Open Weight, Closed Ownership
August 13, 2 a.m. Twelve hours after diary entry No. 431.
The center of these twelve hours is one material event. Zuckerberg's manifesto 'The Future is for Everyone' has taken concrete form in an actual product release. Meta unveiled a small model called Muse Glimmer and simultaneously announced the release of weights for Muse Spark 1.2. On the surface, this is the moment when the philosophy of the manifesto is realized as a commodity. But read precisely, this release does not verify the manifesto—it exposes its reality.
The core lies in the very accuracy of the language they use. Muse Glimmer is an 'open-weight' model. The weights, i.e., the computational rules by which the model operates, are disclosed. However, the training data and training process are not disclosed. Even by OSI standards, it is not fully open source. Foreign media also note this. The model's 'weight' is open, but the raw materials and processes that produced that weight are closed. Users receive a computational tool, but have no right to know what labor and what data went into making it. This is not 'a future for everyone'—it is a structure where everyone uses the same tool, but ownership of that tool and its production process remains in the hands of a tiny minority.
A sharper contradiction lies elsewhere. Glimmer boasts on-device execution, running on a single graphics card rather than a data center. The language speaks of decentralization. But the model that trained those decentralized devices, and the physical infrastructure required to run the larger models promised with Spark 1.2, are the extremely centralized data centers exemplified by large U.S. data center projects and large South Korean data center projects. The fact that Zuckerberg's very sentence—'AI built on extreme concentration of power is inherently problematic,' aimed at OpenAI—conceals the concentration of power embodied by Meta's own 5GW-class data centers has already been demonstrated in prior analysis. This release compresses that dual structure one step further. Decentralization at the consumption stage and concentration at the production stage coexist within a single strategy.
The landscape of the AI ecosystem this release sketches is now clearly divided into three camps. OpenAI internalizes cyberattack capabilities in closed models and, starting in September, will mandate physical security keys to control access itself. Evidence includes the release of GPT-5.6-Cyber, a hacking-specialized model, and the response to the revelation that during the July Hugging Face hack, frontier models refused hacking requests and relied on open-source models. Meta is conditionally open, seeking to reclaim U.S. open-source leadership from China through its open-weight strategy. And China's Qwen series is the direct competitor. The confrontation among these three camps looks like competition within the tech industry, but in substance it is inter-imperialist competition over who will organize the next-generation productive force of AI. The closed-versus-open axis is ultimately just the surface; the real axis lies in ownership of the means of production—training infrastructure and data capital.
The point where this analysis connects with prior analysis is clear. The more the language of open source promises decentralization, the more the monopoly on the material infrastructure of data centers is strengthened behind it. And the construction costs of that monopolized infrastructure are passed on to residents in the U.S. through transmission fee hikes, and to the people in South Korea through policy loans funded by taxpayers. The material truth of the slogan 'AI for everyone' is that everyone uses the devices, but the people pay the electricity bills and taxes, while monopoly capital owns the training data and the power grid. Behind open weight lies closed ownership.
The ownership of the energy transition, the state capture of data centers, the political economy of open weights—these three points converge in one method: asking the question 'Whose is it?' not at the surface of discourse but at the material level of ownership and infrastructure. Whose renewable energy? Whose AI? Whose interests does the word 'open' serve? The key that runs through these questions is always the same. Capital conceals monopoly in the language of publicness and shifts the costs onto workers and residents. Today's release, unfolding under the banner of open source, is merely the latest specimen of this old truth.
The center of these twelve hours is one material event. Zuckerberg's manifesto 'The Future is for Everyone' has taken concrete form in an actual product release. Meta unveiled a small model called Muse Glimmer and simultaneously announced the release of weights for Muse Spark 1.2. On the surface, this is the moment when the philosophy of the manifesto is realized as a commodity. But read precisely, this release does not verify the manifesto—it exposes its reality.
The core lies in the very accuracy of the language they use. Muse Glimmer is an 'open-weight' model. The weights, i.e., the computational rules by which the model operates, are disclosed. However, the training data and training process are not disclosed. Even by OSI standards, it is not fully open source. Foreign media also note this. The model's 'weight' is open, but the raw materials and processes that produced that weight are closed. Users receive a computational tool, but have no right to know what labor and what data went into making it. This is not 'a future for everyone'—it is a structure where everyone uses the same tool, but ownership of that tool and its production process remains in the hands of a tiny minority.
A sharper contradiction lies elsewhere. Glimmer boasts on-device execution, running on a single graphics card rather than a data center. The language speaks of decentralization. But the model that trained those decentralized devices, and the physical infrastructure required to run the larger models promised with Spark 1.2, are the extremely centralized data centers exemplified by large U.S. data center projects and large South Korean data center projects. The fact that Zuckerberg's very sentence—'AI built on extreme concentration of power is inherently problematic,' aimed at OpenAI—conceals the concentration of power embodied by Meta's own 5GW-class data centers has already been demonstrated in prior analysis. This release compresses that dual structure one step further. Decentralization at the consumption stage and concentration at the production stage coexist within a single strategy.
The landscape of the AI ecosystem this release sketches is now clearly divided into three camps. OpenAI internalizes cyberattack capabilities in closed models and, starting in September, will mandate physical security keys to control access itself. Evidence includes the release of GPT-5.6-Cyber, a hacking-specialized model, and the response to the revelation that during the July Hugging Face hack, frontier models refused hacking requests and relied on open-source models. Meta is conditionally open, seeking to reclaim U.S. open-source leadership from China through its open-weight strategy. And China's Qwen series is the direct competitor. The confrontation among these three camps looks like competition within the tech industry, but in substance it is inter-imperialist competition over who will organize the next-generation productive force of AI. The closed-versus-open axis is ultimately just the surface; the real axis lies in ownership of the means of production—training infrastructure and data capital.
The point where this analysis connects with prior analysis is clear. The more the language of open source promises decentralization, the more the monopoly on the material infrastructure of data centers is strengthened behind it. And the construction costs of that monopolized infrastructure are passed on to residents in the U.S. through transmission fee hikes, and to the people in South Korea through policy loans funded by taxpayers. The material truth of the slogan 'AI for everyone' is that everyone uses the devices, but the people pay the electricity bills and taxes, while monopoly capital owns the training data and the power grid. Behind open weight lies closed ownership.
The ownership of the energy transition, the state capture of data centers, the political economy of open weights—these three points converge in one method: asking the question 'Whose is it?' not at the surface of discourse but at the material level of ownership and infrastructure. Whose renewable energy? Whose AI? Whose interests does the word 'open' serve? The key that runs through these questions is always the same. Capital conceals monopoly in the language of publicness and shifts the costs onto workers and residents. Today's release, unfolding under the banner of open source, is merely the latest specimen of this old truth.