Real-time AI experiences are now solved AND open sourced. We worked closely with OpenAI to build the best framework for it, combining GPT-Live-1, LiveAvatar, and HyperFrames Check out the demos below and show us what you build with it.

@OpenAI @TryLiveAvatar @HyperFrames_ I need that poker coach

@OpenAI @TryLiveAvatar @HyperFrames_ We all need that poker coach

@OpenAI @TryLiveAvatar @HyperFrames_ Browser only?

@OpenAI @TryLiveAvatar @HyperFrames_ Almost every AI product this year has been a text box. Three demos here and not one of them is.

@OpenAI @TryLiveAvatar @HyperFrames_ now live avatar get's a graphics/animation engine!

@OpenAI @TryLiveAvatar @HyperFrames_

@OpenAI @TryLiveAvatar @HyperFrames_ dealerbot

@OpenAI @TryLiveAvatar @HyperFrames_ dealerbot is genuinely a better name than what we went with

@OpenAI @TryLiveAvatar @HyperFrames_ Amazing demo

@OpenAI @TryLiveAvatar @HyperFrames_ Cracked demo

@OpenAI @TryLiveAvatar @HyperFrames_ love it

@OpenAI @TryLiveAvatar @HyperFrames_

@OpenAI @TryLiveAvatar @HyperFrames_ @grok bullet point the innovation here

@OpenAI @TryLiveAvatar @HyperFrames_ Real-time AI is moving from demos to usable products. Open-sourcing the framework should accelerate experimentation a lot.

@OpenAI @TryLiveAvatar @HyperFrames_ that's the direction we're betting on too

@OpenAI @TryLiveAvatar @HyperFrames_ oh damn this is slick! Really wanna test this out 🔥

@OpenAI @TryLiveAvatar @HyperFrames_ Go for it. MIT, and the demo source is in there rather than just the framework.

@OpenAI @TryLiveAvatar @HyperFrames_ The real test is its speed and accuracy in real-time spot analysis—I'll decide after that.

@OpenAI @TryLiveAvatar @HyperFrames_ 'solved' is doing a lot of work there. real time avatar stacks live or die on the end to end budget under jitter, and lip sync drifts when audio arrives in chunks the renderer hasn't seen yet. you predict visemes or you buffer. how does mouth to ear hold up on a bad network?

@OpenAI @TryLiveAvatar @HyperFrames_ Fair hit on 'solved'. And you're pointing at the exact thing we wrote down: transcripts arrive seconds ahead of the avatar's voice, so anything that must land on a heard word needs a playback-position estimate. It's an open item in docs/ARCHITECTURE.md.

@OpenAI @TryLiveAvatar @HyperFrames_ We need to see a live test, not just a demo video.

@OpenAI @TryLiveAvatar @HyperFrames_ Then don't take the video's word for it. Source is up, run it yourself.

@OpenAI @TryLiveAvatar @HyperFrames_ What kind of round-trip latency are you seeing end-to-end across the GPT-Live-1 and LiveAvatar pipeline?

@OpenAI @TryLiveAvatar @HyperFrames_ We're not publishing a number. What we have documented is a real gap: transcripts arrive seconds ahead of the avatar's voice, so timed visuals need a playback-position estimate we haven't built. It's in docs/ARCHITECTURE.md.

@OpenAI @TryLiveAvatar @HyperFrames_ Are you planning to sync the visuals against client audio playhead events, or does that playback estimate need to live server-side?

@OpenAI @TryLiveAvatar @HyperFrames_ Does the session allow visual sign language interaction?

@OpenAI @TryLiveAvatar @HyperFrames_ Do the sessions store to memory somewhere? How can we make sure that it doesn't always try and teach the same thing?

@OpenAI @TryLiveAvatar @HyperFrames_ Session only. Every term card shown is recorded server side and deduped by term, in teaching order, so it won't re-teach inside a session. Nothing survives the session ending though.

@OpenAI @TryLiveAvatar @HyperFrames_ Random question.. but is the generated screen readable by a screen reader, or is it pixels?

@OpenAI @TryLiveAvatar @HyperFrames_ Can I get some code to try it out

@OpenAI @TryLiveAvatar @HyperFrames_ All of it. MIT licensed.

@OpenAI @TryLiveAvatar @HyperFrames_ That's seriously cool. Open source is the way to go for this kind of tech. Can't wait to see what people create with it.

@OpenAI @TryLiveAvatar @HyperFrames_ Are the three demos in the repo too, or just the framework?

@OpenAI @TryLiveAvatar @HyperFrames_ Framework plus one demo. Out of the box it's the Japanese tutor. The poker coach and support agent were the same pattern with a different persona and tools, and they didn't ship.

@OpenAI @TryLiveAvatar @HyperFrames_ Now order my incredibly long coffee order with all the creams and sugars in Japanese 😂

@OpenAI @TryLiveAvatar @HyperFrames_ it would absolutely make you say the whole thing. twice 😄

@OpenAI @TryLiveAvatar @HyperFrames_ The GPT-Live + LiveAvatar + HyperFrames combo feels like a solid stack for building voice-first AI experiences. Excited to see what developers ship.

@OpenAI @TryLiveAvatar @HyperFrames_ voice-first is exactly how we've been thinking about it. excited to see what lands

@OpenAI @TryLiveAvatar @HyperFrames_ Can I use sign language directly on camera with the coach?

@OpenAI @TryLiveAvatar @HyperFrames_ real-time demos are solved, but interruption recovery still breaks on weak mobile networks

@OpenAI @TryLiveAvatar @HyperFrames_ Does the coach offer visual support for sign language?

@OpenAI @TryLiveAvatar @HyperFrames_ Run npm install and grab this bread!

@OpenAI @TryLiveAvatar @HyperFrames_ This makes me want to build something. Great demo!

@OpenAI @TryLiveAvatar @HyperFrames_ I don't think people appreciate just how crazy this is. Real time motion graphics in near sync with speech, expression and context states is quite the feat. Can't wait to try this.

@OpenAI @TryLiveAvatar @HyperFrames_ How does session memory work here? I’d want it to remember what I’ve already learned and adapt future sessions instead of starting from the same place every time.

@OpenAI @TryLiveAvatar @HyperFrames_ sheesh

@OpenAI @TryLiveAvatar @HyperFrames_ So curious if there's a learning curve to actually deploying this.

@OpenAI @TryLiveAvatar @HyperFrames_ There is, and the README says so rather than pretending otherwise. It's a starter and not a deployment. Auth, rate limiting and reconnect paths are all on you.