Yes although you would probably have to choose another model to build on top of. So the bot would be a fusion between your ex and like Makoto from Persona 5
I'll jump on the AI waifu train when a continuous training model gets released. Until then, I refuse to develop a relationship with an intelligence that isn't actually molded by its interactions with me and will forget me as soon as I'm out of its token window
yeah I can see continuous training probably being the next MoE style leap
It's not too hard to save a persistent identity and have it included at the top of the prompt each time, and you could even get GPT4 to review the most recent conversation and decide if the identity needs to be updated a little, so that the identity isn't completely static. I'm expecting that OpenAI have basically solved the continuous training problem and GPT5 will never be more than a month out of date.
Problem is that 'top of the prompt' space is limited and won't fit a lot of tokens. We'll need something that doesen't need to evaluate all million tokens of your lifetime conversations for every answer
>all million tokens of your lifetime conversations for every answer
I hate to break it to you, but your lifetime conversation history is very compressible. A good AI can probably extract they key points in less than 1% of the total number of tokens. Also, a million tokens is a realistic context window soon anyway.
https://arxiv.org/abs/2304.11062
https://manifold.markets/IsaacKing/will-any-llm-have-a-context-window
4 months ago
Anonymous
Why would you hate to break it for me? It's absolutely fabulous if you can (and maintain any reasonable quality)
4 months ago
Anonymous
>Why would you hate to break it for me?
If someone's conversation history is very compressible, it means they repeat themselves a lot and rarely have any new ideas. I thought you might take offence at that suggestion. It was all in jest, though. I don't know what the upper limit is on the quality of experience that this technique can achieve, but I imagine that something similar is used by platforms like BabyAGI and Replika.
I have like 3 years of daily texts with my ex (about 50000 rows last I checked) I'm saving to train a model one day. I also have like 10 years of texts with my female best friend that I'll probably use too (she's cool with it)
I have like 3 years of daily texts with my ex (about 50000 rows last I checked) I'm saving to train a model one day. I also have like 10 years of texts with my female best friend that I'll probably use too (she's cool with it)
it's possible the ex was very good at pretending to be a decent woman, right up until her disappearance
that would mean that the training data should be free of any toxicity
Both male and female are susceptible to return to their abusive exes. They deserve life being living hell for not saying "no". Don't try to understand normie brains. Normies don't understand themselves. They just exist to fill our world and amuse us.
>Break up with me? >Well frick you I'll replace you with an AI >An AI of you! A better version of you!
Last I checked /lmg/ wasn't quite there yet, but I look forward to seeing the existential horror of woman and normalgays realizing men can just re-create them and improve them
>using training data from 3DPD when there exist squintillions of hours of dialogue from Japanese visual novels
Someone has to save the normalgays from themselves... the weeb man's burden
I don't think this is a good idea, if you succeed, now you're talking to an echo of a person that betrayed you(r trust), is that really gonna be good conversation? 'she' is not returning with this, and should you really want that, even? she chose to end it, and in a shitty way, I would avoid her as the plague
Yes although you would probably have to choose another model to build on top of. So the bot would be a fusion between your ex and like Makoto from Persona 5
Wouldn't a perfect replica of her also leave you?
You're absolutely right
You need a therapy, dude
Does it bruise up and cry on the floor after I punch it in the face? Does it turn purple after I throttle it? If not it can't replace a real woman.
You're describing a gynoid. AI is what you put in the gynoid
ChatGPT (and AI generally) is a fricking SEARCH ENGINE, not your waifu. It’s just coomer shit, same as always.
If a PNG can be your waifu, so can a search engine. Next are you going to tell me movies aren't real?
I'll jump on the AI waifu train when a continuous training model gets released. Until then, I refuse to develop a relationship with an intelligence that isn't actually molded by its interactions with me and will forget me as soon as I'm out of its token window
We'll either need an order of magnitude of more computing power or a lot more efficient models to do that. Both are possible this decade
yeah I can see continuous training probably being the next MoE style leap
It's not too hard to save a persistent identity and have it included at the top of the prompt each time, and you could even get GPT4 to review the most recent conversation and decide if the identity needs to be updated a little, so that the identity isn't completely static. I'm expecting that OpenAI have basically solved the continuous training problem and GPT5 will never be more than a month out of date.
Problem is that 'top of the prompt' space is limited and won't fit a lot of tokens. We'll need something that doesen't need to evaluate all million tokens of your lifetime conversations for every answer
>all million tokens of your lifetime conversations for every answer
I hate to break it to you, but your lifetime conversation history is very compressible. A good AI can probably extract they key points in less than 1% of the total number of tokens. Also, a million tokens is a realistic context window soon anyway.
https://arxiv.org/abs/2304.11062
https://manifold.markets/IsaacKing/will-any-llm-have-a-context-window
Why would you hate to break it for me? It's absolutely fabulous if you can (and maintain any reasonable quality)
>Why would you hate to break it for me?
If someone's conversation history is very compressible, it means they repeat themselves a lot and rarely have any new ideas. I thought you might take offence at that suggestion. It was all in jest, though. I don't know what the upper limit is on the quality of experience that this technique can achieve, but I imagine that something similar is used by platforms like BabyAGI and Replika.
Roastie rage
Good
You can easily finetune a local model.
Training from scratch is not needed and awaste of ressources anyway.
Use any recent 7B model and that‘s it.
>just coomer shit
So better gf than a real woman
I have like 3 years of daily texts with my ex (about 50000 rows last I checked) I'm saving to train a model one day. I also have like 10 years of texts with my female best friend that I'll probably use too (she's cool with it)
You don't necessarily need to train it provided the memory is large enough. Mostly they're limited to 5K-10K tokens though
you mean just use it as a character card? nah its a good converational dataset. I'm talking rows not tokens, god knows how many tokens it is.
BTW has anyone tried plugging ChatGPT into Monika? Seems like an obvious mod…
>gf breaks up and ghosts you
>still want to talk to her again
Why? These kinds of women shouldn’t be given any attention.
it's possible the ex was very good at pretending to be a decent woman, right up until her disappearance
that would mean that the training data should be free of any toxicity
Both male and female are susceptible to return to their abusive exes. They deserve life being living hell for not saying "no". Don't try to understand normie brains. Normies don't understand themselves. They just exist to fill our world and amuse us.
>Break up with me?
>Well frick you I'll replace you with an AI
>An AI of you! A better version of you!
Last I checked /lmg/ wasn't quite there yet, but I look forward to seeing the existential horror of woman and normalgays realizing men can just re-create them and improve them
>using training data from 3DPD when there exist squintillions of hours of dialogue from Japanese visual novels
Someone has to save the normalgays from themselves... the weeb man's burden
I don't think this is a good idea, if you succeed, now you're talking to an echo of a person that betrayed you(r trust), is that really gonna be good conversation? 'she' is not returning with this, and should you really want that, even? she chose to end it, and in a shitty way, I would avoid her as the plague
Wow. Is this the next isekai goyslop from our based nippon bros?