Models & Hardware
Full-Duplex AI Listens While It Talks
OpenAI, Thinking Machines and China’s Telos are building models that take a correction mid-answer.
AI that hears you change your mind halfway through.
Full-duplex models keep listening while they speak, so a correction can change what comes next, @kimmonismus writes in a thread. A real-time agent can take in “Actually, use the other file” while it is already responding or working, with no wait for a turn.
| Who | What @kimmonismus reports |
|---|---|
| Thinking Machines Lab | Research preview processes audio, video and text in 200ms chunks; a background model handles deeper work |
| OpenAI | GPT-Live: continuous voice, with reasoning and tools delegated to a background model |
| Nuance Labs | Face-to-face full-duplex audiovisual model; $50M Series A led by Lightspeed, with Accel |
| OpenBMB | MiniCPM-o 4.5, a 9B open model designed to see, listen and speak simultaneously |
| Telos | Omni-modal, full-duplex models in China; seed backing from HongShan and Hillhouse |
He sees room for Telos to compete on how well those pieces work together. A tutor needs to notice confusion while explaining, and a coding assistant needs to take a correction mid-task, he writes.
Sources
- @kimmonismus on X · Oct 22, 9:11 AM PT
- @kimmonismus on X · Oct 22, 9:11 AM PT
- @kimmonismus on X · Oct 22, 9:11 AM PT
- @kimmonismus on X · Oct 22, 9:11 AM PT
- @kimmonismus on X · Oct 22, 9:11 AM PT
- @kimmonismus on X · Oct 22, 9:12 AM PT