AI labs outside the frontier seem keen on making sure that a handful of companies don’t get to control AI.
Mira Murati’s Thinking Machines has published a lengthy mission statement laying out why it believes concentrating AI power in a small number of labs is a genuine danger, and what it plans to do about it, framing the entire argument around a single word: distribution. The post stresses not distribution of compute or distribution of revenue, but distribution of knowledge, values, and ownership across as many hands as possible.
The company’s argument starts from an economics lesson rather than an AI one. It leans on Friedrich Hayek’s old case against central planning, the idea that most useful knowledge in the world is tacit and local, sitting inside the heads of chefs, shopkeepers, and specialists who acquired it through years of doing the work. No central planner, however intelligent, can gather that knowledge into one place and act on it better than the people who hold it. Thinking Machines applies the same logic to AI labs. A model trained once in a lab and shipped everywhere, the post argues, ends up flattening the very expertise it’s meant to serve, replacing what makes an organization distinct with “a standard offering.”
It’s a pointed jab at the current shape of the industry, where a handful of labs, OpenAI, Google, Anthropic, and a couple of others, train frontier models and rent access to them through APIs. Thinking Machines wants organizations to own and fine-tune their own weights instead of leasing intelligence from someone else’s server. This is also, not coincidentally, the entire premise of Tinker, the fine-tuning product the company launched in October 2025 to let organizations adapt models to their own industries without building from scratch.
The post’s most interesting turn comes when it moves from knowledge to values. Thinking Machines points out that most flagship models today are trained using their own predecessor to generate training data and a reward signal, meaning each generation inherits the temperament of the last one, and every user of that model inherits the same temperament too. “A single alignment spec suppresses creativity and diversity,” the company writes, arguing that this is closer to a monoculture than to genuine safety. Its proposed fix isn’t a better centralized alignment team. It’s model weights that individual organizations can shape according to their own values, so that no single lab’s judgment becomes the default judgment for everyone using AI.
This isn’t a hypothetical concern in the industry. Geoffrey Hinton has made a related but opposite-facing argument for years, warning that open-sourcing big AI models is akin to selling nuclear weapons at RadioShack, and that keeping powerful models closed and controlled is the safer path. Thinking Machines’ own reasoning runs almost in reverse: leaving all frontier capability in closed, centralized hands is what creates the real risk, because power that answers to no one outside a handful of boardrooms eventually stops caring what anyone outside those boardrooms wants.
There’s a market backdrop to this too. Proprietary labs currently still command roughly 70 percent of token volume on marketplaces like OpenRouter, but that share has been under sustained pressure from open-weight models that now trail the frontier by just four months on independent capability benchmarks. Chinese labs in particular have used openness as a competitive weapon, not an ideological stance, and the gap between what a closed frontier lab can charge for API access and what an open alternative costs to run has been narrowing every quarter. Thinking Machines is essentially betting that the same forces reshaping the open-source model market, cheaper, more distributed, harder to monopolize, will eventually reshape how alignment and customization work too.
Where Thinking Machines differs from the typical open-source pitch is that it isn’t arguing for models to be given away freely to the public. Its case is for organizations to own and adapt weights privately, tailored to their own data and judgment, which is a subtler position than either the fully-closed API model or the fully-open weights release. It sits somewhere between Anthropic’s tightly controlled Claude and China’s open source range (with their own biases), closer in spirit to a world of thousands of specialized, privately-owned models than to one dominant model everyone rents.
Whether this becomes more than a well-written manifesto depends on whether Tinker and whatever follows it can actually make fine-tuning cheap and reliable enough for ordinary companies to bother, and whether Thinking Machines’ own research keeps pace with labs like OpenAI and Anthropic that have vastly more compute at their disposal. Half the company’s founding team, including former CTO Barret Zoph, has already walked back to OpenAI this year, a reminder that ideas about decentralizing AI still depend heavily on the fortunes of one very centralized startup.