would include such a small random detail in the description. Now it seems like she can specifically query for what information she wants to be able to see from the image?
@@soasertsus During the Detroit: Become Human playthrough Vedal showed that Neuro saw 240p screenshots with her "vision"
Where's waldo is a fun idea, though does Neuro know what Waldo looks like, or what anyone looks like for that matter? I guess she's seen people like Layna during cam streams, but does she know what say Anny's model looks like? Does she even know what Evil looks like? I'm curious how that works for her, since she only sees what vedal shows her, so she might only know people by voice.
I wonder if neuro is multimodal. She has to be, right?
oh.
Where is Vedal?
Didn't she have vision capabilities already?
dougdoug is never winning any geo guesser ever again
he just needs to upgrade his ai again, and tell it that it's not kyoto.
4:00 She only got the direction wrong. It's bottom right. But yes, Waldo was under the umbrella, near (under) the horses.
Does neuro even know who Waldo even is?
Very likely. For the same reason most image generators know who Hatsune Miku is: It's a distinct in name and appearance, character that shows up in training data a lot, compared to others.
Not born too early to never see AI
Not born too late to watch AI take over the world
Born at the perfect time to watch Anime girl slowly gain sentience
This isnt a new feature, shes done this for about a year now
Not to this detail. As for the fan art sections, it was clear that there is a secret description she reads.
Downvoted for camila content
@@AiOinc1 can you shut up? seriously, what is up with your hatred against her??
Just as long as she isnt the one making the images
cry about it
she made ascii art of a cat that was like a 3 year old's doodle once and that's about it