Peter Bakkum

I work on Multimodal API @openai, for fun:

[Peter Bakkum @pbbakkum]

https://t.co/NM21RApvGv

[Peter Bakkum @pbbakkum]

In retrospect it’s hard to believe we wrote all that code by hand with such excruciating detail

[Peter Bakkum @pbbakkum]

the actual worst thing in the entire universe is that when you're on airpods on a macbook and open the mic channel the output audio cuts out for ~2s and switches from 48khz to 24khz

[Pauline P. Narvas @paw_lean]

The real unlock of GPT Live is being corrected in real time and learning in real time that I’m actually terrible at languages lol

Here’s me testing it in Tagalog and Greek 🇵🇭🇬🇷

[Peter Bakkum @pbbakkum]

@undefined_void @kevinwhinnery realtime-2.1 in the api now, live-1 coming soon

[Peter Bakkum @pbbakkum]

GPT Live is a full duplex audio model launching today in ChatGPT -- it's coming to API soon, sign up here to be notified
https://t.co/bnjPNNm3no

[Justin Uberti @juberti]

Today, we're launching our third-gen voice model and architecture, GPT-Live. GPT-Live is a full-duplex model with built-in async delegation, which allows it to deliver incredibly natural conversation along with the intelligence you expect from ChatGPT.
https://t.co/6P71MwTfUa

[OpenAI @OpenAI]

Introducing GPT-Live, a new generation of voice models for natural human-AI interaction.

Rolling out in ChatGPT starting today.

You’ll want to turn the sound on for this one.

[Peter Bakkum @pbbakkum]

The 2.1 is our best speech-to-speech model so far, and 2.1-mini is the first small audio model we've had in a while -- please try them and send feedback

[Peter Bakkum @pbbakkum]

Audio model releases today:
gpt-realtime-2.1 -- An incremental update on realtime-2 that should significantly improve handling of alphanumeric details in calls (e.g. phone numbers, emails)
gpt-realtime-2.1-mini -- Small version that still has reasoning!

[OpenAI Developers @OpenAIDevs]

GPT-Realtime-2.1-mini is now available in the API, bringing reasoning and tool use to our Realtime mini lineup at the same cost as GPT-Realtime-mini.

[jason @jxnlco]

Computah! Activate Firewall!

with gpt-realtime-2 you can in context prompt your wake words, reasoning, and build some silly games

check out me playing a game simon says...

spoiler: it beat me

[Peter Bakkum @pbbakkum]

We've been rolling out a TTFT improvement for all gpt-realtime models -- for long sessions the p95 TTFT improvement is really significant (this chart splits out requests with >= 20 turns)

[Peter Bakkum @pbbakkum]

@jxmnop Just to frame this — everything is gated and most large customers are ZDR, meaning there literally is no query retained to look at. I often ask people to repro on a separate debugging account.

[Greg Brockman @gdb]

GPT-Realtime-2 is something new

[Peter Bakkum @pbbakkum]

@vinhocent Tri where do you work again?

[OpenAI Developers @OpenAIDevs]

Here are the experiences developers are building with the Realtime API:

[Peter Bakkum @pbbakkum]

A few demo videos from the OpenAI Voice Hack Night last week, these are great examples of what you can do with gpt-realtime-2

[OpenAI Developers @OpenAIDevs]

🧵 Our Voice Hack Night finalists are here.

4 projects. 6 hours. Realtime voice agents in real-world builds.

Now it’s your turn to vote for your favorite. We’ll announce the winner on Monday.

https://t.co/tLllqhl9Tj

[Kai @hqmank]

Downloaded Clicky and I can't stop playing with it. I just talk and my Mac does stuff, hands-free, instant response.

This is what Siri should've been from day one.

Peeked at the architecture too:
voice in → GPT-Realtime 2.0 → local tool call (shell_exec) → runs locally. Screen reading only when needed.

There's free credit to start, so just try it.