Amazing product -- great experience so far. Biggest hurdle for us is figuring out how to best ingest legal documents larger than the token limit (can sometimes be 10x the token limit)

Let me (somehow) pay for generated tokens but not context tokens

I think a way to streamline few shot, iterative prompt training would be extremely helpful. It would work like this…enter your prompt, apply it to the data, auto save each experiment, rate the output, mark the better solution. Repeat until satisfied.

Using embeddings for semantic similarity / vector search could be muchhhh easier than it is today

The ability to save a long prompt and reference it, similar to a trained model. It would be more cost effective for even the higher models.

Would some kind of caching option be possible so that if you wanted to continue a long conversation, you would not need to keep sending all of the history for context. Currently, I guess people are truncating the history to keep context. Could this be improved/automated?

How about an open ai + zapier dealio

I love the GPT-3 API. Will DALLE-2 be available via API? My company wants to help YouTubers generate thumbnails, so I would love to be added to the DALLE API beta list.

There's still a lot of handwaving on best practices. It's expensive to benchmark prompt usage and not centralized anywhere; would be amazing to have datasets and scoreboards using hold out test sets with free/subsidized compute for community to figure out best usage.

Hmm I’ve been really happy with gpt-3. I wanted to do a bit more by utilizing long term and mid term memory, not sure if there’s some advanced documentation on that?