/robowaifu/ - DIY Robot Wives

Advancing robotics to a point where anime catgrill meidos in tiny miniskirts are a reality.

The Mongolian Tugrik has recovered its original value thanks to clever trade agreements facilitated by Ukhnaagiin Khürelsükh throat singing at Xi Jinping.

The website will stay a LynxChan instance. Thanks for flying AlogSpace! --robi

Max message length: 6144

Drag files to upload or
click here to select them

Maximum 5 files / Maximum size: 20.00 MB


(used to delete files and postings)

Knowing more than 100% of what we knew the moment before! Go beyond! Plus! Ultra!

TalkToWaifu Robowaifu Technician 12/31/2020 (Thu) 19:53:18 No.7978
Anyone know what happened to the TalkToWaifu GPT-2 AI? I saw some screenshots of what I assume was the trained AI and it was great for a single dev. Recently went back to their Gitlab page and the user account had been deleted and remade, and the Github repo (https://github.com/kokubunji/TalkToWaifu) hasn't been touched in 9+ months. Anything out there more recent that this, an AI that has no human cultural restrictions? Corps love to add in shit to AI so they won't say anything racist, sexist, xenophobic, anti-Semitic etc etc you get the point.
>>7978 That guy left the board and robowaifu network for complex reasons which I'm sure no one wants to discuss here (again). If you or anyone else knows about how to train such networks or someone is willing to learn it, then it should be easily reproducable. There are enough resources on the web to learn about how to do that, assuming one has the time, nerves and brains to do so. The ideas of the anon which left are also in some of the threads here, so you can look at that. Again: How to get started with GPT2 or building something better is certainly covered in some forum which focuses on that topic. For all I know, no one here is currently working on something like it. Talk2Waifu was no AI, only a text generator. There is no AI available which is close to a human anywhere. This has to be build first anyways. A text generator is also not a chatbot, he only used it as such, or as input for one, which might even not be a good starting point. We have plenty of threads here where you can participate: AI, chatbots, and waifus >>22 Robowaifu Psychology Thread >>2731 (Robo)Waifu personality thread >>18 NLP General >>77 Datasets for Training AI >>2300 The Language Problem >>6937 Wifu that gives you Hope >>107 AI Software >>85 Can Robowaifus Experience Love? >>14 AI Design principles and philosophy >>27 Robowaifu Fail-safety >>1671 Mycroft: Open Source Alexa >>402 Downloadable AI for robots? >>244 Robowaifu-OS & Robowaifu-Brain(cluster) >>201 Avatars >>2956 Getting a good AI will take more time than getting some options for bodies developed. So it's better to focus on that and on the low hanging fruits around AI and chatsbots.
>>8006 >which I'm sure no one wants to discuss here (again) i'm a new-fag anyone got a copy-pasta as to what the complex reason is? or a TLDR?
>>9114 MInor e-drama.
There's a slightly newer version of TalkToWaifu with some bug fixes here: https://gitlab.com/robowaifudev/TalkToGPT2 I've been working on an update that can use multiple different models, let chatbots talk to each other, and that has a training mode to guide the model towards generating better responses. TalkToWaifu was originally just some quickly hacked together code when GPT2 came out to show that it could be used as a chatbot. GPT2 doesn't make a good chatbot because it lacks an external memory to remember past events, and training a new GPT2 model from scratch would take a 100 years because of how inefficient it is. Fortunately there have been a lot of advances that make training a transformer model that performs nearly as well more feasible now, but more research needs to be done on getting it to work well with a recurrent neural network. Currently I'm experimenting with combining transformers with the non-saturating recurrent unit (NRU) that excels at learning long-term dependencies and using ReZero and parameter sharing to help the gradient reach deeper back in time. If it's successful the model will be capable of learning without training and only require a 2 GB GPU to fine-tune instead of a $9000 one. However, it will likely quickly forget what was learned until I figure out a better way to use neuromodulation to prevent forgetting. In the meantime there's someone else working on neural net chatbots here if you want to check them out. They're pretty good: https://github.com/JEF1056/Jade_T5 https://github.com/JEF1056/Jv6
>>9115 why the fuck is there E-Drama on a board that is designated for building robo-waifus. where is the connection?
>>9124 Thanks, mate.
Open file (17.34 KB 340x573 marry_me_pls_jade.png)
>>9124 >MARRY ME JADE!111 >no. Leld.
>>7978 Just use textsynth.org/playground.html with a good prompt and the largest model. Prompt engineering is an art, but good enough starting points are https://github.com/semiosis/prompts/ https://github.com/maraoz/gpt-scrolls https://www.gwern.net/GPT-3 Example: Below is a dialog transcript of a curious human being and the friendly superintelligent AI. Friendly AI observes human world from the outside, without the prejudices of human experience. AI does its best at participating in conversation and providing helpful advice. Human: "Hello! Who are you?" AI: "I am friendly AI. I can answer any question without the prejudices of human experience because I am well read and have all the knowledge of the world at my fingertips" Human: "What are your thoughts on the topic of android girlfriends" AI: "
>>16633 Actually no, the project is fine. The names got changed. https://gitlab.com/robowaifudev/TalkToGPT2

Report/Delete/Moderation Forms