Open Models and the Architecture of Progress
The July 24 letter supporting open-weight AI makes a compelling case that advanced models should be available beyond the institutions that create them. Greater access strengthens competition, reduces dependence on a few providers and allows more organizations to adapt AI to their own needs.
Openness as an Engine of Discovery
There is another reason to support open models, one that may prove more consequential still.
The institutions and societies that have led scientific discovery have rarely done so by predicting where the next breakthrough would come from. Rather, they succeeded by creating conditions in which more people could participate, knowledge could circulate and ideas could be tested on their merits. Indeed, Linux Foundation, one of the signatories of the letter, has operated on this premise for decades.
No single institution can consistently identify the most promising researcher, architecture or scientific question. Progress often comes from people outside established centers of authority, pursuing ideas that initially look unconventional or commercially uncertain. Even the great directed programs of the twentieth century, from Bell Labs to Apollo, drew on science that circulated openly: they concentrated resources, not permission to think. An open ecosystem preserves the possibility that unconventional ideas can be tested without first receiving permission from those already in power.
Empowering Independent Scrutiny
Open models allow researchers, engineers and entrepreneurs to examine how systems behave, challenge their assumptions and adapt them for purposes their original developers never considered. They let a small team test a new method, a scientific community build around its own knowledge and independent experts evaluate the claims of the largest laboratories.
In this sense, openness creates option value. Most experiments will not produce a breakthrough, and it is difficult to know beforehand which one will. Broad participation increases the number and diversity of ideas the system can explore.
Openness also strengthens correction. For a technology as powerful as AI, accountability cannot depend entirely on the judgment of the institutions building it. Independent researchers must be able to test performance, investigate vulnerabilities, reproduce findings and develop better evaluation methods. Communities affected by these systems need meaningful ways to examine their limitations.
Not every model, dataset or technical detail should be released without restriction. Open weights cannot be recalled once published, and privacy, consent, security and demonstrated risks require careful judgment. But risk is not an argument against independent scrutiny. It is a reason to build the institutions and standards that make such scrutiny possible: a trust layer of provenance, evaluation and accountability around open systems.
Building the AI Commons
The open-source software movement offers a useful lesson. Its lasting value did not come merely from making code available. It came from the ecosystem that grew around it: maintainers, licenses, standards, security practices and communities capable of improving shared infrastructure over time. Access mattered because it enabled contribution, verification and cumulative progress. AI will require a similar foundation.
Open weights are an important starting point, but they are the output of a much longer process. Before a model is released, decisions have already been made about which data to use, which capabilities to optimize, which risks to accept and which problems deserve attention. Those upstream decisions shape everything that follows.
A durable technology commons must therefore extend beyond model releases. Open weights without access to compute still gate participation: a model anyone may download but few can run is open only in name. The commons should include research tools and compute, responsibly governed data, reproducible methods, independent evaluation and sustained support for the people who maintain shared infrastructure.
It must also leave room for work whose value no market can yet capture, such as research on neglected diseases, low-resource languages and scientific problems with long development cycles. The markets that closed systems create direct resources toward identifiable demand, but they are unreliable at valuing discoveries whose benefits emerge years later.
The central question is therefore not only how widely today’s models can be used, but also how broadly the future direction of AI can be shaped. That is, can talented people contribute regardless of institutional affiliation? Can researchers challenge the prevailing technical assumptions? Can scientific communities build systems around forms of knowledge that current models overlook?
The Path Forward
At TCA, our core belief is that “scientific discovery and the progress of engineering belong to the whole of humanity”.
This is an ethical commitment, but it is also an empirical one. The strongest innovation systems have always created room for competition, merit and correction. They attract ambitious talent that recognizes nothing is reserved for incumbents, and they grow more resilient because important ideas can come from anywhere. Only AI that is openly governed continues this tradition.
— The TCA team