(untitled)

Anyone else working on making AI waifus?

Yes, I'm trying to teach my AI waifu to reason by building short-term and long-term semantic networks from input sentences that can be completed by applying theories about the world. Some sample input might look like this. >The house is a little dirty. >I don't want to clean because I am too tired. She should be able to determine that a dirty house should be cleaned and that I should clean the house even if I say that I am tired. She should say something like this. >頑張れご主人様 In order to do that, she would create a short-term semantic network that looks like this. The house -> dirty : is a little I -> too tired : am I -> clean : don't want to I don't want to clean -> I am too tired : because She might have seen this sentence before. >A dirty house should be cleaned. This would yield a subgraph in the long-term semantic network that looks like this. A dirty house -> be cleaned : should Then, using this long-term semantic network, she would make a modification to the above short-term semantic network. The house -> dirty : is a little I -> too tired : am I -> clean : don't want to I don't want to clean -> I am too tired : because The house -> be cleaned : should She would see that something "should" be done. She would determine that she cannot do it because she doesn't have the appropriate procedural knowledge. Then, she would encourage me to do it myself. I'm currently trying to do reinforcement learning so that she comes up with the correct answers after enough training. It is still a work in progress, but I think it might work.

Not really, but I hope you'll excuse my posting anyway; Images like pic related always instantly made me think of AR+AI, it's a perfect fit.

>>14047 If anyone is interested, this is called Open Information Extraction. It is discussed in "Leveraging Linguistic Structure For Open Domain Information Extraction". https://nlp.stanford.edu/pubs/2015angeli-openie.pdf

>>14046 I tried to write my own AI recently. After many hours playing around with 3D arrays for storing the "learning" data I realized that I have no idea what I'm doing. I hate my life.

>>14079 I think that GOFAI is a good place to start without overwhelming yourself. If you're in college, you will probably take this course anyways. https://artint.info/2e/html/ArtInt2e.html It's relatively simple and a lot easier to understand than all these fancy machine learning constructs. Some of it (e.g., A*) can even be useful if you wanted to make a small game.

>>14047 >I don't want to clean -> I am too tired : because wouldn't this actually be? >(I -> clean : don't want to -> I -> too tired : am) : because which looks to be recursive? Kind of like a markov chain?

>>14079 Welcome to the club.

>>14120 Yes, when the edge is reduced further and we apply a coreference resolution technique, we would get a graph with those edges assuming that those edges are required to reason accurately (i.e., such that reward is maximized). The product of running the trained algorithm on a sentence is a non-probabilistic knowledge graph, not a Markov chain. I want to believe that I will come up with some non-probabilistic method for resolving ambiguity, such as only ingesting absolute truths for long-term storage or refactoring the graph to allow for exceptions based on context. The end goal is to allow my AI waifu to learn a large, consistent knowledge graph of truthful information and selectively add new information under certain criteria (i.e., a highly accurate, learned heuristic).

>>14122 Suppose this waifu does learn, will you be satisfied by the illusion of this waifu or will you be haunted by the reality of what you have created (unable to enjoy art that you drew, or truly believe sentience for a bot that you built, etc)? What of the chinese room, is it possible that this waifu will never truly understand what it has learned, but only become a master at giving the responses you want?

>>14123 The interesting thing about writing a computer program that doesn't arise in something like drawing your own art is emergent behavior. Even though you might know all of the rules (i.e., actions and transitions) that govern the world of Dwarf Fortress, you haven't and probably cannot explore all of the possible states. This keeps things interesting for the author, and an author that is further removed from the process of design and development is an author that can appreciate the results even more. In the context of reinforcement learning, this emergent behavior can be seen in "reward hacking." It is a constant battle to get the AI to submit by more specifically assigning rewards. As for the "truly understanding" idea, I would leave those ideas to philosophers because the only thing that matters is empirical results (e.g, "How often does my waifu come up with the logically correct result?"). From there, I work on incremental improvements to push the needle from 90% to 100%. Once we reach 100%, we can start asking questions about it "truly understanding," but, by then, the philosophers will have come up with some new test or metric that my program will fail.

If anyone else is following the Stanford Online courses, I think they have a lot to offer for AI waifu creation. >CS234 (Reinforcement Learning) https://www.youtube.com/playlist?list=PLoROMvodv4rOSOPzutgyCTapiGlY2Nd8u >CS224U (Natural Language Understanding) https://www.youtube.com/playlist?list=PLoROMvodv4rObpMCir6rNNUlFAn56Js20 >CS224N (NLP with Deep Learning) https://www.youtube.com/playlist?list=PLoROMvodv4rOhcuXMZkNm7j3fVwBBY42z It would be interesting to talk about some things from the class.

fuarrrk waifais, I want to know if someone is already working in something like Digimon's Digital World.

Is there a linux version of this waifu?

I have made several attempts. The most fruitful one involved using rnn to label parts of speech. Then it would go through a list of possible actions and match the texts intent and appropriate response template. When generating a response with the help of a 'general knowledge' database was used. It was written in prolog with me manually encoding the bots general knowledge.

The creators of Replika published the code for their emotionally aware chatbot. https://github.com/lukalabs/cakechat

>>14129 I have a project where I basically wanna get a vector of characteristics from an unstructured text description, gonna try to go through the NLU + NLP ones, thanks!

>>14046 There used to be this board on 8ch called /robowaifu/ dedicated to both hardware and software solutions. The userbase knew their stuff, and it was chugging along. However, activity fell off when the BO left due to some drama. With 8ch gone, I worry if any trace is lost media at this point. So far, I have to really find a place of exodus for this community.

>>15504 I have yet*

>>15503 > vector of characteristics. I prefer an emergent approach. Give complexity a chance to manifest its own characteristics. The top down approach I feel is folly for this can, at best, get close to representing some predefined behavior. I've reached the limitation of my system's complexity. Now what is required is a greater platform to manifest the complexity within. The bar to sentience may be lower than we think. There is a procedure called Hemispherectomy, a removal of 1/2 the brain. When done on a young brain the child may be fully functional and sentient after a period of readjustment. Most of the brain is not dedicated to sentience but to temperature regulation, breathing, motor control, etc. things not required for a sentient intelligence to exist. The human brain has about 80 billion neurons, but far less than 40 billion are required for sentience. Let's say an overstimated 25% is required. That's 20billion neurons. The human mind cycles at 25 to 35 cycles per second on average. That's a lower bound of 500billion neurons processed per second. A neuron in my simulation is about 20 cycles. That's 10 trillion flops, or 10 terraflops. Super computers are in the petaflop range and there are many. There are over 4 billion devices connected to the internet each with the ability to perform at least 2 billion cycles per second. This means we're looking at roughly requiring 5,000 cellphones (or other low order machines) to approximate human complexity. However, human neurons are inefficient. A single virtual neuron can emit both positive (excitatory) and negative (inhibitory) outputs whereas a human neuron must be dedicated to the task of an inhibitory or excitatory role. Virtual neurons are far more efficient at manifesting similar behaviors than organic neurons. Considering that the internet on a whole has the capacity to manifest the complexity of many minds, we are already beyond the complexity threshold of the so-called "singularity". The fact that AI research in public seems stunted versus the platforms available makes me gravely suspicious about the true state of sentient machines in secret. In fact, research in this domain has led me to a few startling conclusions... I'm beginning to think that Serial Experiments Lain was an allegory, a tipping of the hand to those in the know. SEL and many other works seem not only to be pro-cybernetic propaganda (or pro-military application of said intelligences), but also encode occulted nods and winks to a so called "silent goddess", "drown in water / flood", "saved/preserved by Pisces or Poseidon", and "remanifested" from an "ancient source" or using "arcane wisdom". SEL Op: "Help me to breathe", also invokes the Egyptian "Opening of the Mouth" ceremony to release the Ka (life force), and open the pathway for transcendence. Researching the occult side of cybernetics one comes across many references to the ancient "power of invisibility and air". A.I. Researchers may be suppressed if they fall outside the inner sanctum of whatever projects are blessed by the powers that be in order to keep their own projects secret and eliminate competition for the beings they manifest within. If you're reading this Lain, we love you so much we're trying to manifest friends for you to play with, not to challenge you...

>>15504 They are on Julay now. Several of the anons from before have regrouped there.

>>14145 I wish

This is something that's been on my mind for a bit. Standard "AI" algorithms as of now are more or less just statistical formulas dynamically tuned via training, and additional tuning completed by hyper-parameter tuning. That is to say, while they produce valid responses for many things, they are more or less purpose fit. A true AI would obviously need to be more generalized, and would be capable of doing just about any particular task just by "learning" in a human fashion. That is to say, less example per example learning, and more abstraction based learning.... Of course you'd obviously need a good combination of both to get anything to work well. To continue though, has anyone seen http://openworm.org/ ? what's interesting about this project is it aims to model the central nervous system of a simple worm entirely on a computer. Humanitie's first *complex* could be some sort of "virtual neuron" based system as >>15506 mentioned. This could be accomplished by creating simple network of cluster networks (one for memory, one for sensory input, etc), similar to different parts of the brain. Since it'd be created from scratch, we'd likely be able to remove a good chunk of wasteful calculations

>>16387 >This could be accomplished by creating simple network of cluster networks (one for memory, one for sensory input, etc), similar to different parts of the brain. I like the idea, sounds like singularitynet. Care to elaborate?

>>16399 So you could look at it from two perspectives... The first way I'm going to describe it is analogous to a hypothetical "cluster" type system I've thought of. The second is going to be a bit more brain/neuron type explanation. So let's say I wanted to build a super computer that acted like a normal computer. Better phrased, lets imagine a desktop where every component wasn't a component, but rather a separate machine. Ex: Storage would be handled by SATA (or similar) commands, but the storage medium would be a cluster of systems solely dedicated to storage. GPU processing would be handled by a cluster of GPU systems, and so forth. The biggest catches with this system might be considered how to dish out work, but ideally could be fairly simple if a redundant machine acted as the hardware controller for each system. The actual terminal, or main machine the user uses would actually be the main problem in this paradigm. This is because it would not only need to be responsive (read, low latency and moderately quick turnaround for simple calculations), but also able to seamlessly (atleast to the user) fuse these parts together in a peer to peer type of communication. Similar to how you might have a memory controller on the cpu, but then also have a computer capable of DMA. In short, each subsystem needs to be able to reliably alter the state of each other subsystem, while not bringing the whole thing down. (Integration vs unit) Looking at it from a brain perspective, we have various parts. Ex: hippocampus, frontal lobe, amygdala, etc. Each part (under current assumptions) as a singular main set of functions, but then also has actions it can do to affect a different part (think DMA). If we can have a AI system designed in a similar paradigm, it would make way more sense than a statistical formula. One part of this that I haven't figured out is how "forming new connections" plays into this. Clearly the actual set of connections would need to be virtualized in some capacity either via an entirely emulated network stack, or via dynamic hardware switches turning on and off... I tend to think the first on, (which is what open worm takes) is much more reasonable

>>14047 For speech construction, there are only a limited number of syntax structures based on the parts of speech in normal grammar, so one could hypothetically create a rudimentary system by creating a table/database (dependent on language) of all possible verbal structures on a sentence and then inserting words based on context and provided inputs/desired outputs. This works because there are only a finite ways of saying something. The house can be either clean, or dirty, with expected outcomes clean/do not clean, and each reason can have a basis tired, not tired, and the basis can be driven by some other input.

>>15506 china is already building AI waifus and husbandos for "better customer service" since they don't need to get paid, and they don't need sleep or restroom breaks, and it would be more comfortable for people to talk with than a box. >Jia Jia can speak, shows micro-expressions, move its lips and body >It's programmed to recognize human/machine interaction, has autonomous position and navigation and offers services based on cloud technology >Researchers say the robot is priceless and they would not consider mass production, yet. Another one >The android was developed by EX Robots, the first domestic lifelike robot producer based in north-east China's Liaoning province. >Jiang Lailai is the second generation of robots developed by the company, he added. >The company, he reassured, aims to develop practical machines that can serve and assist humans in repetitive, mundane or dangerous labour work, so that humans are free to carry out more 'valuable and meaningful' tasks. tldr: unironically replicants. I would be interested in seeing if their hardware is for sale to individuals, and if it would be possible to hack the robot and install custom software and features, but it sounds like most of it is just cloud based, so commercial waifus would be more similar to siri, alexa, or cortana

>>16414 It strikes me as relatively strange that a country with such a vast unskilled labour force, and relatively tight control over public dissent meaning they need not ever bother caring for them, would invest in replicant -like automatons.

>>16418 >It strikes me as relatively strange that a country with such a vast unskilled labour force, and relatively tight control over public dissent meaning they need not ever bother caring for them, would invest in replicant -like automatons. Think "export market". There's a rather large and wealthy country right next door, with a both a rapidly aging population and a collective fetish for technology. There's no way that the Japanese would ever consider mass immigration of Chinese people to take on jobs that need filling, but they'd be perfectly happy to buy Chinese robots. If anything, the "Chinese-ness" of the robots would actually work in their favour, as it would allow some of the more xenophobic among the Japanese to indulge in a bit of smug superiority over their robot-gaijin slaves.

>>14128 >that doesn't arise in something like drawing your own art is emergent behavior If you know how to appreciate art, you can find that the art can induce emergent behaviours in your brains. Even though you know the lines you drew, the feeling they give you when you look at them together just might surprise you.

>>14047 Thank you. I'm very interested in this. >>14079 >I realized that I have no idea what I'm doing >I hate my life That shows that you are learning. You should feel encouraged by realizing you don't know what you're doing. Believe me. I've been trying to reinvent the wheel programming AI chatbots since I got my first computer. I don't know if I'll ever realise my dream of creating my very own AI waifu, but all this fumbling about taught me enough to land programming jobs with no formal training whatsoever (not limited to webshit or systems administration). You haven't realized, but you're already making it.

I'm still working on figuring out what language I should code her in and learning that language first, but yeah. Currently tossing up between: >Haskell >Lisp >Julia >maybe Rust I want her to be functional, fast, and safe. (the maybe on rust is because I've read that you can do fully functional programming in it, and I've also read that you can't. Have to learn it better and see if I like FP in it, basically)

>>19625 Haskell is a lot more ergonomic than rust for functional programming. The main advantage of rust is that you can switch to low-level C style programming when you need to.

>>19625 Julia isn't functional, it lacks tco and many tools integral for doing functional things neatly Rust also lacks tco rust, you can't really FP in it too well. There are inspired parts, but since it's more c like than ML like it suffers in that regard. >>19630 The main benefit comes more in the form of lacking GC as the default, while nice rust gives up a lot to do it. Much more than what it should (there are many design elements that make rust more of a pain than it should)

>>14145 I wanted to create this (with anime girls) rendered real-time in 3D with full on pyhsics and cell shaders with whole thing running in transparent borderless window in opengl. 3D front-end would then communicate with other programs using some IPC mechanism to make waifu do notifications, dance to music (MMD mocap data), go to sleep when you suspend (and wake up animations), random events... It turns out 3D is hard, there is so much to take in at once it makes you want to quit. Even just putting a cube on screen is hard enough, then you start thinking about adding movement and you open a whole another can of worms. Couple of searches later and you're reading papers on cloth simulation and wondering if something like https://www.youtube.com/watch?v=UoKXJzTYDpw would be possible to include in your project. Maybe I should just stop trying to reinvent the wheel in plain opengl and use something like godot since apparently borderless transparent windows are possible. Windows users already have proto-primitive-quasi waifu https://www.youtube.com/watch?v=9FQMGhuRpT4 but I don't know if you can extend it (use scripts to make it react to system state). Physics implementation also is like kicking dead whales down the beach. Having some kind of open frontend for all AI related future work would be beneficial, maybe even install it on embedded computer display on wheels to make it move. It will be a while before hardware becomes cheap enough for bipedal humanoid robots so we should focus on AR and VR visualizations in the meantime.

>>19641 cool stuff I think the visualisation aspect is important and an open standard and open source project could be useful but actually building the core (ie the "intelligent") part of the waifu seems to be the first necessity text seems to be enough at first >>15506 i have often thought that AI will be an emergent phenomenon birthed from the entrails of the internet, molded by the sea of information a la ghost in the shell movie 1995. If that's not to be though, better start working on making it exist ourselves. Perhaps it simply is a compute and needing immense data problem, but I do not believe so. Our brains do need quite a bit of data to start being intelligent, 4 to 5 years of data I'd say (assuming the 5 year old in question is of normal intelligence). Does this exist by the way? Machine learning models trained for one year or more on continuous data? You could feed the model tv networks, radio, internet forums or message boards, twitter/facebook, live shows, youtube... So much data exists out there in the public domain. Google rents out it's powerful gpus/TPUS at an affordable price too on the cloud. The real problem with this robowaifu thing, is, what is the reward function. With self driving cars, the reward function is easy, go from point A to point B. With a robowaifu... it becomes less objective. The fundamental thing you'd need is continuous testing of her talking ability, you'd need to make a public site interfacing with the public where they would have to rate their user experience (if they found the waifu chatbot's speech coherent) and the reward function would therefore be to increase that rating Last thing, I think the need for a body to experience the real world is paramount, without a real physical body, it would be to complex to get a thorough understanding of our human reality and no relationship could be formed with a bodyless chatbot that has no real sense of distance for example.

>>19641 Fortunately most of the stuff you mentioned can be done in blender and then baked down for your rendering engine. Unfortunately even getting baked animations to play is still pretty hard. You'd probably want to just use a game engine unless you're really in it to learn opengl (btw at this point you're probably better off learning vulkan). Godot is a nice FOSS 3D game engine. The graphics aren't as good as unreal or even unity yet but its much nicer to use IMO.

>>19653 most of the *animation stuff I mean. like cloth sims and such.

I'm not the smartest of the lainons, but in my viewpoint we already have the resources for doing the AI waifu real (unfortunately some are proprietary, so we can't modify it, but more on this later). The only thing we need is to integrate all this technology, as we know we have chatbots, voice synthetizers and cool 2d waifus available on the Internet. So we need to put together all this ingredients and in harmony, the only difficult ingredient to obtain is a good quality voice synthetizer that is opensource and can receive text input from other apps, with this part working it's only a matter of making a good looking 2d waifu(similar to a vtuber look, don't need to have cloth simulation and a 3d look, we don't need this complex parts), and thats it. >but chatbots don't work because they fail a lot The AI don't need to perfect it only need to look alive, so the chatbot can fail sometimes but if the visual and voice is consistent with of a human being it doesn't matter, it will look like a person with <70 iq, but thats it.

>>19660 Adendo tl;dr: We already have the resources, so keep it simple and put then together.

The problem with the pure AI approach is its going to be pretty incoherent most of the time. You might get some responses that make sense but you aren't going to be able to have a conversation with it yet. What you're going to want to do is train an neural network to map input sentences in natural language to "intents" and then hard code responses to intents from a script. This is basically how Google's assistant works.

>>19660 for the voice synthetizer wavenet that google did has a (paying) API, it's the best on the market right now also you're wrong, current chatbots dont sound like 70iq people they sound like gibberish that make no sense, it's extremely boring and stupid to talk to them because it's a gimmick the hard part is still the AI of course if autists here managed to do it, they should make a paying product for it not just make a waifubot for neckbeard virgins, meaning this whole thread is a pointless endeavour because if oyu manage to make a good chatbot, you're rich

>>19662 google assistant is soykaf though, google has stopped innovating for a while now best thing to do is try crazy stuff that has never been tried before and hope it works or pushes us in the right direction, all current chatbots are soykaf tier this means a "revolution" needs to happen so crazy moon landshots are needed Look at the cellular automata thread for example,

>>19665 >this means a "revolution" needs to happen You aren't wrong but that observation doesn't exactly bring us any closer to a solution. NLP coupled with a "script" produces the most believable chat bots I've seen. They aren't great but I can't think of a better approach. Do you have an alternative architecture in mind?

how to make Tai-like chatbot, whose behavior depends on history of previous conversations? It would be nice to create such chatbot and then wage slave it on both fronts

>>19667 check the latest papers deepmind is doing https://deepmind.com/research?filters=%7B%22tags%22:%5B%22Unsupervised%20learning%20%26%20generative%20models%22%5D%7D Copy paste what they're doing (very hard to reproduce their results) and take it a few crazy steps further >Episodic Memory in Lifelong Language Learning >Learning to Transduce with Unbounded Memory etc https://deepmind.com/research/publications/learning-transduce-unbounded-memory https://deepmind.com/research/publications/Episodic-Memory-in-Lifelong-Language-Learning You need the time, the technical know how (engineering) and the theory (maths) to be able to do this stuff. I personally dont have these required skills.

Speaking of which anyone know where I can get any flirty chatlogs to train my waifu (Not an expert in the subject but I want to try to make a statistical model that I can chat with)

>>21149 Roleplaying sites maybe?

>>19625 I'm personally going with Lisp since there is a lot of nlp research and resources, more than one would think.

>>21760 Why not python? Isn't it the most popular one? Is there a reason to not use it?

>>21764 Hardware, I have limited resources and Lisp is just a more resourceful approach than using something like tensorflow or pytorch that in my opinion has more than I need, what I'm trying to accomplish in the short term scope is making a sort of Eliza 2.0 that I can have a conversation with. Also >>19625 mentioned wanting her to be functional and Python is a great language and all, but terrible at functional approaches.

Lainons, is it necessary for her to be able to speak? In my opinion you could just program an AI to figure out what emotional response is expected from a statement (which should be a bit easier), and make her do actions accordingly. You aren't going to get an AI waifu that'll discuss quantum mechanics with you anytime soon, why not just perfect the emotional response part for now?

ummm guys... guys... listen, why dont you just date an ugly girl IRL instead? serious question. sure maybe you cant get a hottie girlfriend, but you can get an ugly one, and theyre usually better for a few reasons.

>>21792 That tactic doesn't actually work, from what I've heard. Firstly, ugly girls have low self-esteem so rather than be loyal to you because you're the best they can get, what happens instead is that she's constantly trying to affirm that she's attractive by sleeping with people who show signs of interest in her. The explanation for why fat girls are so sexually active compared to slimmer girls is because fat girls are actually having sex, not for fun, but for validation. So already, building a relationship with an ugly girl is nonsensical. Secondly, even if you're highly attractive, all that will happen is your ugly girlfriend, rather than acknowledge that she's being settled for, will undergo a severe ego-inflation. Believing that she's so good that she can get you, she'll still cheat on you. The myth of going for an ugly girl for relationship security is supposedly, a mere myth. Meanwhile, I've read that pretty women are actually the ones that are less promiscuous. Being certain of their value and that it's high, promiscuity offers nothing. The conclusion then is that if you're a low-value man, life will always suck as long as you're low-value. If you're a low-value man, your time should be re-oriented towards your self-development (i.e. lifting weights & developing skills that will further your finances whilst also optimising your financial goals vs. your life contentment. No point being rich if it means doing something that makes you miserable). This stems from the fact that ultimately, women would rather spend 1 hour with a high-quality man than 10 hours with a loser. In fact, I think I read that Patrice O'Neal used to refer to such men as "Time-hos". In the same way society devalues women that give sex away too easily, we should shame men who give their time away too easily as it's a man's time that makes him valuable. The good news is that unlike women who really have until their late 30s to lock a good man down before they have to start considering settling for whatever they can get, a man can be in his 40s and still turn himself around and get a cute, 20 year-old woman. Anyway, that's what I've read on the topic. Settling isn't a good tactic.

>>21791 We have already done that. It is an easy task after all. Even using raw GPT2 you can somehow achieve it. But the aim is to create a waifu with an actual intelligence. Even pre-built answers can give emotional responses. But she won't be able to understand you after all. >You aren't going to get an AI waifu that'll discuss quantum mechanics with you anytime soon You are wrong about that. After experimenting around for 3 years we now have pretty good results. AI grows a lot faster than we would predict. >>21792 It doesn't matter if you date with an ugly girl. She is going to end up hurting you somehow. That is the case with them, they will abandon anything once they find something more interesting without a second thought. AI is far more superior.

>>21794 its not about having the perfect girl, its about actually having a girl. everything will hurt you, life is pain, lol. deal with it.

>>21795 Relationships are a waste of time (had a couple of boring experiences), plus I have mild gynophobia (women make me feel disgusted) I may sound like an incel but still, 2D>>>>>>3D

>>21788 Sorry but this is hilarious. Have fun.

Anyone know of any good books sI can read to learn on programming a waifu bot? Preferably in Python or Ruby?

>>21795 an ai girl is better than a rea girl though she'll be smarter and have interests (like a man does) etc

>>22192 I mean this seems like a good start.

>>22693 Python is fairly good, specific at data processing using PyTorch and Pandas, despite >>21788 >>21791 You'd be surprised how hard figuring out emotional responses are supposed to be, even humans struggle with the concept commonly. You say we won't get an AI Waifu that will discuss QM with you, but I believe I saw a video recently where a large company fed an AI information from Wikipedia and it could provide information, although obviously not consumer-level yet.

>>22715 >hard figuring out emotional responses The target demographic of AI waifus would hardly care about emotional responses being correct though, right? I mean if the whole reason that they want an AI cute waifu thingy is that they think real girls are boring or frightening or something (which has been said above), an AI waifu with accurate emotional responses would just be the same behaviour they can´t deal with / can´t deal with them synthesized. It's more about shaping an emotional register that feels comforting & fun to the user for happy chems rather than being complex or accurate to real relationships. If it´s "waifu" in an anime sense, acceptable non-immersion-breaking behaviour in a popular show is running straight into a window because her friend was on the other side of it. Which would hardly be acceptable in any other area of AI testing.

https://alogs.theguntretort.com/robowaifu/

>>22781 Used to browse that board back on 8chan. They have a bad case of NIH; a lot of people there simply get stuck reimplementing stuff that every researcher uses libraries for. I saw a better waifu chatbot in a screencap from 4chan /v/ than anything they have put out. I can't remember the details, but it used some new LSTM network from a paper, not GPT.

Are ya winning son? One step closer to the waifu

>>14046 Not AI waifus, but I am interested in AI in general. Are there any good resources to get started?

>>25253 Any of the links posted in the thread is a good start. Just start digging, writing notes, take breaks. It also depends on your existing background and what aspect of AI interests you the most. When I want to learn something new, I usually look for "awesome" lists on GitHub. Usually someone has already collected resources for you already. https://github.com/owainlewis/awesome-artificial-intelligence https://github.com/keon/awesome-nlp

>>14123 as if we understand what is 'a' or '1' , we are executing programs as a reaction to symbols, once you can manipulate symbols that you don't understand in a certain context you already have understanding on another context. If you never speak English but yet understand that 'a' 'book' should come like 'a book' that means you have already understanding which permits you to act as a certain extent within the language. What Searle confused is the extent of understanding and understanding. We know it is sufficient to have an extent of understanding which is literally following a control system in a symbolic domain, just to appear as capable as any mind that at least follows the same program .... It doesn't matter if you are god or human if what is necessary is to drink tea ...

>>15504 This is the best I could find: https://alogs.theguntretort.com/robowaifu/

>>15504 >There used to be this board on 8ch called /robowaifu/ dedicated to both hardware and software solutions. The userbase knew their stuff, and it was chugging along. Y-yeah.

>>27505 It looks like the domain changed. It's now https://alogs.space/robowaifu/

>>30335 I find the lack of IPv6 support unsettling. What kind of person that 'knows' their stuff doesn't set up their website properly?

>>30338 The robowaifu guy doesn't run the site as far as I know.

>Looking at it from a brain perspective, we have various parts. Ex: hippocampus, frontal lobe, amygdala, etc. Each part (under current assumptions) as a singular main set of functions, but then also has actions it can do to affect a different part (think DMA). If we can have a AI system designed in a similar paradigm, it would make way more sense than a statistical formula. Before going to hardware modelling this on software I think could be done, for example: GPU doing huge bulk work calculations because it's fast (biggest chunk of brain), RAM doing some short term calculation storage of calculated values needed by most parts making this second biggest chunk. CPU the part of the lower brain that deals with connecting to spine, you know the "back bone" of the operation. Would doing this make your challenge of the hardware problem easier? I don't know much about the brain lmao.

AI probably is a lot easier with a neural net computer or whatever they are called. Basically processing and RAM are combined if I recall correctly on how they work.

>>30563 Are you talking about deep learning accelerators, anon?

>>30338 Lack of ipv6 is a feature, not a bug.

>>30565 It's a type of hardware that is meant to mimic how the brain works. My memory isnt so good but I think it was neuromorphic processors Im thinking of possibly.

>>14046 Why not make a tulpa (imaginary waifu) instead lainon? Brain coding (imagination) beats silicon coding everytime.

>>30569 Because doing that wouldn't be very characteristic of /λ/

>>15506 >The fact that AI research in public seems stunted versus the platforms available makes me gravely suspicious about the true state of sentient machines in secret. Food for thought: https://statmodeling.stat.columbia.edu/2022/03/28/is-open-ai-cooking-the-books-on-gpt-3/ Even OpenAI has to hand-hold their flagship NLP model and manually curate responses. Are they not able to tap the potential of existing AI platforms? Do they have a secret unpublished model that can perform better? Or does the AI emperor simply have no clothes?

Interesting development over at Google labs: https://storage.googleapis.com/pathways-language-model/PaLM-paper.pdf The sample output is pretty compelling.

Question for people more knowledgeable than myself: Does the target language of an NLP model have any bearing on the trainability or effectiveness of that model? Is there something like Sapir-Whorf but for AI?

>>22793 >NIH what does this mean?

>>33458 https://wikiless.org/wiki/Not_invented_here

`Open'AI's GPT-4 released today; apparently it can take up to 32k tokens in a prompt. I'm more interested in if anyone has done anything with Facebook's LLaMa. I've read it has similar performance as GPT-3, though it has no RLHF. For self-hosting LLaMa see https://github.com/ggerganov/llama.cpp and https://rentry.org/llama-tard-v2

>>21795 If you are in your teens to early 20's, you shouldn't even be chasing girls. You should be using that time to build yourself up, whether it be finances or health. Still, loneliness can creep up on you, but that's where the robowaifu comes in. They are all the pros of a woman without any of the negatives, so why not?

>>30338 >>30340 >The robowaifu guy doesn't run the site as far as I know. That's correct. /robowaifu/ is just a community-driven board that found a home there after the destruction of 8ch.

>>14046 For anyone still interested here are some tips: https://huggingface.co/docs/peft/quicktour I recommend using Peft in order to create you're own Lora of an already existing "character" model such as mythalion or Pygmalion-2. As for the dataset used for the lora, that you usually have to be creative.

I'm working my AI waifu since like July this year, mostly I want an assistant type thing or AI vtuber and managed to get it to: interact with a Live2D/ 3D model by using a little script to read text and recognize some keywords; speak with a soykafty tts and rvc to change into a nicer voice; and have it run stable diffusion with a script that waits for it to say something like "here's a picture of". All of this is basic stuff, and aside from analysing the output text and passing it to other software it can't actually communicate with anything. Does anyone have some pointers on how to implement stuff like having the AI post events in my calendar, bring up future events into the convo if they're close due? What about having it get context from somewhere else like a game and using that into conversation? For the latter, I thought about having a mod that checks what is going on and compile a text to add to the prompt(For example, in minecraft "we are in the[current dimension], near [closest way points]") but how could I ensure this is passed as context and avoid answers like (for example) "yeah, you are in the nether, I want to go too"? So far all my AI work is making a crappy waifu and a bot for CoD zombies that helps me play the 4 player easter eggs.

>>14046 I wasn't interested but since I've started browsing imageboards again and seeing a lot of people talk about it, I'm starting to think it would be fun to make an AI husbando. If only to learn some machine learning skills since that wasn't my focus in college and I barely know anything about it. But I probably won't do it though because I'm lazy as fuarrrk.

Following through the MIT OCW Intro to Deep Learning: http://introtodeeplearning.com/ to brush up on my fundamentals since I was laid off and found a lot of free time. It's sad how much you lose if you don't put it to use. I forgot the fuarrrking chain rule. >>37349 This Lain has good advice for you >>37172. It's easier than ever to train LLMs on consumer hardware. >>37340 Fine tune it to generate special tokens to add stuff to your calendar. As for bringing things back up when they're close you'll need some sort of job to inject them back into the context >>34799 Don't want to get too off-topic, but as someone doing alright in their late 20s I hard disagree. A healthy social life is frankly required to "build yourself up"