mira murati was openai’s cto she shipped chatgpt, dall-e, codex, and sora.
Inseptember 2024 she walked out with 30 of openai’s best people and raised $2 billion,the largest seed round in tech history
then she went quiet.
Two days ago we finally saw what she’s been building,it’s called interaction models and it doesn’t feel like chatgpt at all.
Every ai you’ve ever used works the same way so you talk and it listens and then it responds, you listen.its called turn based texting, thinking machines built something that works like a phone call so it processes your input and generates responses simultaneously. When you pause to think it knows not to interrupt while you r still asking a question it’s already searching in the background. You get the answer in real time like talking to a friend who happens to know everything.
every chatgpt style Ai has a tell, the 2 second pause before it replies and that pause was the giveaway that you’re talking to a machine so
mira just killed the pause. It has 0.40 second response time which is faster than natural human conversation
.The technical architecture is wild, a 276 billion parameter model split into two pieces so a lightweight interaction model that stays live with you the entire time, listening, backchanneling and reacting to your face through video another one as a separate background model that handles heavy reasoning and tool calls then injects results at exactly the right moment.
it can interrupt you mid sentence when you say something wrong and it can speak at the same time as you for live translation. It watches your screen and reacts to what it sees without you asking. Gpt realtime scored 2.9% on cued response timing in their benchmarks, mira’s model scored 82%
every lab has been racing to make ai smarter,mira bet on making ai present and i think she might be right that presence is the actual moat.
So bottleneck isn’t intelligence anymore, it’s the friction of the interaction.
The same shift already happened in creative tools,the models that generate video and do face swaps are all good enough now,the tools that won (magic hour, kling, the platforms people actually use daily) didn’t win because they had the smartest model.They won because the workflow felt effortless(presence over intelligence) and experience over capability
openai didn’t lose a cto in 2024, they lost the person who understood that how we interact with ai matters as much as how smart the ai is.
i’d bet money this forces openai to completely rebuild their realtime api within six months
what’s everyone’s take on whether presence beats intelligence as the next moat in ai
[link] [comments]