Mira Murati’s Thinking Machines Lab released a report arguing that current AI models are too static to serve human needs. The document asserts that most systems are trained in isolated locations and then locked down, removing the ability for the people who use them to shape the technology. The researchers propose a shift toward distributed systems where model weights are customizable and owned by the users.
The technical approach
The lab outlines four specific areas for development. First, they aim to train powerful models capable of multimodal interaction that allow for custom adjustments. Second, the proposal calls for tools enabling individuals to fine-tune and train model weights directly. Third, the team plans to develop interfaces that expand the communication channel between human and machine. Finally, the group intends to publish research to ensure more engineers understand the construction of these models. Together, these steps move technical knowledge and alignment closer to the end user.
Interactive explainer
The report explores how distributed ownership changes the flow of data and control compared to centralised systems.
Live micro-turns keep silence, overlap, and interruption inside the model’s context — details that turn-based prompting drops.
What it means
For people making things, this means moving away from a single, frozen model that everyone must accept as is. Instead, the focus shifts to systems where an organisation or individual can own their specific version of an AI. This allows values to be encoded directly into the model weights rather than relying solely on prompts to steer behaviour. Knowledge generation becomes a two-way street, with local sites cultivating their own models while retaining authorship.
