On July 10, 2026, Mira Murati's Thinking Machines Lab published "The Future Worth Building Is Human" — a manifesto that lays out the worldview behind the company's mission. Within 48 hours it had become the most discussed lab statement of the year, and for good reason: it is less a marketing document than a technical and economic argument against the way frontier AI is currently built.
The core claim is blunt. Most AI in use today is trained in a handful of places and then frozen. It is not shaped by the people it serves, and it does not learn much from the work they do together. Thinking Machines' bet is the opposite: AI that is distributed, customizable, and continuously fine-tuned by its users — with values encoded in model weights that organizations actually own.
This article breaks down the manifesto's argument, the four technical directions behind it, and what the thesis means in practice for businesses deciding how to adopt AI in 2026.
The economics: why distributed knowledge needs distributed AI
The manifesto's intellectual foundation comes from two twentieth-century thinkers rarely cited in AI papers: Michael Polanyi and Friedrich Hayek.
Polanyi's observation, from The Tacit Dimension (1966), is that most productive know-how is tacit — a chef crafting a recipe or a shopkeeper rearranging prices applies knowledge that is not legible to outsiders and cannot be written into a database. Hayek's argument, from The Use of Knowledge in Society (1945), is that central planning fails not because planners lack intelligence, but because productive knowledge is local, fleeting, and held privately by the people doing the work.
Thinking Machines applies both directly to AI: attempting to aggregate the world's knowledge into one centralized intelligence faces the same structural problem as a planned economy. A single frozen model, however capable, cannot capture knowledge that is generated continuously inside each organization.
The manifesto concedes there are domains where centralized intelligence alone wins — chess and mathematics, where goals are static and expressible and nothing about the domain is hidden. Chess engines train on pure self-play; frontier models now prove long-standing theorems on their own. But outside those special cases, the report argues, intelligence alone is not enough. For AI to benefit from distributed knowledge, it must itself be distributed.
The Toyota example anchors the point. In 2014, the company famous for automated plants deliberately brought expert craftsmen back onto the line. As Mitsuru Kawai, who led the initiative, put it: "To be the master of the machine, you have to have the knowledge and the skills to teach the machine." Knowledge production and applied intelligence lift each other; they are not substitutes.
Four technical directions
The manifesto translates its philosophy into four concrete workstreams:
- Train strong models with native multimodal interaction and customizability. The lab is explicit that human judgment needs to shape models that compete on the frontier — a weak but customizable model does not extend anyone's will.
- Build tools for ownership, including the ability to train model weights directly. This is Tinker, the lab's fine-tuning product, which went generally available with support for trillion-parameter models like Kimi K2 Thinking and vision input through Qwen3-VL — research-grade customization through a simple Python training loop, no distributed training stack required.
- Develop interaction models that widen the communication channel between human and machine, so personal judgment can continuously influence AI's work.
- Publish research, because the power to shape AI requires understanding how it is made.
Values belong in weights, not prompts
The most technically interesting section is the argument about where alignment should live. Today, every lab trains its next flagship model using the previous flagship to generate training data and reward signals. Whatever character emerges from that loop, everyone gets the same one — and each generation inherits its parent's traits, raised on its parent's outputs and judged by its parent's tastes.
The manifesto argues that a single alignment specification suppresses diversity the same way a planned economy suppresses market signals. Its alternative: organizations and individuals should encode their values directly into model weights through fine-tuning — producing portable artifacts (in practice, LoRA adapters) that the user keeps and owns.
The reason prompts are insufficient is subtle but well-observed by anyone who has deployed LLMs in production: a system prompt changes surface behavior while deeper habits remain fixed. And a centralized model malleable enough to deeply change via prompt becomes a security liability, vulnerable to repeated adversarial attacks. Weight-level customization sidesteps both problems — the deep behavior actually changes, and the change is authenticated by whoever controls the fine-tuning process.
The endgame the lab envisions is what it calls decentralized alignment: safety as a property of an ecosystem of AIs raised in different places, disagreeing, competing, and learning from each other — rather than one central spec that becomes, in the manifesto's words, "a locus of power to be captured." The document even quotes Pope Leo XIV's Magnifica Humanitas (2026): "A more moral AI is not enough if that morality is determined by a few."
The benchmark critique: measuring the wrong thing
The manifesto takes direct aim at the industry's favorite progress metric: the autonomous task time horizon tracked in METR's charts, which measures how long a software task a model can complete on its own. Thinking Machines expects progress on that curve to continue — but argues it measures only what AI does alone, not what humans and machines accomplish together.
This is more than a philosophical quibble. Incentives follow benchmarks. A lab optimizing for autonomous horizons benefits when customers outsource entire workflows to the model — absorbing what makes each customer distinct. A lab optimizing for customization and collaboration benefits when customers leverage their unique advantages. The manifesto is explicit that this second incentive structure is the one Thinking Machines chose, and that measuring human-AI joint performance is something every organization ultimately has to do for itself.
What this means for your AI strategy
Strip away the philosophy and the manifesto makes three practical claims any business evaluating AI should weigh:
Owning fine-tuned weights is becoming a real option. Until recently, weight-level customization of frontier-class models required an ML platform team. Tools like Tinker — and the broader open-weight ecosystem covered in our guide to fine-tuning LLMs with LoRA and QLoRA — collapse that cost. If your competitive advantage lives in tacit organizational knowledge, renting an identical model to your competitors' is a strategic ceiling.
Vendor concentration is an alignment risk, not just an availability risk. We have argued before that businesses need multi-model fallback strategies for resilience. The manifesto adds a second reason: a model shaped in one place encodes the values of its owner, not yours. For MENA organizations especially, this echoes the case for sovereign and Arabic-native AI models — locality of values is not a nice-to-have when your market's norms differ from a lab's defaults in San Francisco.
Cultivate knowledge, don't extract it. The sharpest line in the manifesto is the distinction between AI that helps an organization cultivate its unique knowledge and AI that extracts a snapshot of it and replaces it with a standard offering. Applied practically: deployments where your experts continuously correct, teach, and fine-tune the system compound your advantage. Deployments that merely replace your experts' output freeze it — and hand the compounding to your vendor.
The honest counterpoints
The manifesto is a company's worldview, and it conveniently aligns with that company's product. Three caveats are worth keeping in view.
First, distributed fine-tuning has real costs: every custom model is a model to evaluate, monitor, and secure. Most small teams struggle to operate one model well, let alone a fleet of bespoke ones. Second, decentralized alignment cuts both ways — the same tools that let a hospital encode its ethics let a bad actor encode theirs, a tension the manifesto acknowledges via von Neumann's 1955 observation that the useful and harmful aspects of technology can never quite be separated. Third, frozen centralized models keep getting better fast enough that "just use the frontier API" remains the right call for many workloads.
But as a directional bet, the manifesto names something real. The gap between what generic models offer and what your organization actually knows is precisely where your margin lives. The labs that treat that gap as something to absorb, and the labs that treat it as something to amplify, are now pursuing visibly different futures.
Conclusion
"The Future Worth Building Is Human" is the most intellectually serious lab manifesto in recent memory — grounded in Hayek and Polanyi rather than vague safety language, and backed by a coherent product thesis: strong models, weight-level ownership, wider human-machine interfaces, published research.
Whether Thinking Machines wins its bet or not, the questions it raises are the right ones for 2026: Who owns the weights your business runs on? Whose values are encoded in them? And is your AI making your organization's knowledge compound — or quietly extracting it?
If you are working through those questions for your own organization, talk to our team — helping businesses build AI that amplifies their advantage rather than flattening it is exactly what we do.
Sources: Thinking Machines Lab — The Future Worth Building Is Human, Mira Murati's announcement, MarkTechPost analysis