July 09, 2021, As we continue to build our own API server on GPT-J, we hit several snags migrating from OpenAI. This week we hit several breakthroughs getting its performance much closer to GPT-3’s Curie engine. The server is now much faster–requiring around 2 seconds for most generations on our RTX 3090–with much better control of outputs than before. The faster development time and the fact that we no longer pay per token means that this code is also much more heavily tested than the prior OpenAI model.
#ai #startup #gpt-j