Mira Murati's Thinking Machines Bets on 'Interaction Models'
Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, has announced it's building what it calls 'interaction models' — a new category of AI that goes beyond static question-and-answer exchanges. The company has yet to release a product, but the framing suggests a bet on persistent, dynamic AI interactions rather than one-shot completions.
Original sourceThinking Machines Lab, founded by Mira Murati after her departure from OpenAI in late 2023, broke its relative silence this week by announcing a directional focus on something it's calling 'interaction models.' The term is deliberately distinct from 'language models' or 'foundation models,' signaling that Murati's team is framing its core research thesis around the mechanics of how humans and AI exchange information over time, not just the quality of a single response.
The company has not released a product, an API, pricing, or a technical paper. What exists publicly is a framing document and Murati's reputation — the latter being considerable, given her role overseeing GPT-4, DALL-E, and Whisper at OpenAI. Thinking Machines has raised significant funding and recruited from top AI labs, but the 'interaction models' announcement is still squarely in the concept-and-direction phase rather than the ship-and-iterate phase.
The core idea, as described, is that current AI models are optimized for isolated exchanges — a prompt goes in, a response comes out — but real-world use involves ongoing context, evolving goals, and feedback loops. Interaction models, by Thinking Machines' framing, would be designed from the ground up for this kind of sustained, adaptive engagement. Whether that means a new architecture, a new training paradigm, or a new product layer on top of existing models is not yet clear.
The announcement positions Thinking Machines in a crowded but genuinely unsolved space: making AI useful across a full workflow rather than a single query. It's a real problem. The question is whether 'interaction models' is a novel technical contribution or a rebrand of what retrieval-augmented generation, memory systems, and agent frameworks are already attempting — just with a cleaner name and a better-known founder attached.
Panel Takes
The Skeptic
Reality Check
“'Interaction models' is a term, not a product, not a paper, and not a demo — and terms without implementations are just branding exercises. Retrieval systems, agent memory, multi-turn context windows: the field has been working on sustained AI interaction for years under different names. What kills this in 12 months isn't a competitor, it's the possibility that OpenAI, Anthropic, or Google ships persistent, adaptive interaction natively into their APIs and the 'category' Murati is trying to define gets absorbed before Thinking Machines ships a single thing.”
The Futurist
Big Picture
“The thesis here is falsifiable: static completion models will be replaced by systems that are architected around ongoing human-AI relationships, and whoever builds the right primitives for that layer first will own a structural position in the stack. The dependency that has to hold is that interaction complexity doesn't get solved at the model layer alone — that there's a durable abstraction above completions worth owning. The second-order effect if this works is interesting: shifting value away from raw model capability toward relational continuity, which reshapes who controls the user relationship and makes context a competitive asset, not just a technical parameter.”
The Founder
Business & Market
“The buyer is completely unidentified right now, and that's the whole problem with this announcement — there's no pricing, no product, no clear enterprise wedge, and no indication whether this is a developer platform, a consumer app, or a research lab with aspirations. Murati's pedigree buys time and funding, but it doesn't buy a distribution channel. The moat question is the real one: if 'interaction models' is an architectural insight, it needs to live in a paper or a model weight that others can't easily replicate; if it's a product layer, someone needs to name the buyer before the runway runs out.”
The PM
Product Strategy
“There is no job-to-be-done I can evaluate here because there is no product — just a named direction, which means Thinking Machines is asking users and the market to hire a concept. The 'interaction models' framing is genuinely interesting as a product thesis: users don't want better single answers, they want AI that stays useful across a workflow. But until there's an onboarding flow, a use case, and a moment where a user gets value in under two minutes, this is a roadmap slide dressed up as an announcement.”