if you’re an openai api customer, i’d love your feedback. what annoys you the most? what features do you want us to add? if you looked at using the api and didn’t, why not?

Aug 20, 2022 · 4:10 PM UTC

253
27
6
537
Load newest
Replying to @sama
It’s amazing, thanks! 1. For support apps: Improve grounding in fact, reasoning. Reduce hallucinated “facts”. 2. Improve performance on complex/multi-step problems. 3. Predictable scaling + cost.
Replying to @sama
We had started using OpenAi in the Beta and then suddenly we were told Open Ai is not available in your country. Kindly asking what the reasoning for country restrictions is.
Replying to @sama
The playground UX isn't very intuitive for seeing the possibilities (for a non-engineer). Takes too long to get to the magic moment (but once you get there, 🤯). Build some more interactive/fun/mindblowing demos to make it easier to see the possibilities.
Replying to @sama
Token limit but that’s not an api limitation but a model one, looking forward for GPT-4. One thing I’d love would be to be able to do a “continuation mode” which you don’t have to send the previous prompt and completion but only an ID of the session and it would to use previous
1
Replying to @sama
Just thought of something else. Fine-tune by conversation. You give a prompt and add the best (subjective) answers to a dataset to be used for fine-tuning. Almost like a single player Reinforcement Learning akin to what you did with InstructGPT.
Replying to @sama
In the playground a find & replace would be super helpful
Replying to @sama
It’s impossible to use fine tuned models in a real time app (eg chat). API waits for load and then times out. Need one or more of 1) pay to keep model loaded 2) api option to fail immediately if model isn’t loaded (so I can fallback to non FT) 3) api to query model status
1
2
Replying to @sama
-Better prompt suggestions eg: autocomplete for prompts ? -A way to upload my own data (structured, tabular & unstructured) to "co-parse" it with GPT. -Clear guidelines for what is ok / not ok for GPT3 apps that have wide, open ended use cases like bots or summarization.
2
Replying to @sama
Why is Dalle more expensive and worse than current (free) diffusion models? Also why has openai switched to profit mode? Wack
1
Replying to @sama
text-davinci-002 seems heavily biased towards giving “I don’t know” completions, unless I crank the temperature up to 1.0. Same with “There are many possible answers to _” instead of GPT3 picking one.
1
Load more