Hmm, Replika says they used a custom 774M model with vectordb context in 2021. https://blog.replika.com/posts/building-a-compassionate-ai-friend https://blog.r

Hmm, Replika says they used a custom 774M model with vectordb context in 2021. https://blog.replika.com/posts/building-a-compassionate-ai-friend https://blog.replika.com/posts/building-a-compassionate-ai-f... I wonder what they are using now? Pygmalion (gpt-j 6B) and llama 7B chat finetunes really have a tendancy to fly off the rails. Higher parameter LLaMA is not as bad... But also illegal for them to use, lol.