some of the links there point to github repositories that are not public yet! what a time to be alive! this dev day is going 1000% better than my expectations!
3 Likes
1400$ and I’ll be paying cash hahahahah that was funny, great find @MrFriday
4 Likes
Also the pricing page has been updated to reflect model prices that benefit from prompt caching. Seems 50% across the board for eligible models (gpt-4o, gpt-4o-mini, o1-preview, o1-mini).
5 Likes
thank you for sharing and keeping us updated!!
4 Likes
Oh neat! Since Advanced Voice Mode is still not released in the EU, any idea if this is going to be delayed in the same regions as well? The release post makes no mention of it (I just checked and it’s not appearing on the website on my end, but it was just announced so it may just be a rollout thing)
1 Like
I imagine that the EU will be dalayed on this too considering how AVM was, as in, the API might not allow requests based on the information of the account creation and cellphone associated region… not sure how they do it with vpn users that create or use bank accounts in other regions (for both API and ChatGPT)
2 Likes
3 Likes
Thank you so much @anon22939549 for keeping us in the loop. Man. Feels like last year’s DevDay was only some weeks ago. I am EQUALLY excited. Already revving the engines
3 Likes
Thanks for the updates @anon22939549 
3 Likes
seems like the playground is getting updated!
evals just dropped at platform.openai.com/evaluations
edit: no sign of real time in the playground yet, tried a couple of urls in case it was just a missing button and nothing so far, I guess we’ll just have to be patient
6 Likes
turns out the links in the openai-realtime-console are broken and have a typo! I’ve fixed it and created the first pull request of this repository! Hopefully it will be fixed soon!
to access the repository make sure to go to the correct url, github.com/openai/openai-realtime-api-beta
5 Likes
Giving a little first mover advantage to y’all on the ground. 
2 Likes
11:00 PM for me, guess I will have to wait until tomorrow to build what would have been black magic 3 year ago
Life is just so unfair!
It’s very refreshing to have OpenAI ship so much, I was getting worried there for a bit that they had become a safety blogging company!
4 Likes
By this time tomorrow, “caching strategy” will be the new black.
https://platform.openai.com/docs/guides/prompt-caching/best-practices
(Also, thank you very much Elmstedt!)
3 Likes
Distillation is cooooool!
It looks like they did one of their low-key understated releases: metadata is now searchable for completions!
2 Likes
Distillation is going to be huge!
I’d love to see them add synthetic data generation to the interface…
Use something like CriticGPT on the responses, then given the critique, generate n alternate responses, and keeping the top k.
Get those small models outperforming the larger ones in narrow tasks.
3 Likes
It looks like that’s exactly what the new Evals, which are accompanying Distillation, do!
https://platform.openai.com/docs/guides/evals
Then there’s a way of using those evaluated completions in the Playground to fine tune a model. Now that you bring it up, you’re right, this must be how they created CritiqueGPT in the first place.
1 Like
Another issue on the openai-realtime-console repository is that you’ll get this error when running on firefox: AudioNodes from AudioContexts with different sample-rate is currently not supported
the best way to go about this for now is to use another browser. I’ve created a pull request that fixes this, but lets first wait for the OpenAI staff to review my code.
Edit: Had to resample the audio PCM data as well, hopefully it’ll get reviewed soon. Also, the realtime endpoint doesn’t seem to have been released yet as some suggested it could be at 3pm SF local time.
3 Likes