A new manifesto called Open Source AI Must Win makes a stark claim: if intelligence becomes something people can only rent from a few closed institutions, the public does not just lose software freedom. It loses operational freedom.
The document, published by an individual using the handle @TheAhmadOsman, is short. It runs a few paragraphs. But its framing is worth sitting with because it names a risk that the tech industry mostly dances around: the possibility that frontier AI becomes a subscription economy for cognition.
That phrase is the manifesto’s sharpest line. It captures something concrete. Today, the most capable models — GPT-5, Claude 4, Gemini 2.5 — are accessible primarily through APIs or chat interfaces controlled by their developers. You cannot download them. You cannot fork them. You cannot run them on your own hardware without paying per-token or per-seat. If the provider changes its terms, removes a model, or raises prices, users have no recourse.
The manifesto argues this is not a temporary state. It is a structural feature of how frontier AI is being built and distributed. The labs that train these models spend billions on compute. They have no incentive to give the weights away. The result is a world where the most powerful intelligence infrastructure is always someone else’s property, running on someone else’s servers, subject to someone else’s rules.
Open-source AI should remain usable, understandable, reproducible, locally deployable, economically viable, and community-governed, the manifesto states, even if today’s dominant labs, foreign labs, hardware vendors, cloud platforms, or open-weight model providers change direction or disappear.
That is a long list of dependencies to insulate against. It is also a realistic one. The past two years have shown how quickly the landscape shifts. OpenAI has restructured, changed its safety approach, and altered its API pricing multiple times. Meta has released open-weight models like Llama 3 but has also signaled that future versions may be more restricted. Google has open-sourced Gemma but kept Gemini proprietary. The manifestos point is that none of these decisions are permanent, and none are controlled by the people who depend on the technology.
The document frames AI as civilizational infrastructure for work, education, science, software, creativity, public services, and national capacity. That framing is not hyperbole. It is descriptive. Governments are already using AI for benefits processing, medical diagnosis, and military targeting. Schools are adopting AI tutoring tools. Software engineers rely on AI code generation. If these systems are all rented, the fragility is systemic.
The manifesto also introduces a geopolitical dimension. It says America should not fall behind on the freedom to run, inspect, modify, benchmark, teach, and preserve intelligence infrastructure. The practical posture, it argues, is American capacity with global open standards. This is a notable departure from the dominant policy conversation in Washington, which has focused on export controls and keeping advanced chips out of Chinese hands. The manifesto suggests that building domestic open-source capacity is as important as restricting foreign access.
There are obvious counterarguments. Open-weight models carry risks. They can be fine-tuned for harmful purposes. They can be used to generate disinformation at scale. They can be deployed without safety guardrails. The manifesto does not address these concerns. It treats openness as an unqualified good, which is a position that requires more nuance than the document provides.
But the core thesis is hard to dismiss. The AI industry is consolidating around a small number of closed providers. The number of labs training frontier models has shrunk. The cost of entry is rising. If the trend continues, the infrastructure of intelligence will be owned by a handful of companies, answerable to shareholders, not users.
The manifesto ends with an invitation to help make this real, directed to an email address. It is not a call to action in the traditional sense. It is more like a signal. Someone is trying to start a conversation about what open-source AI needs to become, not just what it is today.
The question the document leaves open is whether the open-source ecosystem can keep pace. Today’s open-weight models are good. They are not frontier. Llama 3 405B is competitive with GPT-4 on some benchmarks. But GPT-5 and Claude 4 are another step ahead. The gap may widen if the closed labs continue to scale their compute investments faster than the open community can match.
That is the tension the manifesto names but does not resolve. Open-source AI must win. The question is whether it can.