About
Since the emergence of large language models, the world has been moving around this new technological frontier, attempting to accommodate their potentialities and their impacts on the economy, society, and the environment. The earliest companies that discovered this frontier have pursued an inconsistent pursuit of progressively larger models, lacking a viable budgeting strategy for the long term. Moreover, since these models are closed and proprietary, there was no possibility for society to engage in a broader debate about the technology.
Another approach, more aligned with the values of the open-source (OSS) community, advocated opening the model weights, thereby significantly expanding possibilities for own infrastructure use, fine-tuning, and model distillation. This approach is more sustainable over the long term because it aligns with shared societal interests, such as scientific research and socio-environmental discourse, whereas companies may concentrate on model innovation and commercial technology deployment.
For society, this scenario creates particularly fertile ground for deepening knowledge and mastery of the technology. The climate emergency and the urgent need to democratize access to the technological frontier call for a collective effort aimed at developing rational and efficient uses of AI models, thereby establishing a cost-benefit relationship that makes sense from economic, social, and environmental perspectives. More than a mere research stimulus, it constitutes one of contemporary society's major challenges.
This blog is a modest contribution to these efforts. I have been working on the hypothesis that careful context curation, combined with fine-tuning techniques, can improve the inference quality of smaller models, reduce token consumption by larger models, and in both cases diminish or perhaps eliminate noise and hallucination during inference.
This led me also to initiate the project Datorum, a framework of context engineering, designed to serve both as a testing laboratory and scalable to production deployment upon achieving a stable version. Here you can follow the project's development, tests, use cases, and some reflections on the broader context, sharing both advances and dead-end paths. You are also welcome to provide feedback on Datorum through the project's issues page (link) and to interact with me on Fediverse (link).
My links: - CodeBerg: alicebonafe - GitHub: alicecbonafe
Datorum links - CodeBerg: datorum - GitHub: datorum - PyPI: datorum