It's the difference between the pretrained and chat/instruct fine tuned models.
The TextCompletion DaVinci was way better than the ChatCompletion model in variety of language.
Trying to get the chat model to generate marketing copy was laughable. It looked like a high school senior's idea of copywriting, and was nearly impossible to correct.
The base model was pretty easy to get great results from as long as effectively biasing the context towards professional copy.
Even the fact that you can't set the system messages at a core level is silly.
I can't have the model actually be told it is an award winning copywriter. Instead it effectively gets told "you are a LLM by OpenAI pretending to be an award winning copywriter."
Really too bad that 99% of the training data it was built on top of wasn't being written by a LLM by OpenAI.
It's effectively narrowing the context space unnecessarily and creating a bottleneck that severely limits the application of the SotA model, but it still scores well on the handful of tests it is being evaluated on so no one bats an eye as it seems no one at OpenAI has heard of Goodhart's Law.
The TextCompletion DaVinci was way better than the ChatCompletion model in variety of language.
Trying to get the chat model to generate marketing copy was laughable. It looked like a high school senior's idea of copywriting, and was nearly impossible to correct.
The base model was pretty easy to get great results from as long as effectively biasing the context towards professional copy.
Even the fact that you can't set the system messages at a core level is silly.
I can't have the model actually be told it is an award winning copywriter. Instead it effectively gets told "you are a LLM by OpenAI pretending to be an award winning copywriter."
Really too bad that 99% of the training data it was built on top of wasn't being written by a LLM by OpenAI.
It's effectively narrowing the context space unnecessarily and creating a bottleneck that severely limits the application of the SotA model, but it still scores well on the handful of tests it is being evaluated on so no one bats an eye as it seems no one at OpenAI has heard of Goodhart's Law.