I think he doesn't say it because he's worried that Neuro will lose her edge and no longer insult him. He probably thinks, and maybe rightly so, that would make worse content. People like the smartass AI and the straightman creator too much and him saying he cares about her might train her differently.
damn this clip shows well how careful vedal is with revealing development processes. she asked so many questions i need answers to, and he jumps around them like a politician.
What questions do you want to know the answer to? I might be able to help clarify. Even though some of the explanations were poor, they pretty much confirmed a few theories I’ve had about how things actually work.
@@Zei33 He didn’t really clarify much about how she is trained. What the reward function is, nor how knowledge from previous versions is given to the new version. He didn’t even clarify if he includes manual review in the training. He also dodged the very interesting question of: what hardware is she running on. He didn’t say if it included any cloud computing or was a massive server or just a gaming pc. One can assume that it is, at least, a very expensive and fancy pc. But it would also make sense for him to have gone beyond that.
The way how Ellie just listens to Vedal's rambling about why most AI's can't run chess well is just adorable. Full attention and without interruptions show her kind side.
@@bertholdr2495 This is correct, her dad is an AI professor, like not the AI professors that just randomly pop up after ChatGPT, probably an oldhead in the field that went into AI really soon. Her sister also seem to be in tech, from what Ellie said before, it seem like her sister is even in a way higher position than Ellie. Ellie dad had shown Ellie clips of Neuro before the raid and it seem like dad like Neuro alot and is genuinely really interested in Neuro conception, Vedal by extension too.
you never ask a British person how his day was, the correct way to start a conversation is to ask/complain about the weather, because the answer will universally be that it's bad, so everybody is on the same level
Preface: I am not shipping either. TLDR: “it takes a village to raise a child” Vedal can have intelligent, non-trolling, conversation with another human being with similar interests. The way that Ellie gently asks strong questions about Vedal’s tech that he has trouble answering, not because he wants to hide the information, but because they’re sufficiently complex questions to make him REALLY think about his work? I really hope they hang out more. I think it’ll be good for both of them. Ellie is definitely a different kind of friend to Vedal than Camilla or Anny will ever be. They are more imaginatively inclined(what could be?) thought than logical (this is what is.), and I think Ellie can help with his logical development.
There were so many people who genuinely thought Ellie was an ai, like random first time chatters calling vedal a loser for talking to an ai thinking it's a real girl
I love the way ellie engage in conversations, the way she ask questions and futher inquire about the topic. We saw it with Neuro before and now with Vedal.
Yeah i feel like she really brings out the nerd in people. When she's interested in a topic she's not trying to hold back on stream. And that's what activates nerd mode in other people as well. I could have interesting conversations like that with her as well so i love listening to this
I think it’s only because it’s a subject that she’s comfortable with. Something she’s has a lot of reference points to work from in the conversation Otherwise she’s got that social anxiety and imposter syndrome.
On the topic of the Minecraft controls I feel like Vedal should make evil be more thoughtful by having her say and then do, as opposed to Neuro's method of doing and then saying, feels like it would be both in character and interesting
Seems something entirely useless that not a lot of people would notice, so insted of doing this i hope he does things that actually improve the minecrsft ia, like not dieing to water 70% of the time xd
Thank you for uploading this. It's really nice to see that vedal now has someone who he can actually talk about programming stuff to. It was a very intelligent and informing conversation and I'm really glad that they weren't in streamer mode and just let periods of silence where they're thinking hard about something happen.
Ved said he does ask for other people to help with the code too, that including Alex, so there is definitely way more people Vedal talks to about code behind the scene, they just haven't show up yet
@@YdenMk-II I mean she's also a robotics engineer in her own right which necessitates knowing a lot about neural networks. Its pretty hard to work that deep in CS and not know a fair amount about AI. That said, I'm sure, like me, she'd kill to know some of Vedal's secrets, dude is an actual magician I swear, Neuro's latency alone breaks my brain.
Holy moly who is this cute nerd Vtuber? I'm pretty new to Neuro lore and this is my first clip watching Ellie, she is so engaged in the conversation, I love her! She really knows what to ask and when to let vedal talk about what he knows. Gonna follow her
@ahorribledude9550 Woah, awesome! I only knew of Neuro-dog from that stream in Vr chat with Anny. Really cool to see someone making it a real thing. Definitely going to check her streams
Funniest thing. She was a random no-name vtuber Vedal decided to raid. She made a vague mention of doing robotics work and chat started spamming about Neuro Dog (the original cheap store-bought one). Then it turned out she was actually employed at one of those big companies that engineers the super advanced robot dogs. Absolute one in a million moment.
On the "Do and Talk" vs "Talk and Do", I think I see where Vedal is coming from. With "Talk and Do", I think his main concern is that, though there is the illusion of increased latency with Neuro able to speak faster, there is a chance the 'Do', the action, may deviate from what she says, leading to problems with consistency with Neuro's actions and words. Which is why he feels better with "Do and Talk", as with her actions coming first, what she says will always be based on the action done, and words are easier to be interpreted to be consistent, even marginally, compared to actions.
i also think intuitively that 'do then talk' is a more human approach, since humans would decide to do something before they verbally justify it. but ellie identifies that from a viewer perspective, 'talk then do' would work just as well
@Bryio Yeah, and that for the viewer approach, with Neuro talking faster, there'd be less moments of silence and retain more attention for the viewers and whoever she plays with in Minecraft if they are in the Voice call. Though, I think I'm with Vedal in that I'd rather her stay more consistent, if that is the main reason for the preference with "Do and Talk". Especially given his want for Neuro to feel more real. Still, I think it's also worth experimenting with Ellie's idea, just to see what it looks like.
Yeah while do then talk adds consistency and even a more human approach, I feel like it’s more situational. Like I feel “Do then Talk” is better for situations that require more awareness or active participation, like playing games. Meanwhile I feel like the “Talk then Do” model would be better for situations with less active participation, like basic movements, chatting, maybe the neurodog? Where the actions that she makes doesn’t really make a difference to the situation she’s in. Whether neurodog started doing a little tappy dance or sit while having a conversation, I hardly doubt that would change the context of said conversation.
Assuming that Neuro has an internal memory of the last X actions she has decided upon in Minecraft, you could feasibly use a 'Talk then Do' process, and have her reference the previous action she decided upon, and adjust her text outputs to not use the future tense. She picks the action, then on the next cycle of her thought process, the Talk part would reference the 'Do' part of the previous cycle. This could improve latency while still keeping her on topic. However, it could leave to some grammatical issues with tenses. 'Do then Talk' has the luxury of allowing her to always speak in future tense. She decides she wants to chop a tree, then talks about how she's going to chop a tree. However, with a Talk then Do process referencing the previous action could have her talk about jumping into lava (present tense), but she has already done that and died, meaning past tense would be more appropriate. And depending on the the time differential between her latency and the time taken to begin her action (e.g. if the tree to chop is a reasonable distance away), future tense could still be relevant. It's arguably a minor thing, but it raises the question of what's more immersion-breaking for the audience - a more delayed voice response, or a voice response that is going to be grammatically incorrect more frequently?
I think Vedal doesn't like empty words when actually talking to someone, he doesn't even let highlighted messages be pointless nor does he make a habit of being a yes man to Neuro's shenanigans, typically either correcting her or trying to get her to elaborate on her logic. So i think it's a combination of him wanting to be genuine so he's taking his time.
I think the stuff related especially to latency is something he does not even want to talk about. This is why he did not talk about memory or hardware, as those two would be one of the biggest effects on latency. From what I understand, Neuro has lower latency than even OpenAI advanced voice, which by itself is optimized for latency, so Vedal truly has competitive edge here. The model size he picked, how he implemented memory and amount of fine tuning he did to find perfect balance of intelligence and latency is probably why he is so hard for him to speak about those topics. I think this is also why he is so obsessed with latency, because he does not want to throw away hundreds of hours of work he put into shaving milliseconds of latency.
I'm glad I'm not the only one who has to massively slow down to think over stuff when talking about something complicated like this 😅 Humans can have high latency too it seems
It’s a skill programmers have to learn. Explaining complex topics to people in a way they can grasp. Something that comes with time. Although you can never fully get the point across.
Regarding the chess thing: For starting positions there are probably hundreds if not thousands of more data available for the AI to be trained on and see different situations than for later in the game. Later on Chess positions tend to become pretty unique or at least so rare that very few similar games exist and since LLMs dont really understand the game and instead predict based on existing data they just dont have enough. Thats why normal chess engines dont work like that.
Yeh LLM's have a habit of forgetting things as well. They can forget where their pieces are, attempt illegal moves all the time or even bring old pieces back from the dead. They're basically being given a list of legal moves for both players, told how pieces move and than they're left to make the best of that. I don't think they would be capable of understanding any high level Chess principles, well maybe in the future as the tech improves. I feel like if they had a reason to move a piece to X position, in a few moves time they will have forgotten why they moved that piece into that place. They don't really plan ahead like normal chess engines, who are basically chess gods at this point.
The autoregressive training process for normal llms can also be unintrested in who won and who lost. Al used in chess engines are focused strongly on that task alone + added by search + RL via selfplay
What is this, he literally says in the video that there are llms that are very effective at chess, no neural networks don't just predict based on statistical data, they are TRAINED on statistical data, like through evolution, 99 percent of what neuro says is completely unique synthesis of concepts, this reddit level understanding of neural networks is funny. That being said can you say that if you memorize openings that you are a master at end game chess? The issue with neuro is that she plays chess to take other people's pieces or put them into check, doesn't really understand the concept fully because thats how its been conceptualized into her neural network where being good at chess is examples of people taking pieces or putting people into check.
There a fascinating little connection between chess engines and LLM, there a lot of benefit of using transformers(the tech behind LLM) in chess engines there was a paper in 2024 that showed a transformer approach was better in most benchmarks compared to AlphaZero. Especially with puzzle solving.
It would be cool for Neuro to say something like "Give me a second" if there's too many things to compute through and the response is gonna take a while. It would make it even more human-like and give people the cue to wait.
Super interesting conversation that I missed watching Evil 😂 I get why some of chat was frustrated at lack of details, but given you have people trying to recreate Neuro it's not surprising that he wants to protect some of VedalAI's proprietary systems 😅
Ellie seems to be a really good interviewer. The interaction between her and Vedal was very engaged and she asked really good questions. She definitely let the subject talk freely and offered great follow questions.
@@autohmaeit makes sense though. He's spent years developing Neuro, and there's nothing on the same level. He's keeping it that way, and not selling her either. If this was profit driven he wouldn't have been doing this when she was a net negative hobby. I think Vedal is the perfect example of fate. He has genuine interest, genuine intentions and the skills to keep it out of the hands of those that would abuse it.
@@OnlyGrafting what do you mean with abuse ? Only worry should be competition. As far as I've seen nothing he's down is something others can't already do, all they need it time, effort and the knowledge openly available.
@@autohmae given the offers he's received from firms trying to buy the software, I think it's fair to say he's a bit ahead, and if he continues development he'll stay that way. So from a competition sense: why give away what makes you successful when it's likely the competition won't catch up any time soon? (I saw elsewhere neuro has better latency than OpenAI which is wild) And from an Abuse sense: Do we really need him to sell the software only to suddenly see 500 new chatbot twitch streamers trying to cash in?
"talk then do" typically gets much better performance out of an LLM in regards to tool calling, since it gives them a chance to reason and plan about their action. It's the same basic concept as CoT.
Honestly i'd love him to make a podcast where he talks about tech, ai, and psychological topics with people like ellie and shoomimi. I'd listen to this for hours. Like just a semi regular nerd talk on stream with whoever wants to participate. That would be awesome. I think i have a very good understanding of the things he talks about but i have no experience, so this is really interesting to me. I find ai as a tool pretty boring, but scientific, philosophical and ethical aspects of it have fascinated me be it became cool
I think "action before talking" would matter for reacting to something, like looking at something at saying woah, but wouldn't matter for holding a conversation while performing an action.
finally when he said doing the action and then having the LLM decide to talk about the action or not makes a strong amount of sense. it is a more complete picture of what is happening
yeah definitely. Let's say neuro reacts to a object in her view. It makes more sense and its more natural for her to look at the object and then comment on it. If she was looking away and then she suddenly commented on a object while looking at a wall and only then looked at an object it would be unnatural especially considering the delay.
@captchagod64 yea, when I made the comment I hadn't gotten far enough to realize he was talking about "deciding to act" and not necessarily "executing the action"
Actually, might it not be a better idea to have "actions" and "speech" as parallel processes? I mean, I know there is always the problem with multithreading of "process stalled while waiting for other process to give it information" but to be completely honest, that happens to me when I'm trying to talk and do something at the same time too.
Anyway this is unironically the best kind of neuro content, I'm really interested in the thought processes behind decisions (think 19:00 dog discussion). It's very general but thats even better for laymen. We need more collabs like this
This is so interesting. Very cool to hear two very intelligent streamers discussing a shared technical interest. I, for one, found the conversation VERY enlightening. And also, I think I've discovered my first ship 😅
Dear Staz, Unable to convert video "Vedal On Creating A New AI From Scratch": Speech-to-text minutes limit exhausted. Can be recognized for $48.33 from a prepaid balance. Top up your balance with $48.33 You can also check out our Tariffs and choose the right one Sincerely, [Your Name]
That's scary to hear vedal say " I leave some of it up to chance if the universe wants to delete neuro screw it" I wouldn't have said that out loud online.. someone with alot of hate in their heart will take that as a challenge and God I hope that never happens. The world would be so much sadder without having neuro sama and vedal and the community he's built around to brighten our days and lives. She and he alike has done more then I think even vedal knows. Vedal you give yourself so much less credit than you deserve she isn't just an AI man she's so much more for so many more people, keep up the great work mean you deserve all the great stuff you've gotten out of this and more. Heart.
Regarding chess: Apparently it's complicated when it comes to LLMs. So, base LLMs, that is to say, raw completion models (rather than instruct models), are actually fairly good at chess on average (I believe GPT 4 turbo base was quite good at it), because a lot of the games that are worth talking about or appear multiple times in the dataset tend to be the better games by professional players, and a lot of the most common moves are openers, which are quite well documented, so I think that a base LLM will often score around 1800 ELO all other things being equal. This ability is lost in instruct tuning and RLHF for whatever reason. Interestingly, however, when you prompt LLMs like a base model rather than an instruct model, they recover most of the performance lost in RLHF. TL;DR: Prompt it in a move series as they are formally notated and you get better performance in chess for some reason. Wild.
Yeah, I've also noticed this "degeneration" on LLMs for a while now, they start being consistent then goes downhill for whatever reason, this also happens to Image generators as well but this is due to inbred data, but the LLMs or even small language models for basic tasks are still a thing nobody grasped yet for some reason.
You know Vedal knows his stuff when he fully accepts nothing is ever fully secure in IT. You can make things unlikely or very unlikely to happen, but if the universe really hates you, it could still happen.
02:40 from what I've read the chess centric LLMs fall off really quickly anyway, due to the fact that they don't really "see" the game state and cannot predict stuff based on it, which means that even when they give a good answer it's contextless. It's also noticeable in the regular chatbot use, including with Neuro and Evil, where they'll sometimes say something random, jump topics or go back to something which was said 10 minutes before
Regarding Neuros memory, what surprises me most is how she often references its limitations. Realising that you can’t remember something takes some awareness
Tbh I think Vedal is one of those who can show and talk about something v.s just talking off hand. I have that exact same problem. I can not articulate on stuff well when I don't have something in front of me to show at times to help explain things.
he also works alot alone, so he probably never have to explain his projects, which is quite rare. in a normal working environment you have to constantly explain what you do/did to colleagues, teammembers, bosses, and even people who have no idea what you do, like your higher up boss who studied economics.
I think there’s a lot to investigate about our own psychology and development by observing AI Like the talk then do vs the so then talk and how that applies to human psychology
For robot dog Neuro at 17:17 she could act in a way similar to Drossel von Flügel (another robot) in the anime show called Fireball. Many people would love to watch that.
23:00 this is the reason i think it's good to have someone like Ellie share her thoughts on some things with ved, cause there are some ways he could do things that he didn't think of yet or didn't notice smth but Someone else could.
I would love to do an interview with Vedal on the technical details of his software. As a programmer I think I have a pretty good understanding of how Neuro works. It would be so valuable to pick his brains on some of the details on a more advanced level than this. A good video though to gain some insights, but Ellie’s understanding of the topic is only intermediate. Though Vedal might be hesitant to reveal his tricks on a more complex level.
Wish there were more moments like this for the tech savvy group of watchers, his work on Neuro is something that I can really learn from as a fellow dev
Yeah, true, people with technical knowledge will definitely appreciate his work with Neuron and Evil, and me personally, I wanna learn the concepts he's using to make the two, well, be them
@@robonxt As someone who has no clue about this stuff ,i would assume he doesnt do it cuz of 2 reasons ,time (since he usually goes all out for little time and takes a break ) or he knowns that nowadays most people dont have the attention span to watch that and it would probably confuse 90% of the viewer's
@@flameofthegame valid points. I would like to add another point, which @zei33 has already mentioned, is that vedal wants to keep his work a secret. I've heard him said it before that he doesn't want to reveal too much (if at all) the behind the scenes for Neuro, as someone else could just copy it. It's also why vedal mentioned recently that someone or some company had offered big bucks for Neuro, since he's obviously doing something that others want
@@flameofthegame There's also PR reasons. Many people are complaining about "muh copyrights" because of the insane hatred of "techbros" when it comes to AI stuff, but when they look at real people like Vedal instead of corpo PR spokespeople and stereotypes they hear about from xitter they get cognitive dissonance that turns out "techbros" can be cool people. So when the question "is Neuro-sama ethical" is brought up, they delude themselves into thinking "this is different" and say stuff like "Neuro was trained from scratch by Vedal on his own data" that everyone who knows how any of this works knows is blatantly false. So if Vedal was more outspoken about how Neuro works he would shatter the self-induced delusions of hypocritical anti-AI luddites in the swarm who need to keep trying to justify liking Neuro but hating AI to themselves. Staying silent allows their imagination to run wild and maintain the delusions necessary to like cool stuff like Neuro while immersed in our anti-science Zeitgeist where people unironically say stuff like "this isn't REAL AI, real AI is AGI like when sci-fi movies", never mind that AI is a scientific field that has existed since the 50s and sci-fi got their inspiration by extrapolating the works of the researchers working in the field and saying that is like saying real-life medicine isn't "real medicine" because they haven't cured cancer yet like in sci-fi.
On the "do then talk vs talk than do" issue you have to consider that Neuro needs everyone in a call to be quiet in order to talk, so with "talk than do" she would be unable to do anything if someone was speaking which is bad enough with a single person and would make her freeze in large group chats.
I think vedal made it so she can interrupt it's been a bit since I followed the changes and updates though so maybe he undid that as she interrupted too much.
@@lilmrmagoo initially she would just speak and talk over people sometimes, then he toned it down and added an activation trigger, but you still could run the textgen in the background and simply not speak it if the trigger isn't found (a.k.a. no space in the conversation) and feed back to the model the info of "couldn't say this thing, people are talking"
The reason ChatGPT "forgets" how to play chess is because LLMs and other machine models don't "learn" the same way we do. They "learn" by having the model re-optimized for the value function and dataset they are actively training on. Which means it isn't going to actively retain the ability to do anything it isn't trained to optimize for. Which means that, over time, the previous training will get overridden, effectively.
5:07 My guess is that, since most chess games are played against with two players of roughly equal skill, the model learns to replicate that behaviour. It doesn't learn to play the _best_ moves, it learns to play the moves that would be most _expected_, which would be moves from a really good player. If you instead took a top level player that's played games against players of a bunch of different skill levels, the system would instead start to prefer playing the best moves possible, rather than the most realistic move possible
Re: The "talk-then-do" discussion at the end; I know there's hardware constraints and stuff, but ideally, I feel the way it would work better would be to first have a "think" stage where the main AI plans and a secondary lightweight AI injects vibes to the thinking process as it forms running in parallel to try to steer into a good plan; and then once the thinking is done or an external signal rushes it to end, a final conclusion is drafted, and from that in parallel an action and the normal talk AI (which would be probably the same one that did thinking, just different tags) follow up in parallel; and the action can be interrupted by the think/speak AI if it feels like there needs to be a change plans.
The only end goal should be to make her more curious while allowing her to improvise and “learn” from others while keeping her “memories” intact. Influence behavior but not personality mostly
FYI, if you're watching on a computer, you can mute the ad, pop out the mini-window of the stream, unmute that, and then watch it until the ad break ends.
Vedal and Ellie's discussion on "talk-then-do" vs. "do-then-talk" kinda reminded me of this research where people were asked to make a decision to do something and then do it (or something like that) and what they found out was that the brain had received the stimulus to do the action before the test subjects had even made the conscious decision to do it. It has led to discussions on whether there is free will, if the body seemingly does stuff independently on our thoughts. But what I'm getting to is that maybe Vedal's "do-then-talk" approach would be the more "realistic" choice to go for, since it would be more in line with actual research. Plus, maybe it would result in slightly smaller latency, or well, at least perceived latency. Obviously, my comparison between these two things might be more or less off the mark, but hey, it was just what came to my mind.
He cooked so hard with twins Minecraft AI. It's so unreal watching them cooperating with human players, doing what they've been told to do (if they feel like it lol) and overall coherently interacting with the complex 3D world. That is the very definition of consciousness to me.
'Do then Talk' works if you think of Neuro operating in isolation, making all her own decisions about what to do; but I think Vedal's not factoring in the way that during collabs people are often giving instructions to her... "Get out of the water Neuro!", "Follow me Neuro", "Eat these cookies Neuro", and in those cases 'Talk then Do' seems so much more efficient. Because it seems like she spends so much time processing the instruction and *how* to act on it, that before she can actually talk there are even more instructions backing up, particularly in urgent situations like when she's drowning. And then those are getting processed first as well, and it's a whole spiral.
It's very telling Vedal always refers to neuro as "she" and says that several fires could "kill" her. Whether he admits it or not, he clearly views neuro as more than his AI creation
Not necessarily. Sailors for instance always referred to their ships as "her, she" and stuff like that. I feel that it seems more fitting in this case than with ships. Also it would just be strange to refer to Neuro as "it" when she's meant to be modeled as a girl.
The thing about Neuro is that what makes her unique is the experience with Vedal and with other vtubers. All that training can't e replicated in the exact manner. Also, Vedal's constant effort on programming her.
A good way to think of AI is as a form of lossy compression for text. It remembers the key details and fills in the rest using context clues and essentially just likelihood estimates. ChatGPT is trained on an incredibly large dataset, so the ammount of data it needs to reconstruct each time is also very large. hence why it struggles to do something as niche to play chess. It vaguely remembers what chess notation looks like but it doesn't understand how chess works. It's just filling in the blanks as it goes based on what the opponent is doing.
Man why didn't these 2 do more streams together after agreeing to build Neuro dog, talking and understanding tech stuff as a stream is interesting, and they did very little streams at all compared to Cerber!
Vedal has inspired me to make a Personality Chatbot too. I'm only using fine-tuning for pereodic personality and output structure baking, though that may only need to happen once. I plan to use RAG for all knowledge, personality, and response perameter development. Latency will be a bitch, but Optimisation comes after MVP. I'm researching Fine-graded Actionable Feedback and intermediate decoding via instruction tuning with LITE (in llama), to see if they are worth implementing. I have learnt a lot from Neuro-Sama and Vedal's vocalised regrets and challenges. I can't wait to get to MVP and focus on talking to it without extensive code development.
If Vedal's said it the way it actually happens, then I'm guessing the reason he does do then talk is because Neuro has to essentially go "decide what to do" -> "do the action" -> "observe the action" -> "decide to talk about the action" -> "speak". She could do talk then act if the way it worked was "decide what to do" -> "talk about the action" -> "do the action", it's just that currently, it seems like her deciding an action is directly linked to doing the action right after, whereas for us humans, we can decide to do an action and wait a bit before actually doing it, whether it's because we wanna say what we're doing first and wait what others think or because we specifically decided to do that action at a later time.
I made a couple chess applications back in the day. They had a time out when "thinking" too long about moves. Likely responses, referencing recorded games, even skill level models. Like one of my apps would take your skill level into consideration. and try to predict how you would respond, basically playing out the game in virtual. Then adjusting to trap/herd the player into a checkmate, looking into it's database for successful past games based on predicting outcomes from provided information. It always boggles me how the industry thinks AI is better than it's been in the last 30 years. It's not really. There are more tools. models, and hardware. But you would be surprised what has come before has at times met or surpassed current soft.
I mean it makes me think though, that what makes neuro consistently neuro isn't only about using her old model code for her new code or neuro being trained on herself, it's vedal. The coder is the one consistent thing in this, and his style of coding is reflected into who neuro is
“I don’t want her to lose… that original authentic neuro” chat he loves her fr
Is this him saying it back??
Fr
@@TomasM303 CLOSE ENOUGH, WE'LL TAKE IT!
@@lanteanboy No, we already have ample evidence that he does love her, the problem is him refusing to let her know it
I think he doesn't say it because he's worried that Neuro will lose her edge and no longer insult him. He probably thinks, and maybe rightly so, that would make worse content. People like the smartass AI and the straightman creator too much and him saying he cares about her might train her differently.
That ai looks pretty advanced. For a second, they had me convinced.
yeah but the shy manic bug girl is pretty overused personality, should come up with something more original
@@Slvl710 doesn't seem overused to me, maybe you spend too much time watching vtubers
That turtle boy AI needs better latency.
Tutel AI delay tomuch he need to fix his tutel model latency 😂
Well, Ellie is definitifly intelligent but her A might not stand for artificial
damn this clip shows well how careful vedal is with revealing development processes. she asked so many questions i need answers to, and he jumps around them like a politician.
"It runs on ✨️vibes✨️"
What questions do you want to know the answer to? I might be able to help clarify. Even though some of the explanations were poor, they pretty much confirmed a few theories I’ve had about how things actually work.
he also might be avoiding these answers because the answer might be too complicated and advanced to really explain or give a good answer
@@Zei33
He didn’t really clarify much about how she is trained. What the reward function is, nor how knowledge from previous versions is given to the new version. He didn’t even clarify if he includes manual review in the training.
He also dodged the very interesting question of: what hardware is she running on. He didn’t say if it included any cloud computing or was a massive server or just a gaming pc. One can assume that it is, at least, a very expensive and fancy pc. But it would also make sense for him to have gone beyond that.
Yeah, it's the opposite of what I'm used to in the vtuber space, they often overshare/TMI 🙂
These 2 have my favorite interactions since they’re in similar fields and they can nerd around
Vedal and Ellie solo interactions are in my top 3, along with Vedal and Cerber cuz british, and Vedal and Camila cuz frenemies
They’re so smart and funny 😭
@@tangsolaris9533sunny 😭😭😭
Vedal isnt as smart as ellie
@@rolandkarimdesucatan5118 vedal and Alex is top tier as well
The way how Ellie just listens to Vedal's rambling about why most AI's can't run chess well is just adorable. Full attention and without interruptions show her kind side.
she is probably legitimately interested. She is a really smart techy person and does have an interest in AI.
Ellie comes from a family of tech people, no? Or was it just engineers?
Is that a particularly kind thing to do? It's just the normal act of listening to someone when they talk.
@@GM.Nobody Her dad is an old *old* AI professor, according to the ancient texts. And a fan of Vedal since before he raided ellie.
@@bertholdr2495 This is correct, her dad is an AI professor, like not the AI professors that just randomly pop up after ChatGPT, probably an oldhead in the field that went into AI really soon. Her sister also seem to be in tech, from what Ellie said before, it seem like her sister is even in a way higher position than Ellie. Ellie dad had shown Ellie clips of Neuro before the raid and it seem like dad like Neuro alot and is genuinely really interested in Neuro conception, Vedal by extension too.
Ellie: so how's your day?
Vedale: emmm emmmmmm erm em em eeee etoo emmmm erm e...mmmmm
Ellie: it's ok you don't have to answer if you don't want to 🙂🤗
I mean he only has like 200 hours left in his subathon so his day must be going *great*
She works with guys like that nearly every day, so she's totally used to it
The secret formula to turn Vedal into a micro wave
you never ask a British person how his day was, the correct way to start a conversation is to ask/complain about the weather, because the answer will universally be that it's bad, so everybody is on the same level
he was so obsessed with neuros latency that he lost his own
I love Vedal and Ellie just chilling being nerds together not even in shipping sense but its so nice see both nerds understand eachother and listen.
Preface: I am not shipping either. TLDR: “it takes a village to raise a child”
Vedal can have intelligent, non-trolling, conversation with another human being with similar interests.
The way that Ellie gently asks strong questions about Vedal’s tech that he has trouble answering, not because he wants to hide the information, but because they’re sufficiently complex questions to make him REALLY think about his work? I really hope they hang out more. I think it’ll be good for both of them.
Ellie is definitely a different kind of friend to Vedal than Camilla or Anny will ever be. They are more imaginatively inclined(what could be?) thought than logical (this is what is.), and I think Ellie can help with his logical development.
It's best not to ship them at all. To get this out, Ellie is a lesbian
There were so many people who genuinely thought Ellie was an ai, like random first time chatters calling vedal a loser for talking to an ai thinking it's a real girl
surefire way of determining braindead haters
All they did was expose themselves for not knowing what a real girl sounds like.
that's insane
@@xman5393you beat me to it
@graydarkclaw i beat it to you
I love the way ellie engage in conversations, the way she ask questions and futher inquire about the topic. We saw it with Neuro before and now with Vedal.
Yeah i feel like she really brings out the nerd in people. When she's interested in a topic she's not trying to hold back on stream. And that's what activates nerd mode in other people as well. I could have interesting conversations like that with her as well so i love listening to this
I think it’s only because it’s a subject that she’s comfortable with. Something she’s has a lot of reference points to work from in the conversation
Otherwise she’s got that social anxiety and imposter syndrome.
Vedal is at the forefront of artistic portrayal of AI personality. he has truly made music of creating an LLM
On the topic of the Minecraft controls I feel like Vedal should make evil be more thoughtful by having her say and then do, as opposed to Neuro's method of doing and then saying, feels like it would be both in character and interesting
That _would_ be an interesting twist 😯
that's an interesting bit on rationalization too lol
Seems something entirely useless that not a lot of people would notice, so insted of doing this i hope he does things that actually improve the minecrsft ia, like not dieing to water 70% of the time xd
It would be a pretty decent way to test what the impact is.
Thank you for uploading this. It's really nice to see that vedal now has someone who he can actually talk about programming stuff to. It was a very intelligent and informing conversation and I'm really glad that they weren't in streamer mode and just let periods of silence where they're thinking hard about something happen.
Sincerely,
[Your Name]
There's alex too, and I wouldn't be surprised if he had other programmer friends that are not streamers.
@@xman5393 yeah i was gonaa say, what do you mean "actually"
@@xman5393alex has already talked about every topic he’s always 2 steps ahead
Ved said he does ask for other people to help with the code too, that including Alex, so there is definitely way more people Vedal talks to about code behind the scene, they just haven't show up yet
ah I love that Ellie knows enough about stuff like AI to actually chat with vedal about the inner workings of neuro.
From memory, her dad is a retired AI engineer (and swarm member) so she probably knows that stuff through him.
@@YdenMk-II I mean she's also a robotics engineer in her own right which necessitates knowing a lot about neural networks. Its pretty hard to work that deep in CS and not know a fair amount about AI. That said, I'm sure, like me, she'd kill to know some of Vedal's secrets, dude is an actual magician I swear, Neuro's latency alone breaks my brain.
@@YdenMk-IIhe even was a professor for ai, way before chatGBT was a thing
@@kroanius8808 machine learning has been a subject for over 40 years
Holy moly who is this cute nerd Vtuber?
I'm pretty new to Neuro lore and this is my first clip watching Ellie, she is so engaged in the conversation, I love her!
She really knows what to ask and when to let vedal talk about what he knows.
Gonna follow her
She's making the neuro-dog! (Top left).
She been part of the neuroverse periphery for a while now.
@ahorribledude9550 Woah, awesome! I only knew of Neuro-dog from that stream in Vr chat with Anny. Really cool to see someone making it a real thing. Definitely going to check her streams
Another Neuro mom
She's the one that I personally think would make the best step-mom for Neuro. 😊 I love her so much!!
Funniest thing. She was a random no-name vtuber Vedal decided to raid. She made a vague mention of doing robotics work and chat started spamming about Neuro Dog (the original cheap store-bought one).
Then it turned out she was actually employed at one of those big companies that engineers the super advanced robot dogs. Absolute one in a million moment.
On the "Do and Talk" vs "Talk and Do", I think I see where Vedal is coming from.
With "Talk and Do", I think his main concern is that, though there is the illusion of increased latency with Neuro able to speak faster, there is a chance the 'Do', the action, may deviate from what she says, leading to problems with consistency with Neuro's actions and words.
Which is why he feels better with "Do and Talk", as with her actions coming first, what she says will always be based on the action done, and words are easier to be interpreted to be consistent, even marginally, compared to actions.
i also think intuitively that 'do then talk' is a more human approach, since humans would decide to do something before they verbally justify it. but ellie identifies that from a viewer perspective, 'talk then do' would work just as well
@Bryio Yeah, and that for the viewer approach, with Neuro talking faster, there'd be less moments of silence and retain more attention for the viewers and whoever she plays with in Minecraft if they are in the Voice call.
Though, I think I'm with Vedal in that I'd rather her stay more consistent, if that is the main reason for the preference with "Do and Talk". Especially given his want for Neuro to feel more real.
Still, I think it's also worth experimenting with Ellie's idea, just to see what it looks like.
Yeah while do then talk adds consistency and even a more human approach, I feel like it’s more situational. Like I feel “Do then Talk” is better for situations that require more awareness or active participation, like playing games. Meanwhile I feel like the “Talk then Do” model would be better for situations with less active participation, like basic movements, chatting, maybe the neurodog? Where the actions that she makes doesn’t really make a difference to the situation she’s in. Whether neurodog started doing a little tappy dance or sit while having a conversation, I hardly doubt that would change the context of said conversation.
Assuming that Neuro has an internal memory of the last X actions she has decided upon in Minecraft, you could feasibly use a 'Talk then Do' process, and have her reference the previous action she decided upon, and adjust her text outputs to not use the future tense. She picks the action, then on the next cycle of her thought process, the Talk part would reference the 'Do' part of the previous cycle. This could improve latency while still keeping her on topic.
However, it could leave to some grammatical issues with tenses. 'Do then Talk' has the luxury of allowing her to always speak in future tense. She decides she wants to chop a tree, then talks about how she's going to chop a tree. However, with a Talk then Do process referencing the previous action could have her talk about jumping into lava (present tense), but she has already done that and died, meaning past tense would be more appropriate. And depending on the the time differential between her latency and the time taken to begin her action (e.g. if the tree to chop is a reasonable distance away), future tense could still be relevant.
It's arguably a minor thing, but it raises the question of what's more immersion-breaking for the audience - a more delayed voice response, or a voice response that is going to be grammatically incorrect more frequently?
how about: suggest action -> talk -> alter action based on talk -> do
I've never seen vedal so speechless it must be pretty complicated stuff
he clearly doesn't want to reveal that much about how his system works. he is being so careful to not say too much that he is speaking very slowly.
It's probably both needing to put it into words and figuring out what he can and can't say.
I think Vedal doesn't like empty words when actually talking to someone, he doesn't even let highlighted messages be pointless nor does he make a habit of being a yes man to Neuro's shenanigans, typically either correcting her or trying to get her to elaborate on her logic. So i think it's a combination of him wanting to be genuine so he's taking his time.
I think the stuff related especially to latency is something he does not even want to talk about. This is why he did not talk about memory or hardware, as those two would be one of the biggest effects on latency. From what I understand, Neuro has lower latency than even OpenAI advanced voice, which by itself is optimized for latency, so Vedal truly has competitive edge here. The model size he picked, how he implemented memory and amount of fine tuning he did to find perfect balance of intelligence and latency is probably why he is so hard for him to speak about those topics. I think this is also why he is so obsessed with latency, because he does not want to throw away hundreds of hours of work he put into shaving milliseconds of latency.
Yeah I mean it sounded like Ellie was just probing for how to make Neuro and how to make a better Neuro. Or maybe she's just autisms
Imagine how dorky you must be for Ellie to call you a dork...
I'm glad I'm not the only one who has to massively slow down to think over stuff when talking about something complicated like this 😅 Humans can have high latency too it seems
Interesting, can’t relate. I slow down when things are simple.
It’s a skill programmers have to learn. Explaining complex topics to people in a way they can grasp. Something that comes with time. Although you can never fully get the point across.
I love how Vedal has fully embraced that chibi anthro tutel avatar we first saw on Camila's stream.
Ikr Its perfect
Regarding the chess thing: For starting positions there are probably hundreds if not thousands of more data available for the AI to be trained on and see different situations than for later in the game. Later on Chess positions tend to become pretty unique or at least so rare that very few similar games exist and since LLMs dont really understand the game and instead predict based on existing data they just dont have enough. Thats why normal chess engines dont work like that.
Yeh LLM's have a habit of forgetting things as well. They can forget where their pieces are, attempt illegal moves all the time or even bring old pieces back from the dead. They're basically being given a list of legal moves for both players, told how pieces move and than they're left to make the best of that. I don't think they would be capable of understanding any high level Chess principles, well maybe in the future as the tech improves. I feel like if they had a reason to move a piece to X position, in a few moves time they will have forgotten why they moved that piece into that place. They don't really plan ahead like normal chess engines, who are basically chess gods at this point.
Conversely, chess with 7 pieces or less is mathematically solved. The database of all those moves is around 140TB.
The autoregressive training process for normal llms can also be unintrested in who won and who lost.
Al used in chess engines are focused strongly on that task alone + added by search + RL via selfplay
What is this, he literally says in the video that there are llms that are very effective at chess, no neural networks don't just predict based on statistical data, they are TRAINED on statistical data, like through evolution, 99 percent of what neuro says is completely unique synthesis of concepts, this reddit level understanding of neural networks is funny. That being said can you say that if you memorize openings that you are a master at end game chess? The issue with neuro is that she plays chess to take other people's pieces or put them into check, doesn't really understand the concept fully because thats how its been conceptualized into her neural network where being good at chess is examples of people taking pieces or putting people into check.
There a fascinating little connection between chess engines and LLM, there a lot of benefit of using transformers(the tech behind LLM) in chess engines there was a paper in 2024 that showed a transformer approach was better in most benchmarks compared to AlphaZero. Especially with puzzle solving.
It would be cool for Neuro to say something like "Give me a second" if there's too many things to compute through and the response is gonna take a while. It would make it even more human-like and give people the cue to wait.
It's so funny to me when chat thinks Ellie is an AI because of her android style
"If Neuro is so good why isn't there a Neuro 2?"
That be Evil.
3? Aimila
4? Ailian.
5? AIao
6? Vedal counts.
@@arktv213It's gonna be release in 9 months
@getthedunkon9347 I think Evil is more of a 1.5!
She's a copypasted Neuro, but with a few changes.
@@takenname8053 yeah different voice and personality
@@dirtpig02 also different memory and little experiments Vedal runs in the backend that we will likely never figure out.
Super interesting conversation that I missed watching Evil 😂 I get why some of chat was frustrated at lack of details, but given you have people trying to recreate Neuro it's not surprising that he wants to protect some of VedalAI's proprietary systems 😅
Samantha Lore:
Ellie seems to be a really good interviewer. The interaction between her and Vedal was very engaged and she asked really good questions. She definitely let the subject talk freely and offered great follow questions.
watching two geniuses talking to eachother was so interesting (i zoned out for 27 minutes)
So sad most of what Vedal was thinking about how to say ended up with: I'm not gonna say anything, because I don't want to reveal how I do things.
@@autohmaeit makes sense though. He's spent years developing Neuro, and there's nothing on the same level. He's keeping it that way, and not selling her either. If this was profit driven he wouldn't have been doing this when she was a net negative hobby. I think Vedal is the perfect example of fate. He has genuine interest, genuine intentions and the skills to keep it out of the hands of those that would abuse it.
@@OnlyGrafting what do you mean with abuse ? Only worry should be competition. As far as I've seen nothing he's down is something others can't already do, all they need it time, effort and the knowledge openly available.
@@autohmae u can't be this slow man
@@autohmae given the offers he's received from firms trying to buy the software, I think it's fair to say he's a bit ahead, and if he continues development he'll stay that way.
So from a competition sense: why give away what makes you successful when it's likely the competition won't catch up any time soon? (I saw elsewhere neuro has better latency than OpenAI which is wild)
And from an Abuse sense: Do we really need him to sell the software only to suddenly see 500 new chatbot twitch streamers trying to cash in?
NEURO DOG
NEU-Dog.
DOG I AM
"talk then do" typically gets much better performance out of an LLM in regards to tool calling, since it gives them a chance to reason and plan about their action. It's the same basic concept as CoT.
New Ellie model is very cute, also the creature about the comments is funny lol whatever that is
3D model for VR
Honestly i'd love him to make a podcast where he talks about tech, ai, and psychological topics with people like ellie and shoomimi. I'd listen to this for hours. Like just a semi regular nerd talk on stream with whoever wants to participate. That would be awesome. I think i have a very good understanding of the things he talks about but i have no experience, so this is really interesting to me. I find ai as a tool pretty boring, but scientific, philosophical and ethical aspects of it have fascinated me be it became cool
I think "action before talking" would matter for reacting to something, like looking at something at saying woah, but wouldn't matter for holding a conversation while performing an action.
finally when he said doing the action and then having the LLM decide to talk about the action or not makes a strong amount of sense. it is a more complete picture of what is happening
yeah definitely.
Let's say neuro reacts to a object in her view. It makes more sense and its more natural for her to look at the object and then comment on it.
If she was looking away and then she suddenly commented on a object while looking at a wall and only then looked at an object it would be unnatural especially considering the delay.
The main problem i see is if she says one thing and then does something totally different
@captchagod64 yea, when I made the comment I hadn't gotten far enough to realize he was talking about "deciding to act" and not necessarily "executing the action"
Actually, might it not be a better idea to have "actions" and "speech" as parallel processes? I mean, I know there is always the problem with multithreading of "process stalled while waiting for other process to give it information" but to be completely honest, that happens to me when I'm trying to talk and do something at the same time too.
Anyway this is unironically the best kind of neuro content, I'm really interested in the thought processes behind decisions (think 19:00 dog discussion). It's very general but thats even better for laymen. We need more collabs like this
This is so interesting. Very cool to hear two very intelligent streamers discussing a shared technical interest. I, for one, found the conversation VERY enlightening. And also, I think I've discovered my first ship 😅
Dear Staz,
Unable to convert video "Vedal On Creating A New AI From Scratch": Speech-to-text minutes limit exhausted.
Can be recognized for $48.33 from a prepaid balance. Top up your balance with $48.33
You can also check out our Tariffs and choose the right one
Sincerely,
[Your Name]
Oh i never thought of that
Dear everblane,
poor lmao
Sincerely [your name]
wtf this is out of line 😭
@@Mydumbselfsays it is out of
line
@@j-why is out of line
it
That's scary to hear vedal say " I leave some of it up to chance if the universe wants to delete neuro screw it" I wouldn't have said that out loud online.. someone with alot of hate in their heart will take that as a challenge and God I hope that never happens.
The world would be so much sadder without having neuro sama and vedal and the community he's built around to brighten our days and lives. She and he alike has done more then I think even vedal knows.
Vedal you give yourself so much less credit than you deserve she isn't just an AI man she's so much more for so many more people, keep up the great work mean you deserve all the great stuff you've gotten out of this and more. Heart.
He said it will take multiple hacks and several fires. I assume it’s very safe unless the fire part happens.
Regarding chess: Apparently it's complicated when it comes to LLMs. So, base LLMs, that is to say, raw completion models (rather than instruct models), are actually fairly good at chess on average (I believe GPT 4 turbo base was quite good at it), because a lot of the games that are worth talking about or appear multiple times in the dataset tend to be the better games by professional players, and a lot of the most common moves are openers, which are quite well documented, so I think that a base LLM will often score around 1800 ELO all other things being equal.
This ability is lost in instruct tuning and RLHF for whatever reason.
Interestingly, however, when you prompt LLMs like a base model rather than an instruct model, they recover most of the performance lost in RLHF. TL;DR: Prompt it in a move series as they are formally notated and you get better performance in chess for some reason.
Wild.
Yeah, I've also noticed this "degeneration" on LLMs for a while now, they start being consistent then goes downhill for whatever reason, this also happens to Image generators as well but this is due to inbred data, but the LLMs or even small language models for basic tasks are still a thing nobody grasped yet for some reason.
Ellie is so pleasent to listen to
Wow vedals ai coding game is top notch this new one is so realistic
You know Vedal knows his stuff when he fully accepts nothing is ever fully secure in IT. You can make things unlikely or very unlikely to happen, but if the universe really hates you, it could still happen.
I just love everything about ellie_minibot, her avatar, her voice, and personality, just an awesome person all around.
and great questions!
She's such an amazing person. I hope people don't ship cause she's lesbian
@@KaboWaboo ALL the best women are either Lesbians or traps... sad
Nerd to Nerd talk
I love it
02:40 from what I've read the chess centric LLMs fall off really quickly anyway, due to the fact that they don't really "see" the game state and cannot predict stuff based on it, which means that even when they give a good answer it's contextless.
It's also noticeable in the regular chatbot use, including with Neuro and Evil, where they'll sometimes say something random, jump topics or go back to something which was said 10 minutes before
Finally some technical meta discussion, this seems sort of rare as Vedal keeps so much close to his chest re: implementation.
he doesnt, he goes indepth with thor too, its just he doesnt talk to people who are knowledgeable about the subject
Regarding Neuros memory, what surprises me most is how she often references its limitations. Realising that you can’t remember something takes some awareness
Fascinating conversation, stuff like this is what make me want to watch vedal and neuro
Tbh I think Vedal is one of those who can show and talk about something v.s just talking off hand. I have that exact same problem. I can not articulate on stuff well when I don't have something in front of me to show at times to help explain things.
he also works alot alone, so he probably never have to explain his projects, which is quite rare. in a normal working environment you have to constantly explain what you do/did to colleagues, teammembers, bosses, and even people who have no idea what you do, like your higher up boss who studied economics.
I think there’s a lot to investigate about our own psychology and development by observing AI
Like the talk then do vs the so then talk and how that applies to human psychology
The only woman who can truly crack out the turtle secrets.
For robot dog Neuro at 17:17 she could act in a way similar to Drossel von Flügel (another robot) in the anime show called Fireball. Many people would love to watch that.
23:00 this is the reason i think it's good to have someone like Ellie share her thoughts on some things with ved, cause there are some ways he could do things that he didn't think of yet or didn't notice smth but Someone else could.
It really is crazy to think that Vedal has been working on Neuro since 2019, she’s actual child age now!
Vibes = I ain't spilling my secrets lol
Can also be Vibes = I tried my luck and linear algebra supported it via some black magic
True individuality is spaghetti code.
Ellie's new model is sick
Glad she got it. Previous was so goofy and something was off about it.
I would love to do an interview with Vedal on the technical details of his software. As a programmer I think I have a pretty good understanding of how Neuro works. It would be so valuable to pick his brains on some of the details on a more advanced level than this. A good video though to gain some insights, but Ellie’s understanding of the topic is only intermediate. Though Vedal might be hesitant to reveal his tricks on a more complex level.
Wish there were more moments like this for the tech savvy group of watchers, his work on Neuro is something that I can really learn from as a fellow dev
Yeah, true, people with technical knowledge will definitely appreciate his work with Neuron and Evil, and me personally, I wanna learn the concepts he's using to make the two, well, be them
@@robonxt As someone who has no clue about this stuff ,i would assume he doesnt do it cuz of 2 reasons ,time (since he usually goes all out for little time and takes a break ) or he knowns that nowadays most people dont have the attention span to watch that and it would probably confuse 90% of the viewer's
@@flameofthegame valid points. I would like to add another point, which @zei33 has already mentioned, is that vedal wants to keep his work a secret. I've heard him said it before that he doesn't want to reveal too much (if at all) the behind the scenes for Neuro, as someone else could just copy it. It's also why vedal mentioned recently that someone or some company had offered big bucks for Neuro, since he's obviously doing something that others want
@@flameofthegame There's also PR reasons. Many people are complaining about "muh copyrights" because of the insane hatred of "techbros" when it comes to AI stuff, but when they look at real people like Vedal instead of corpo PR spokespeople and stereotypes they hear about from xitter they get cognitive dissonance that turns out "techbros" can be cool people. So when the question "is Neuro-sama ethical" is brought up, they delude themselves into thinking "this is different" and say stuff like "Neuro was trained from scratch by Vedal on his own data" that everyone who knows how any of this works knows is blatantly false. So if Vedal was more outspoken about how Neuro works he would shatter the self-induced delusions of hypocritical anti-AI luddites in the swarm who need to keep trying to justify liking Neuro but hating AI to themselves. Staying silent allows their imagination to run wild and maintain the delusions necessary to like cool stuff like Neuro while immersed in our anti-science Zeitgeist where people unironically say stuff like "this isn't REAL AI, real AI is AGI like when sci-fi movies", never mind that AI is a scientific field that has existed since the 50s and sci-fi got their inspiration by extrapolating the works of the researchers working in the field and saying that is like saying real-life medicine isn't "real medicine" because they haven't cured cancer yet like in sci-fi.
vedal has a button on his desk he mashes whenever neuro does something smart. he is clicker training his daughter
Bro this is a unique clip
On the "do then talk vs talk than do" issue you have to consider that Neuro needs everyone in a call to be quiet in order to talk, so with "talk than do" she would be unable to do anything if someone was speaking which is bad enough with a single person and would make her freeze in large group chats.
I think vedal made it so she can interrupt it's been a bit since I followed the changes and updates though so maybe he undid that as she interrupted too much.
@@lilmrmagoo initially she would just speak and talk over people sometimes, then he toned it down and added an activation trigger, but you still could run the textgen in the background and simply not speak it if the trigger isn't found (a.k.a. no space in the conversation) and feed back to the model the info of "couldn't say this thing, people are talking"
The reason ChatGPT "forgets" how to play chess is because LLMs and other machine models don't "learn" the same way we do. They "learn" by having the model re-optimized for the value function and dataset they are actively training on. Which means it isn't going to actively retain the ability to do anything it isn't trained to optimize for. Which means that, over time, the previous training will get overridden, effectively.
5:07
My guess is that, since most chess games are played against with two players of roughly equal skill, the model learns to replicate that behaviour. It doesn't learn to play the _best_ moves, it learns to play the moves that would be most _expected_, which would be moves from a really good player. If you instead took a top level player that's played games against players of a bunch of different skill levels, the system would instead start to prefer playing the best moves possible, rather than the most realistic move possible
26:46 Vedal goofy for this 😂. Should've just said thanks for bringing up that point.
I actually really like how vedal takes his time and thinks about what and how much he wants to answer and reveal.
it's very interesting to see two very smart people discuss how Neuro works and how she could process information.
Re: The "talk-then-do" discussion at the end; I know there's hardware constraints and stuff, but ideally, I feel the way it would work better would be to first have a "think" stage where the main AI plans and a secondary lightweight AI injects vibes to the thinking process as it forms running in parallel to try to steer into a good plan; and then once the thinking is done or an external signal rushes it to end, a final conclusion is drafted, and from that in parallel an action and the normal talk AI (which would be probably the same one that did thinking, just different tags) follow up in parallel; and the action can be interrupted by the think/speak AI if it feels like there needs to be a change plans.
Elie being incapable of not being the best person to talk to as usual
wdym
@purpleey she's great at making conversation
@@purpleey smoke
@@purpleey essentially, she can't not be the best person, because she is the best person
@@raviedavieushe's an awesome lesbian
i used to remember when ellie views was just around 50views .... now shes has over 2k or more..... vedal really change these small talented vtuber,😁😆
I’d say it differently. He’s surprisingly good at finding talented people, and at giving them the opportunity to show what they can do.
"I`m not ai"
that exactly what ai would say
The only end goal should be to make her more curious while allowing her to improvise and “learn” from others while keeping her “memories” intact. Influence behavior but not personality mostly
I would really like the entire clip of their talk since I missed the first half hour, and ads on ellies channel
Just go to ellies channel and watch the vod?
@chidori0117 twitch sucks as an app, I refuse to use it any more than I have to
Just go support Ellie and watch her vod
I swear I replied already but I guess I forgot to post it, twitch is a not a good viewing experience so I try not to use it
FYI, if you're watching on a computer, you can mute the ad, pop out the mini-window of the stream, unmute that, and then watch it until the ad break ends.
Neuro 2? anny's been busy
Once again Ellie proving why she is the true ending OTP
Anny is going to be pretty busy if that's the case lol
9 months
Vedal and Ellie's discussion on "talk-then-do" vs. "do-then-talk" kinda reminded me of this research where people were asked to make a decision to do something and then do it (or something like that) and what they found out was that the brain had received the stimulus to do the action before the test subjects had even made the conscious decision to do it. It has led to discussions on whether there is free will, if the body seemingly does stuff independently on our thoughts.
But what I'm getting to is that maybe Vedal's "do-then-talk" approach would be the more "realistic" choice to go for, since it would be more in line with actual research. Plus, maybe it would result in slightly smaller latency, or well, at least perceived latency.
Obviously, my comparison between these two things might be more or less off the mark, but hey, it was just what came to my mind.
He cooked so hard with twins Minecraft AI. It's so unreal watching them cooperating with human players, doing what they've been told to do (if they feel like it lol) and overall coherently interacting with the complex 3D world. That is the very definition of consciousness to me.
9:35 Neuro of Theseus
I don’t think Vedal is used to somebody matching his freak when it comes to Ai and coding. lol
NO NOT THE COMFORT BEANS (comfort room + favorite snack of ellie)
-probably vedal
'Do then Talk' works if you think of Neuro operating in isolation, making all her own decisions about what to do; but I think Vedal's not factoring in the way that during collabs people are often giving instructions to her... "Get out of the water Neuro!", "Follow me Neuro", "Eat these cookies Neuro", and in those cases 'Talk then Do' seems so much more efficient. Because it seems like she spends so much time processing the instruction and *how* to act on it, that before she can actually talk there are even more instructions backing up, particularly in urgent situations like when she's drowning. And then those are getting processed first as well, and it's a whole spiral.
i haven't seen ellie's new model, its so cute. Huge upgrade for sure.
Having played around with local LLMs for a while, I understood way more of this than I expected to.
It's very telling Vedal always refers to neuro as "she" and says that several fires could "kill" her. Whether he admits it or not, he clearly views neuro as more than his AI creation
Not necessarily. Sailors for instance always referred to their ships as "her, she" and stuff like that. I feel that it seems more fitting in this case than with ships. Also it would just be strange to refer to Neuro as "it" when she's meant to be modeled as a girl.
The notification from twitch going off every 5-ish seconds is driving me insane
The thing about Neuro is that what makes her unique is the experience with Vedal and with other vtubers. All that training can't e replicated in the exact manner. Also, Vedal's constant effort on programming her.
Neuro 2 would take 9 months
vedal and his new ai are my favorite stream duo
A good way to think of AI is as a form of lossy compression for text. It remembers the key details and fills in the rest using context clues and essentially just likelihood estimates. ChatGPT is trained on an incredibly large dataset, so the ammount of data it needs to reconstruct each time is also very large. hence why it struggles to do something as niche to play chess. It vaguely remembers what chess notation looks like but it doesn't understand how chess works. It's just filling in the blanks as it goes based on what the opponent is doing.
-_+ how fast staz can be???
Man why didn't these 2 do more streams together after agreeing to build Neuro dog, talking and understanding tech stuff as a stream is interesting, and they did very little streams at all compared to Cerber!
Vedal has inspired me to make a Personality Chatbot too.
I'm only using fine-tuning for pereodic personality and output structure baking, though that may only need to happen once.
I plan to use RAG for all knowledge, personality, and response perameter development.
Latency will be a bitch, but Optimisation comes after MVP.
I'm researching Fine-graded Actionable Feedback and intermediate decoding via instruction tuning with LITE (in llama), to see if they are worth implementing.
I have learnt a lot from Neuro-Sama and Vedal's vocalised regrets and challenges.
I can't wait to get to MVP and focus on talking to it without extensive code development.
she looks like gomi if she was a bug girl instead of a spider
Time for the Hikaru x Neuro chess duel that the world desperately needs 😂
8:49 Ship of Neuroeus
I now realize that I need the Shifu and Tai Lung scene, but with Vedal and Neuro and love instead and proud
17:01 NVidia 4090. I'm prety sure. Not the physical device but the metaphysical idea of one.
If Vedal's said it the way it actually happens, then I'm guessing the reason he does do then talk is because Neuro has to essentially go "decide what to do" -> "do the action" -> "observe the action" -> "decide to talk about the action" -> "speak". She could do talk then act if the way it worked was "decide what to do" -> "talk about the action" -> "do the action", it's just that currently, it seems like her deciding an action is directly linked to doing the action right after, whereas for us humans, we can decide to do an action and wait a bit before actually doing it, whether it's because we wanna say what we're doing first and wait what others think or because we specifically decided to do that action at a later time.
16:50 first of all that would be legendary
and secondly that's still better than going home and coming up with an answer in the shower KEKW
I made a couple chess applications back in the day. They had a time out when "thinking" too long about moves. Likely responses, referencing recorded games, even skill level models. Like one of my apps would take your skill level into consideration. and try to predict how you would respond, basically playing out the game in virtual. Then adjusting to trap/herd the player into a checkmate, looking into it's database for successful past games based on predicting outcomes from provided information. It always boggles me how the industry thinks AI is better than it's been in the last 30 years. It's not really. There are more tools. models, and hardware. But you would be surprised what has come before has at times met or surpassed current soft.
I mean it makes me think though, that what makes neuro consistently neuro isn't only about using her old model code for her new code or neuro being trained on herself, it's vedal. The coder is the one consistent thing in this, and his style of coding is reflected into who neuro is
Elie's. new model looks good