If you can't reproduce the model then it's not open-source
Whether AI models should be called “open source” is contested, with many arguing that releasing only model weights—without datasets, data collection pipelines, or full training recipes—falls short of the transparency and reproducibility traditionally expected in open-source software. Others counter that what most users care about is the ability to download, run, and fine‑tune models, and that full data release is often impractical due to cost, legality, and privacy concerns. The exchange highlights an emerging split between “open weights” and truly reproducible models, and ongoing efforts by groups like the Open Source Initiative and Debian to clarify what “open” should mean in the AI era.
What “open source” means for AI models
- Many argue that if a model cannot be reproduced from data + code + training recipe, it should not be called “open source,” only “open weights” or “model available.”
- Others counter that classic open source never guaranteed reproducibility or affordability; access to code (or weights + runner) is sufficient, even if most people lack the compute.
- There is debate over what counts as “source” for ML: weights, datasets, training scripts, or the whole pipeline. GPL/OSI definitions about “preferred form for modification” are frequently invoked but applied inconsistently.
Role of data, code, and reproducibility
- Some see datasets and collection/filtering scripts as the crucial missing pieces; without them you can’t retrain, audit bias, or verify claims.
- Others claim the trained weights are the practical “source,” because most useful work happens via fine-tuning or adding layers, not retraining from scratch.
- Reproducibility is further complicated by non-deterministic training, huge compute costs, and disappearing or private data sources.
Practical value vs philosophical purity
- One camp focuses on user freedoms: download, run locally, modify (via fine-tuning), and share derivatives. They see current “open” LLMs as an enormous gain versus API-only services.
- Another camp stresses long‑term autonomy: without data and recipes, the community cannot truly fork or continue models if original sponsors stop releasing versions.
- Comparisons are made to firmware blobs in Linux: better than nothing, but not fully open.
Legal and ethical concerns about data
- Many believe datasets remain closed to avoid copyright lawsuits and public backlash over controversial sources (e.g., books, social media, pirated content).
- There is concern that users of opaque models might unknowingly rely on infringed or biased data.
Compute, feasibility, and emerging efforts
- Several note that training state-of-the-art models is beyond most individuals, but smaller, useful models are feasible for nonprofits and labs.
- Some projects (e.g., Pythia, StableLM, RedPajama-like efforts) are cited as attempts at fully documented or public-data training, though true end‑to‑end reproducibility remains rare.
- Standards work (e.g., OSI “deep dive,” Debian ML policy) and ideas like “Dockerfiles for models” or zero-knowledge proofs of training are mentioned as possible paths forward.