Open Source
A development methodology where source code is made public, allowing anyone to freely view, modify, and distribute it. In the AI field, according to the 'Open Source AI Definition (OSAID)' announced in 2024, it refers to systems where not only model weights but also training data information and training code are transparently disclosed, enabling replication and improvement.
Detailed explanation
Why it matters in tool selection
It allows organizations to escape vendor lock-in without relying on a specific company's APIs and enhances security by deploying directly on internal servers. Additionally, it is advantageous for building customized solutions optimized for business needs through fine-tuning on proprietary company data.
What to check
- Is the license a standard OSI-approved license (Apache 2.0, MIT, etc.), or does it have custom restrictions?
- Are there terms that limit the number of users or require separate royalties for commercial use?
- Are the sources and processing methods of the training data disclosed, allowing for bias and security reviews?
- Is the project actively maintained by the community, as indicated by GitHub stars or contribution activity?
Key Examples
PyTorch (framework), Transformers (library), OLMo (a model with fully open training data). In contrast, models like Llama 3 or Mistral have open weights, but behave more like 'Open Weights' in terms of licenses and data.
Confusing Terms
Open Source AI
A state where code, weights, and training data information are all public, enabling anyone to reproduce the model at the same level.
Open Weights
Models where the trained output (weights) can be downloaded and used, but the training process or data remains private (e.g., Llama, Gemma).