Open Source and AI for Sustainable Development: Digital Infrastructure, Open Data, and Multistakeholder Cooperation

Wei Wang draws a parallel between Richard Stallman's foundational question about who controls your computer in the closed-source software era and a new, more pressing question in the AI age: who controls your mind and knowledge. He argues that open source AI is critical for making AI more transparent to end users, giving them the right to understand how large models retrieve and process information. This transparency is essential for both users and developers.

Wei Wang invoked Richard Stallman's question 'who controls your computer?' as a historical reference point for the open source movement , then reframed it for the AI era by asking 'who controls your mind and your knowledge?' . He argued that open source AI would be a great tool to make AI more transparent to end users and give them the right to understand the mechanisms by which big data models retrieve information from open source code, models, and datasets .

Wei Wang contends that while AI has made it easy for anyone to write code, the most important engineering skills — judgement, taste, and quality control — cannot be developed through AI alone. Open source projects provide real, industry-relevant tasks that students can engage with on campus, making open source communities the ideal practical course for training future engineers. He integrates open source projects directly into his university teaching.

Wei Wang described how he incorporates open source projects into his university courses and conducts research by mining data from GitHub and HuggingFace using his own open source project, OpenDigger . He argued that in the age of AI, where anyone can write code, the most important abilities are judgement, taste, and quality control, and that open source projects provide real industry-relevant tasks students can work on while still on campus .

Wei Wang highlights that open source AI is far more complex than open source software, encompassing models, data, publication reports, and source code, yet there is no agreed standard for what constitutes openness across these elements. He notes that many companies market their models as open source without genuine transparency, which misleads users. His lab is therefore working to build a spectrum or catalogue that classifies the actual level of openness of AI models.

Wei Wang noted that open source AI includes at minimum models, data, publication reports, and source code, and that there is currently no agreement on what level of openness is required across these elements . He observed that many corporations claim their models are open source as a marketing story, without genuine transparency . He described his lab's effort to build a standard spectrum that catalogues the transparency level of models so users know what they can actually do with them, whether for commercial use, research, or other purposes .

Wei Wang points out that in the open source software era, contributors invested significant time without receiving commercial benefit, and this problem is amplified in the AI age where AI can generate content easily. He argues that constructing a system that provides commercial feedback or benefit to contributors would make open source AI ecosystems more sustainable and vibrant. He frames this as an opportunity to build a new, better system.

Wei Wang noted that in the open source software age, many contributors dedicated their time to projects without benefiting commercially from them . He argued that in the AI age, if a system could be constructed whereby contributors to open source receive commercial feedback or benefit, the ecosystem could run even better, and framed this as an opportunity to build a new system since 'the whole world is broken' .

Wei Wang argues that AI has become a new digital infrastructure for the public, comparable in importance to air and water, and that this is especially true in education. Students and teachers should not have to pay heavily for AI tokens in order to learn and teach effectively. Open source is presented as the best mechanism to implement an economic system that ensures equitable, affordable access to AI for educational purposes.

Wei Wang described AI as 'a new digital infra of the public, just like the air and the water', particularly in public domains such as education . He argued that students and teachers must not pay much for AI tokens and should be free to access AI in order to be more efficient in learning and teaching . He concluded that open source is the best way to implement this vision, despite the many challenges involved .

Yue Gao frames open source not merely as a technical collaboration model but as a governance philosophy with broad societal implications. She argues that open source software and open data form the backbone of digital infrastructure and that their synergy with multi-stakeholder collaboration is essential for achieving the SDGs. She highlights concrete application domains such as climate monitoring, public health, and smart city development.

Yue Gao described open source as 'not merely a technical collaboration model' but 'a governance philosophy that promotes global knowledge sharing and narrowing the digital divide' . She cited examples including Linux, Apache, HyRub, and PyTorch as illustrations of open source's foundational role . She also noted that the achievement of the SDGs relies on open, trustworthy, and collaborative digital infrastructure, with applications in climate monitoring, public health early warning, and smart city development .

Christofer Dutz describes how AI-generated contributions, initially appearing as trivial pull requests about punctuation, have grown into a massive volume of submissions that overwhelm maintainers of large open source projects. He warns of a risk of burning out core contributors to major projects. However, he expresses confidence that AI-based tools developed within the open source community itself, such as Apache Magpie, can help manage this load.

Dutz described observing an influx of trivial pull requests - such as whether a comma or full stop should appear in documentation - which he attributed to people trying to build reputations by having pull requests merged across many projects . He noted that large projects like Apache Airflow receive up to 170 pull requests per day, each requiring review, triage, and feedback, creating enormous stress on maintainers . He referenced the Apache Magpie project as an AI-based tool designed to review, sort, and merge pull requests to reduce this load .

Dutz draws a distinction between open weights, open models, and fully open data, referencing the Linux Foundation's multi-level classification framework. He argues that the Apache Software Foundation's principle of releasing source code — so that any two people building from the same package get identical results — cannot apply to AI, because training the same model on the same data at different times yields different results. This non-determinism fundamentally challenges the open source ethos as applied to AI.

Dutz referenced the Linux Foundation's classification of AI openness into approximately three or four levels: open model (the AI binary), open weights (including settings and build infrastructure), and open data (including training information) . He contrasted this with Apache's principle of releasing source code rather than binaries, so that two people building from the same source package get the same result . He argued that this deterministic reproducibility is impossible in AI, because even running the same training on the same data on the same infrastructure at different times yields different results .

Dutz argues that even companies that would genuinely like to publish their training data are often legally prevented from doing so because that data includes copyrighted material such as books from libraries. He uses the analogy of a person reading all the books in a library and then quoting from them, which is acceptable for an individual but becomes a copyright violation if the content is republished at scale. This creates a structural tension between the desire for open data and intellectual property law.

Dutz referenced reports about Google training its models by feeding entire library collections into the training process . He drew an analogy: just as an individual could read all the books in a library and quote from them without legal issue, a company doing the equivalent is acceptable - but if that company were required to release all the training data, it would be violating many copyrights . He concluded that sometimes, even if companies would like to publish training data, it is simply not possible from a copyright perspective .

Dutz argues that while language models have become relatively mainstream, vision models present a more complex set of legal and ethical challenges. These include determining whether a generated image constitutes a copy of a copyrighted work, and the risks posed by deep fakes, including the non-consensual use of other people's images. He emphasises that the safeguards required for vision models are significantly more demanding than those for language models.

Dutz noted that vision model training is more likely to conflict with copyright law, raising the question of where to draw the line between a generated image and a copy of a copyrighted image . He also highlighted the problem of deep fakes, noting that while uploading a picture of oneself for creative purposes is acceptable, uploading someone else's image is a problem, and detecting this distinction is a significant challenge .

Dutz describes world models — which he also calls physics models — as the ultimate goal for industrial AI, capable of predicting system behaviour from specifications alone. However, he acknowledges that the training data and computational resources required for such models currently exceed available data centre capacity. He presents time series models as a practical near-term alternative, capable of observing existing systems and optimising them over time without needing to model their physics from scratch.

Dutz described world models as his 'holy grail', explaining that for his work in industrial automation with PLC4X, a physics model could predict and calculate system behaviour based solely on device specifications . He noted that the training data and computational capacity required for such a model 'just exceeds our human data center capacity' . He then described time series models as a 'world model light', citing tools like IoTDB as ready-to-use examples that allow observation of existing systems and prediction-based optimisation without requiring full physics modelling .

In response to an audience question, Dutz acknowledges that computing resources represent a significant additional barrier to open source AI participation, beyond the open model versus closed data tension. He explains that the Apache Software Foundation does not own or operate significant hardware infrastructure and relies on contractors, making it currently unable to provide GPU resources for AI training. He suggests that substantial donations from large AI companies would be needed to change this.

Dutz explained that the Apache Software Foundation does not own or run much hardware on its own and relies on contractors for such needs . He stated that he does not foresee the ASF building data centres full of GPUs for AI training, and suggested that only very large donations - in the two-to-three digit million range - from big AI companies might make this feasible .

Dutz argues that while the current wave of AI-generated contributions poses a serious threat to open source maintainers, the projects that survive this period will emerge stronger. Having been reviewed by both human and artificial eyes, surviving open source code will have had its issues identified and addressed at a scale that proprietary software cannot match. This makes openness itself a long-term competitive advantage.

Dutz used the example of advocating for open source in the industrial automation sector, where he often faces the objection that proprietary vendors like Siemens, Schneider, and Rockwell have decades of experience and that open source exposes zero-day exploits . He argued that after surviving the AI tsunami, open source can claim that 'so many human eyes and so many artificial eyes have reviewed all of this code and we've addressed all of the major issues', which is something proprietary software cannot claim at the same scale .

Dutz describes specification-driven development as an approach where project rules and architecture are written in natural language (e.g., Markdown) and checked in with the project, rather than requiring contributors to master a specific programming language. He argues that AI has supercharged this approach, enabling contributors to implement complex functionality in languages they do not know by simply specifying what they want. This dramatically lowers the barrier to participation in open source projects.

Dutz described experimenting with specification-driven development years ago, writing project rules and architecture in Markdown or ASCII and checking them in with the project . He described a concrete experiment in which he wrote a specification for an industrial driver in Java, then asked Claude (an AI assistant) to implement it in Rust - a language he does not know - and the AI produced a driver that communicated perfectly . He argued that being able to formulate intent without dealing with syntax will open many more doors in the future .

Dutz warns that as AI-generated code enters open source repositories and is subsequently used to train future AI models, there is a risk of recursive quality degradation — a feedback loop in which AI trains on its own outputs, progressively degrading quality. He notes that this concern is a significant motivation behind large AI companies donating to open source foundations, as they fear their models will begin training on AI-generated rather than human-generated code.

Dutz acknowledged that while AI helps contributors with weaker software engineering backgrounds to participate in open source projects like PLC4X , scaling this up means large volumes of AI-generated code will enter open source repositories and then feed back into AI training data, which he described as 'recursive poisoning of AI datasets, which everybody's afraid of' . He suggested that donations from large AI companies to open source foundations are partly motivated by these companies' fear of their models 'starting to eat their own dog food all the time' .

Jianmin Wang acknowledges the fundamental tension between open models and closed data, drawing on his experience building IoTDB for industrial applications. He argues that enterprises view their data as private property and will not share it openly, but that open source training code can enable them to train models in their own private environments. As a compromise, he proposes that synthetic or simulated datasets generated by domain foundation models could be shared publicly.

Jianmin Wang described his experience building IoTDB, a database for industrial applications, noting that industrial enterprises consider their data to be private assets and properties . He proposed that open source training code could be downloaded by enterprises to train models in their private environments, allowing models to take on characteristics of their private datasets . He further suggested that open source tools could generate artificial/simulated datasets that could be made public and shared, potentially using domain foundation models to generate simulation data .

Jianmin Wang argues that the focus on large language models overlooks the rich variety of data types generated in industrial and enterprise settings, including relational data from ERP and SCM systems and time series data from machines. He sees an opportunity to combine these data types to train multi-modal, domain-specific foundation models tailored to enterprise needs. He identifies this as a distinct track for open source development alongside general-purpose language models.

Jianmin Wang noted that industries and enterprises generate large volumes of relational data from ERP and SCM systems, as well as time series data from machines . He argued that attention should not be limited to large language models, as there are opportunities to combine multiple data types to train multi-enterprise large language models and large time series models for specific business domains .

Jianmin Wang argues that while world models and physical AI hold great promise, progress is hampered by a lack of standardised, interoperable data formats for physical and embodied AI data. He draws on his own work providing a time series file format (TS file) to Hugging Face as a positive example, and calls for similar standardisation efforts for physical data collections. He frames this as an opportunity for the AI for good community to contribute.

Jianmin Wang noted that unlike language models, which benefit from abundant internet data, world models and physical AI suffer from a lack of data, particularly for space and time . He described his team's contribution of a time series file format (TS file) to Hugging Face as a data format for time series data . He argued that there is currently a lack of equivalent data format standards for physical data collections and called for the development of a general data set format for physical AI and world models as an 'AI for good' initiative .

Jianmin Wang, responding to an audience question, frames computing resource access as a distinct and important dimension of openness in the AI era, separate from the open model versus closed data debate. He suggests that the open source community needs to consider how to address unequal access to the computational power required for AI training, not just access to code and data.

Jianmin Wang characterised the audience question as raising 'another dimension for open source - the open source for the computing power resource' , framing computing access as a distinct and important aspect of openness that the community must grapple with.

Jianmin Wang argues that software engineering has undergone a fundamental upgrade to what he calls 'version 3.0', characterised by three distinct paradigms: traditional high-level programming languages, foundation models as integral components of software systems, and natural language prompting. He contends that future software production will be a combination of all three formats rather than a replacement of one by another.

Jianmin Wang referenced a claim made the previous year that software engineering had been upgraded to 'three point year old' (version 3.0), comprising: high-level programming languages as the first paradigm, foundation models as a part of software and software systems as the second, and natural language prompting as the third . He concluded that future software will be a combination of all three formats as new production styles emerge .

An audience member raises the point that the tension in open source AI is not solely about the divide between open models and closed datasets, but also encompasses the significant barrier posed by access to computing resources. They question whether open source software foundations like the Apache Software Foundation could step in to provide computing infrastructure, such as GPUs, to contributors who lack access.

The audience member explicitly stated that 'the conflict does not only exist between the open model and closed data set, but also the computing resource and the model' . They then directed a question to Chris asking whether the Apache Software Foundation would solve this problem or provide computing resources to contributors .

A second audience member, representing the Chinese Institute for AI Development Strategy and the World Federation of Engineering Organizations, poses a broader question about the long-term technical trends and future vision for open source software. They acknowledge Chris's framing of the current AI wave as a 'tsunami' and ask what the outlook is for open source once that challenge has been navigated.

The audience member asked all panellists for their view on the technical trends of open source software, referencing Chris's tsunami metaphor, and specifically asked 'what is your vision to that if you survive from the tsunami, what is the new outlook?' .