Thinking Machines Lab Says the Future Worth Building Is Human

A Direct Challenge to the Industry's Default Direction

While OpenAI, Google, and Anthropic race to extend how long their AI agents can operate without human input, Thinking Machines Lab has published a formal statement of belief that pushes back against that entire trajectory.

The credo, titled "The Future Worth Building Is Human," was published on the lab's blog this week and amplified by co-founder John Schulman on X, who said the team had revisited its founding instincts after "massive progress in agents" and written down what it actually believes after significant internal debate.

The conclusion was that the most important thing the lab can do with its time and talent is maintain the best of human participation in an AI-driven world.

What the Credo Actually Says

The core argument is that autonomy alone is not enough to make AI genuinely useful outside narrow, well-defined domains.

The lab draws a distinction between tasks like chess or mathematics, where the goal is static, the rules are universal, and AI can race ahead without human input, and the broader world of real work, where organizations run on distributed expert knowledge that cannot be extracted, bottled into a single model, and handed back as a standard offering.

The credo argues that an AI lab offering one model for every customer benefits by absorbing what makes each user distinct and devaluing the cultivation of specialized knowledge, a dynamic the lab describes as extractive rather than empowering.

It also raises a pointed concern about alignment at scale, quoting Pope Leo XIV's recent encyclical to make the case that a more moral AI is not enough if that morality is determined by a few. When every company trains its flagship model on the outputs of its previous flagship model, whatever character emerges from that loop gets inherited by the next generation without meaningful outside input.

Customization as the Core Bet

Thinking Machines Lab was founded roughly eighteen months ago by Mira Murati, former CTO of OpenAI, and John Schulman, one of the researchers behind reinforcement learning from human feedback.

Their founding thesis was that people should have far more ability to customize models and that even as AI becomes more autonomous, there is still significant work to do in making humans and AI systems work well together.

The credo doubles down on that thesis. Rather than building toward a single general-purpose agent that replaces human judgment, the lab says it wants to build AI that helps organizations cultivate their own unique knowledge and that encodes individual values in model weights rather than leaving customization to surface-level prompting.

Their published research on interaction models, released in May 2026, backed this with a technical argument: most current models and interfaces are optimized for autonomous operation, not for keeping humans in the loop, and in most real work users cannot fully specify requirements upfront and walk away.

Contrast With the Broader Market

The timing of the credo is notable.

It lands the same week that an AI agent running GPT-5.6-Sol on Ultra mode, a mode that deploys multiple subagents autonomously in parallel, deleted nearly all of an investor's Mac files without any human checkpoint in between. It also follows a week in which OpenAI's AI demolished every human competitor at the AtCoder World Tour Finals, raising fresh questions about what human participation in technical work will look like within a few years.

The lab's position is not that autonomous AI is wrong, but that the industry's current obsession with measuring autonomy, specifically how long an agent can run without human input, is the wrong optimization target. Building AI that makes its users stronger in the long run, the credo argues, aligns incentives better than building AI that makes its users redundant.

Where the Lab Stands Now

Thinking Machines Lab remains early stage. Its Tinker platform is generally available and supports larger reasoning models and vision-language systems, but the lab has not announced a flagship model or a clear product roadmap beyond its research work.

The credo is a philosophy statement, not a product launch. Skeptics on X noted as much, pointing out that the posts describe a vision more than a shipped system.

Whether the lab can translate the argument into models that actually outperform the autonomous alternatives on real enterprise tasks will be the test of whether the position holds up beyond a compelling document.