Mira Murati’s AI Lab Bets on Models Users Can Train

Thinking Machines Lab has drawn a clear line around its AI strategy: users should be able to train models, not just prompt them.

The lab, founded by former OpenAI chief technology officer Mira Murati, published the position on July 10. It says it will build frontier models, tools for changing model weights, live multimodal interfaces and open research.

What Thinking Machines announced

In its new strategy statement, Thinking Machines says most AI is trained in a few places and then frozen. The company argues that this limits how much people and organizations can shape a model with their own knowledge.

Its plan has four parts:

• Train strong models that can handle multiple forms of input and be customized.
• Give users tools to train model weights for their own needs.
• Build interfaces that can take continuous human feedback, including voice and video.
• Publish research so more people can understand and shape the technology.

This is a strategy statement, not a new model release. Thinking Machines did not announce a launch date, benchmark score or price for a new frontier model.

Why the model weights matter

Most AI products let users change behavior through a prompt, a system instruction or retrieved documents. Thinking Machines is making a larger bet: some knowledge and preferences should be learned inside the model itself.

The company already offers Tinker, an API for fine-tuning open-weight models. Tinker handles the distributed computing while developers control the data and training loop. Users can also download saved model weights.

That makes the new statement more concrete than a general promise to keep humans involved. Thinking Machines is saying that custom training and user control will be central to its product line.

The hard part is whether this can become simple, safe and economical outside research teams. Training can fail because of weak data, poor evaluation or the wrong objective. Changing model weights also creates more safety and quality work than changing a prompt.

The business bet

The strategy puts Thinking Machines on a different path from a single general model sold in the same form to every customer. The lab argues that organizations should own and tailor AI around knowledge that competitors do not have.

Murati described a similar direction in a May interview with WIRED. The company previewed interaction models that process live audio and video, but those models were not publicly released at the time.

For developers, the practical question is whether training APIs become as normal as calling an inference API. If they do, AI infrastructure spending could move beyond serving answers and toward continuously adapting smaller, specialized models.

Arkolith's own guide to agent-native APIs explains the other side of that shift: models still need reliable tools and source-linked data when facts change. Our guide to how AI agents cite sources shows why that provenance remains necessary. Custom weights can preserve behavior and expertise, but they do not replace live data.

What to watch next

The next proof will be a released Thinking Machines model or interface that shows how human feedback changes the product in real time. Pricing, model ownership, evaluation and safeguards will matter as much as a demo.

Until then, the July 10 statement is a declared product direction. It tells developers and investors what the lab intends to compete on, but it does not yet prove that user-trained frontier AI will work at scale.

Arkolith provides source-linked public information for educational and informational use. This article is not investment advice.