Circular AI Economy

This project aims to deliver: (1) comprehensive studies of existing paradigms governing the AI lifecycle—including production (training AI models), distribution (deploying the models), and consumption (decision-making); (2) novel paradigms aligned with the principles of a circular AI economy; and (3) a practical platform that enables practitioners to implement these paradigms in real-world applications.

Problems: Modern Artificial Intelligence (AI) systems—such as predictive models, generative models, and large language models—have shown impressive capabilities, driving progress in science, healthcare, education, and daily life. However, adopting these powerful technologies come with significant environmental costs. By 2025, AI is expected to consume more electricity than Bitcoin mining, potentially making up nearly half of the energy used by data centers worldwide. Training large models like GPT-3 can require up to 700,000 liters of water, and by 2027, data centers may consume as much as 6.6 billion cubic meters of water. 

As resources for developing and deploying AI systems become increasingly scarce, there is a growing need to create more sustainable ecosystems for AI development. [Data] Public data on the internet, which has driven much of the current AI progress, is becoming depleted, leading to a rising demand for private data. This shift raises significant concerns around data privacy and copyright infringement. [Compute Resources] At the same time, the escalating demand for computational resources—particularly GPUs and data centers—not only during training but also at inference time, is placing unprecedented strain on infrastructure and widening the gap in AI capabilities. These challenges call for a paradigm shift in the AI development lifecycle.

State-of-the-art: The traditional train-then-test paradigm is increasingly being replaced by more flexible and scalable approaches. Contemporary AI development often involves pre-training models on large, general-purpose datasets, followed by fine-tuning on task-specific data. API-based access to models—such as Claude 3.5 Sonnet (Anthropic) and GPT-4o (OpenAI)—not only supports integration into external applications but also enables fine-tuning for specialized use cases. Meanwhile, open-weight models like Llama 3 (Meta) and Mixtral (Mistral) allow users to download and run models locally. In some instances, model weights, datasets, and training/inference code are all openly available, empowering users to reproduce and adapt models in their own environments. Additionally, recent advances in federated and distributed learning enable communities to collaboratively train models in a decentralized manner, fostering collective learning and shared benefits.

Toward circular AI economy: While the production, distribution, and consumption of AI models as commodities have significantly expanded AI’s influence in science and society, it remains unclear whether these benefits outweigh the associated environmental costs and human labor demands. To date, there has been limited effort to explore how existing AI models can be reused, recycled, or repaired to support a more circular AI lifecycle. Advancing toward this goal requires more sustainable training and inference procedures and the development of platforms that facilitate the efficient use of raw data, models, and computational resources through reuse, recycling, and repair.