Towards AI That Can Actually Interact

May 12, 2026

29 min episode · 2 min read

Episode

29 min

Read time

2 min

Topics

Artificial Intelligence

AI-Generated Summary

Published May 13, 2026

Key Takeaways

✓Interaction Model Architecture: TML's model splits into two parallel components — a real-time interaction layer that continuously perceives and responds to users, and a background model handling reasoning, browsing, and agentic tasks. The two streams run simultaneously, with results woven into conversation naturally, enabling capabilities like live background search without interrupting dialogue.
✓Visual Proactivity Gap: Current commercial real-time APIs cannot respond to visual world changes without audio prompts. TML demonstrates capabilities no existing model handles: time-aware speech triggers, visual-based counting, and visual cue responses. To measure these, TML created two new internal benchmarks — TimeSpeak and QSpeak — because no existing benchmarks captured them.
✓AI Deployment Bottleneck: Enterprise AI adoption stalls not from model capability gaps but from institutional inertia. OpenAI's Deployco addresses this directly, pairing forward-deployed engineers with enterprise clients. The $4B raise at a $10B pre-money valuation, led by TPG with Bain Capital and Advent International, signals that implementation support is now a multi-billion-dollar market.
✓Private Market SPV Risk: Anthropic explicitly voided all unauthorized SPV-based share transfers and named specific platforms selling what it calls fraudulent or legally void instruments. Retail investors holding tokenized crypto receipts claiming Anthropic exposure may own nothing of legal value, as Anthropic's cap table will not recognize these transfers under its transfer restrictions.
✓GUI Moment Framework: TML researcher Claire Birch frames interaction models as AI's equivalent of the graphical user interface — the shift that ended CLI-era computing. Current chat interfaces still reward prompt engineering fluency over natural communication. The target state is users staying fluent in their task rather than learning to communicate like the tool they are using.

What It Covers

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, releases TML Interaction Small, a model built from scratch for continuous real-time interaction using 200-millisecond micro-turns rather than traditional turn-based exchanges, alongside OpenAI's official launch of Deployco, a $4B consulting joint venture with 19 partners.

Key Questions Answered

•Interaction Model Architecture: TML's model splits into two parallel components — a real-time interaction layer that continuously perceives and responds to users, and a background model handling reasoning, browsing, and agentic tasks. The two streams run simultaneously, with results woven into conversation naturally, enabling capabilities like live background search without interrupting dialogue.
•Visual Proactivity Gap: Current commercial real-time APIs cannot respond to visual world changes without audio prompts. TML demonstrates capabilities no existing model handles: time-aware speech triggers, visual-based counting, and visual cue responses. To measure these, TML created two new internal benchmarks — TimeSpeak and QSpeak — because no existing benchmarks captured them.
•AI Deployment Bottleneck: Enterprise AI adoption stalls not from model capability gaps but from institutional inertia. OpenAI's Deployco addresses this directly, pairing forward-deployed engineers with enterprise clients. The $4B raise at a $10B pre-money valuation, led by TPG with Bain Capital and Advent International, signals that implementation support is now a multi-billion-dollar market.
•Private Market SPV Risk: Anthropic explicitly voided all unauthorized SPV-based share transfers and named specific platforms selling what it calls fraudulent or legally void instruments. Retail investors holding tokenized crypto receipts claiming Anthropic exposure may own nothing of legal value, as Anthropic's cap table will not recognize these transfers under its transfer restrictions.
•GUI Moment Framework: TML researcher Claire Birch frames interaction models as AI's equivalent of the graphical user interface — the shift that ended CLI-era computing. Current chat interfaces still reward prompt engineering fluency over natural communication. The target state is users staying fluent in their task rather than learning to communicate like the tool they are using.

Notable Moment

TML researchers had to invent two entirely new benchmarks to measure their model's capabilities because nothing existing could evaluate proactive audio and visual responses — a signal that the capability shift is substantial enough to fall outside the entire current evaluation framework.

Know someone who'd find this useful?