You know that feeling when AI news starts blending together? Another benchmark, another half-point bump, another “smarter” model that talks the same as the last one. I was getting numb to it. Then this week happened and one demo actually snapped me back to attention.
Matt Wolfe broke down a fresh drop from Thinking Machines Labs, the company Mira Murati started after leaving OpenAI. According to the curator, their new interaction model is the first thing in months that felt like a real leap, not a polish job. And the kicker? Half the magic isn’t in the answers. It’s in how the model handles the conversation itself.
What’s new
Thinking Machines showed off a preview model built around real-time, full-duplex conversation. No more rigid “you talk, then I talk.” It listens, watches, and acts at the same time.
The twist
Here’s the part Matt highlighted that I keep replaying. The model can interrupt you. On purpose. In one demo, a user casually mentioned taking their 80-year-old parents mountain biking on an active volcano. The model cut in mid-sentence with a flat “stop, that’s deadly.” It picks moments to jump in, and it shuts up when you pause mid-thought. That timing is the unlock.
The demos worth knowing
- 🌐 Real-time translation that speaks over the speaker instead of waiting their turn
- 🐑 Story listening with live counting (it tracked every animal mentioned and gave a tally)
- 🧍 Posture coach that watches your webcam and calls out slouching
- 🔄 Tone rewriter that reframes rude speech into calm professional language in real time
- ⏱️ Built-in time awareness, so you can say “end this chat in 4.5 minutes” and it actually keeps the clock
Pro tips from the curator
- ⚡ The model runs simultaneous tool calls. It can search, browse, and generate UI while still talking, then weave results back in mid-conversation.
- 🎯 If you build voice agents, watch the interruption logic carefully. That’s the design pattern most current voice apps miss.
- 👀 It’s not public yet. A limited research preview is coming “in the coming months,” with wider access later this year. Worth bookmarking the page so you don’t miss the signup window.
Why this matters
Most voice AI today still feels like a walkie-talkie. You press, you wait, it replies. Thinking Machines is treating conversation like an actual two-way stream, where listening, watching, thinking, and speaking all happen at once. That’s closer to how humans talk than anything I’ve seen from the big labs this year.
Matt also packed the same video with Codex going mobile, Anthropic’s new agent view, Krea 2, Google’s wild new AI mouse pointer, and a guy who used Claude to recover five Bitcoin he’d locked himself out of for 11 years. But the Thinking Machines drop is the one I’d actually clear time to watch.
Check the full video for every demo and the rest of the week’s AI rundown.