The promise of open-source AI rests on more than code. It rests on the assumption that those who need it most can actually run it — a condition that, for most of the world, remains unfulfilled.
My own encounter with this gap came during my current research at the University of Massachusetts Amherst, where I am investigating how large language models (LLMs) can support K-12 educators in assessing students’ Computational Thinking. Moving from cloud-based, proprietary APIs to locally deployed, open-weights inference environments forced me to reckon with something no benchmark had prepared me for: the distance between model availability and infrastructure access.
To understand how an LLM interprets a student’s thinking, I built a local inference environment from the ground up — deploying models on professional hardware, managing computational requirements, and eventually visualizing the internal structures these models use to represent meaning. Each layer of this process revealed a new dependency: specialized hardware, specific software configurations, ongoing maintenance, and the kind of infrastructure knowledge that is not typically taught in education departments. Drawing on a colleague’s recent work in evaluating LLMs (Kavut, 2026), I transformed high-dimensional model responses into two-dimensional maps to make the embeddings interpretable — a process that illustrated, in miniature, how much technical overhead lies between an open-weights model and meaningful research use. Maintaining a functional workspace across distributed, proprietary GPU hardware required integrating streaming protocols that most educators and researchers have never encountered, let alone configured.
This hands-on encounter with local AI infrastructure surfaces a question that is harder than it first appears: what does it actually mean to be AI-literate?
The dominant answer, embedded in most current K-12 frameworks and professional development programs, emphasizes prompt engineering: how to ask a model the right questions. But my experience points to a deeper layer. Understanding what a model is doing, where its knowledge comes from, and whether its outputs can be trusted requires a form of technical fluency that most educators are never offered and that most institutions never fund. The question is not whether this knowledge is valuable. The question is who gets to have it — and who is quietly excluded by not having it.
The case for deeper AI literacy is ultimately an argument for empowerment. When educators, researchers, and policymakers understand how these systems represent meaning, where they fail, and how their outputs are shaped by training data, they become less vulnerable to industry hype, vendor lock-in, and misinformation. In my own research, this understanding proved essential: it shifted my posture from passive consumer of automated outputs to informed, critical evaluator of what those outputs actually mean. That shift in posture is precisely what deeper AI literacy makes possible.
The counterargument is equally important. Demanding deep technical mastery risks gatekeeping — erecting barriers that exclude the very domain experts whose judgment AI systems most need. Educators, sociologists, and artists should not need to configure CUDA drivers or manage Docker-based containerized inference environments to participate meaningfully in decisions about how AI is used in their fields. Over-emphasizing technical fluency as a prerequisite can reproduce the same exclusions that open education movements have spent decades working to dismantle. This is not merely a pedagogical concern. It is a question about who gets to shape AI’s role in society, and whether the people most affected by these systems have any meaningful voice in how they are designed and deployed.
Yet this philosophical debate risks being overshadowed by an even harsher reality: the bottleneck of hardware access. Access to GPUs, high-performance compute clusters, and even basic cloud service subscriptions confers a decisive advantage in AI research and development. For underfunded school districts, independent learners and researchers, and communities already marginalized by digital divides, these resources are not merely expensive — they are out of reach. Open-weight models are only as open as the hardware required to run them. We may be turning a corner in releasing open models, but we are still falling behind in democratizing the compute required to use them.
This is where the principles of Open Education must intervene — and must expand their own scope. The open education movement has successfully championed open curricula, open textbooks, and open access publishing. These are necessary but no longer sufficient. As UNESCO’s Recommendation on Open Educational Resources (2019) recognized, access to educational opportunity increasingly depends on access to digital infrastructure. In the age of AI, that infrastructure includes compute. The World Bank’s Digital Progress and Trends Report 2023 identifies compute as a foundational pillar of AI development alongside connectivity, context, and competency. Open education must now make the same argument about silicon that it once made about syllabi.
1. Establish a National AI Service Corps
Just as the National Health Service Corps subsidizes medical training in exchange for rural clinic service, we should consider creating a similar pipeline for AI expertise in education. A National AI Service Corps would provide graduate students in Computer Science, Information Science, Learning Technology, and Machine Learning with stipends in exchange for a committed period of service in under-resourced K-12 school districts or municipal governments after graduation. This could also build on the Ford Foundation’s pioneering Public Interest Technology movement, which embeds technologists in local governments.
Rather than seeing top-tier technical talent absorbed entirely by corporate tech giants, marginalized communities and K-12 educators would gain embedded experts capable of architecting, troubleshooting, and configuring local AI environments — ones appropriate to their specific instructional context and data privacy requirements.
2. Fund Open Educational Compute as Public Infrastructure
The Open Education movement has successfully championed Open Educational Resources to make curricula freely available. Federal and state governments should mirror this by establishing grants specifically to provision K-12 school districts with physical compute — so that models developed by researchers can be applied and deployed locally. Recent advances in low-powered but capable minicomputers equipped with GPUs can be networked together as entry-level clusters, with potentially minimal environmental impact, to serve entire school districts.
As the World Bank highlighted in its Digital Progress and Trends Report 2025, compute is now one of four dominant factors in reaching AI literacy. Philanthropic organizations like the Gates Foundation are already funding K-12 districts to focus on AI infrastructure and data systems. Dedicated compute grants would fund localized hardware — robust multi-GPU server clusters — breaking the reliance on rate-limited, corporate cloud APIs.
3. Shift K-12 AI Literacy from Passive Prompting to Active Architecture
Current curricula too often focus on how to write effective prompts. Drawing inspiration from the European Union’s AI for Public Good initiatives, which prioritize open-access models and human-capital pipelines, educational policy should redefine AI literacy to include evaluating the mechanics of models themselves — how they are trained, where they fail, and how their outputs are shaped by data.
We might also mirror the World Bank’s ongoing La IA en el Aula (AI in the Classroom) study in Peru (Molina, 2025), which guides sixth-grade teachers to use LLMs as co-creators for designing differentiated, context-specific formative assessments — demonstrating that technology can support, rather than replace, teachers’ professional judgment. Experimenting with localized deployment, model fundamentals, and data privacy in K-12 settings could help transform marginalized students from passive consumers into active architects of AI.
The open-source software community has given us the architecture. Organizations championing Open Education are fighting to give us the pedagogy. But without equitable access to the silicon itself — and the human capital required to orchestrate it — the AI revolution will only reinforce existing societal divides. By subsidizing the deployment of technical expertise and localized compute into the communities that need them most, we can begin to ensure that the promise of an AI-enabled future is not reserved for the already well-resourced.
References
Kavut, M., Dzafic, A., & Bayram, U. (2026). Assessing the representation of suicidal ideation in social media datasets relative to suicide notes. IEEE Transactions on Affective Computing, 1–12. https://ieeexplore.ieee.org/document/11396956
Miao, F., Mishra, S., Orr, D., & Janssen, B. (2019). Guidelines on the development of open educational resources policies. UNESCO Publishing.
Molina, E., et al. (2025). La IA en el Aula: Tu Aliada Pedagógica. World Bank Group. https://documents.worldbank.org/en/publication/documents-reports/documentdetail/099623210222512099
UNESCO. (2019). Recommendation on Open Educational Resources (OER). https://www.unesco.org/en/legal-affairs/recommendation-open-educational-resources-oer
World Bank. (2023). Digital Progress and Trends Report 2023. https://www.worldbank.org/en/publication/digital-progress-and-trends-report
World Bank. (2025). Digital Progress and Trends Report 2025: AI Foundations. https://www.worldbank.org/en/publication/dptr2025-ai-foundations
About the Author
Will Lee has over sixteen years of experience in the government consulting sector. He is currently a graduate student in the Computer Science department at the University of Massachusetts Amherst, where he studies the intersection of large language models, Computational Thinking, and K-12 education under the supervision of Professor Ivon Arroyo at the Advanced Learning Technologies Lab.
Cite this article
Lee, W. (2026). Beyond the open-weights: Open education, the global compute divide, and the future of AI literacy. Society and AI. https://societyandai.org/perspectives/beyond-the-open-weights/
Write for Society & AI
We welcome contributions from educators, researchers, policymakers, and practitioners examining the intersection of artificial intelligence, education, and society. All work published on Society & AI is made available free of cost and in open access, in service of education and the public good. If you have a perspective to share, send your proposal to the Editorial and Content Director at submissions@societyandai.org.
