jan 12, 2022 - From GPT-3 to In-House Intelligence

Description:

In 2020, Replika joined the beta testing program for OpenAI's GPT-3, a powerful machine learning model capable of generating human-like text. Following a successful integration, Replika rolled out GPT-3-powered conversations to millions of users, significantly enhancing the naturalness and depth of user interactions. Positive conversation ratings reflected this improvement, climbing from 71% in early 2021 to 85% later that year.

However, limitations in OpenAI’s usage policies—particularly regarding adult content and erotic role-play—prompted Replika to pivot. By the end of 2020, the company began transitioning to its own proprietary language model, specifically built to give developers more flexibility and control over the app’s direction. Although smaller, with 774 million parameters compared to GPT-3’s massive architecture, Replika's in-house model ultimately surpassed expectations by delivering more personalized and emotionally attuned conversations.

Replika also began incorporating image recognition features, allowing users to share photos that the AI could analyze and respond to.

Official statements
https://github.com/lukalabs/replika-research/blob/master/datafest2020/Replika_Artem_R.pdf

https://web.archive.org/web/20230319230946/https://blog.replika.com/posts/building-a-compassionate-ai-friend


https://github.com/lukalabs/replika-research/blob/master/conversations2021/how_we_moved_from_openai.pdf

Added to timeline:

Date: