Get the popcorn ready, the drinks filled to the rim. Last years DevDay was jampacked with great information, and I have no doubt this will be a worthwhile watch.
@jr.2509 shared a screenshot from an OpenAI employee hinting of some of “the most requested” API features being announced during the keynote & it got me excited.
Assistant enhancements are also top of my mind, in particular greater control and customization of file search (such as the ability to include and filter by meta data).
Greater control over o1 reasoning tokens would be great although that seems too early.
+1 on just better cost control and reporting, preferably via both a dashboard and API. E.g. simple API access to get stats on utlization - number of input and output tokens, per organization and per API key, and for specified time period.
I also agree, the killer announcement would be Perplexity competitor - web search via API, with citations!
Seriously, though, I don’t think there’s anything specifically planned for online. Last year we did a Bingo thing as you probably recall, but no live presentation put a damper on that…
Will be good to discuss with people as the videos drop, though, I think.
Search bar on ChatGPT history (ok fine, not an API feature, maybe with assistants?)
o1 release?
I don’t think AVM or tts-2 will be out, it would certainly be a surprise
sora-preview would be nice
option to view how much a request would cost (I know there’s a way to do this, but an improvement on this would be nice) - specially on o-models (o1-mini, o1, o1-preview)
search the web would also be nice, but I’m not sure that is something they’ll want to add to the API due to competitors
improvements to the assistant pipeline (I don’t use it much, but I suspect there will be many updates on this)
lower the price of models again?
SearchGPT for everyone would make my day, I know some users were selected to test it during the same time as AVM, so maybe (ok fine, probably not an API thing, but it would certainly create a little bit of excitement)
More image generation functionalities via the API?
Memory? I mean, most of us are using some database for this
“auto-gpt” mode? would certainly make us all spend more tokens, but I feel like there would need to be some IDE plugin to work with it (replit and devin are banking on this, so I guess it would make a lot of devs make happy noises)
GPT generation via the API? If hosted on OpenAI servers to deliver zero friction, that would be amazing, but I don’t think the metrics on GPT is all that strong for this to be entertained
Improvements to abstract more away of the fine tuning process?
simple commands to the api to view models and organization/personal metrics? for example: openai models on the terminal (yes, I’m aware we can do this ourselves with a little bit of code and some aliases, but still, it would be cool to see something like this as a natural evolution of the API’s maturity)
parsing of prompts longer than token limitations? similar to what gpt does (the “continue” button) but native to the API?
lower the costs and increase the rate limit, everyone is always happy if this happens
no breaking changes (as in, no changing from completions to chatcompletions like we had with gpt-4)
streaming on where there is no output streaming
a toggle-able progress bar for when you send a request (right now, when you send a prompt it just hangs for a few seconds before just giving you back the output)
Honestly, the API has progressed so much, I’ve sort of got used to it since gpt-4 (similar to python2 to python3) that I feel like my list is “weak”.
(not happening)