Murati Says Next-Generation Models Will Always Be Listening; Commenters Do the Math First

In a video, former OpenAI CTO Mira Murati mentioned that the models people use today are basically turn-based: you speak, it replies, then it goes into thinking, and while it is thinking it is deaf and blind. Her judgment is that the next generation of models will be listening all the time.

The term “turn-based” is apt. When you use ChatGPT or Claude, after you hit send the interface spins, and during that time the model cannot receive any new input. If you change your mind or add information, you have to wait until it finishes speaking. The model is offline while it reasons; it is just that this offline period is very short—so short that it is easy to overlook.

“Always listening” changes where interaction starts. No wake word, no pressing send; the audio stream flows continuously into the model. Conversation is no longer taking turns speaking; it is more like an always-on listener in the room. It may understand your intent before you finish speaking, and it may offer a hint while you are silent.

Murati left OpenAI in 2024 and later founded Thinking Machines Lab. Her judgment about the shape of interaction is worth hearing, but the comment section offers another perspective. Some first question whether her appearance suits being a CTO, and by the way conclude that they have no chance against Anthropic. Another reply asks: “never stop listening” sounds great in a demo, but what will the bill look like?

The second question is more concrete. Continuous listening means the model has to process an audio stream for long periods, and the compute, bandwidth, and electricity for real-time inference are all billed by time. The cost of turn-based interaction is discrete: you pay once per question. The cost of always-on presence is continuous: even if the user says nothing, the model is still running. In a demo, “always listening” is romantic; finance sees a different spreadsheet.

There is another issue that gets glossed over: privacy boundaries. If the model is always listening, when does it count as starting, when does it count as ending, and who decides which segments enter inference? Turn-based interaction at least has a clear send action; always-on presence hands the boundary over to product design.

It is worth noting that turn-based interaction is not necessarily the essence of intelligence; it is more like an engineering compromise: inference is too expensive and latency too high, so interaction gets cut into segments. As costs fall and streaming inference becomes feasible, models will naturally move toward continuous presence. Murati is talking about interaction, but behind it are changes in compute and product form together.

As for the comment that talks about appearance first, it only shows how easily attention in technical discussion can be led astray. The other reply doing the math is actually closer to the real question.