Date: 2026-05-17 (Sunday)

Pillar: Model release + spicy take

Topic: Mira Murati's Thinking Machines Lab releases TML-Interaction-Small, a 276B MoE that listens while it talks. 0.40s turn-taking latency vs 1.18s for GPT-Realtime-2.0.

Why this story today

• Voice and real-time agents are the fastest-growing LLM app of 2026 (continuation of yesterday's Vapi $50M narrative)
• Mira Murati is a celebrity name in AI — automatic attention
• Three eye-popping numbers to anchor the post: 0.40s response, 276B MoE, beats GPT-Realtime by 3x
• Architecture insight worth teaching: split into interaction model + background reasoning model — a design pattern AI founders can steal
• Spicy take: standard VAD (voice activity detection) just got killed. The interaction layer is its own moat
• Continues last week's thesis ("the layer above the model wins"), refines it ("the interaction layer is the new frontier")

Source list

• Thinking Machines Lab blog — Interaction Models
• TechCrunch — Thinking Machines wants AI that listens while it talks
• VentureBeat — Near-realtime AI voice and video
• The AI Insider — Full-duplex AI responds in 0.4 seconds
• MarkTechPost — Native multimodal architecture for real-time human-AI collaboration
• Latent Space — TML-Interaction-Small 276B-A12B kills standard VAD
• Semafor — Thinking Machines previews interaction models

X / Twitter post

Mira Murati's Thinking Machines just dropped TML-Interaction-Small.

276B MoE, 12B active. Responds in 0.40 seconds.
GPT-Realtime-2.0 takes 1.18.

It doesn't wait for you to finish. It listens while it talks.

3 reasons every AI founder should care:

1. Voice agents were already 2026's fastest-growing LLM app. The model layer just leveled up.

2. Standard VAD is dead. The new model perceives in 200ms micro-turns. Audio, video, text, all at once.

3. Murati split the system: one model live with the user, one running reasoning and tool use async. The agent stack just got a new default design pattern.

The "model doesn't matter" thesis was always too clean. The interaction layer is the new frontier.

What's the first product you'd rebuild on a sub-half-second interaction loop?

@rahilnjain