What exactly is GPT-NL?
A collaboration of non-profit organizations, including TNO, SURF, and the Netherlands Forensic Institute (NFI), has announced plans to develop an open Dutch language model called 'GPT-NL'. This model will be accessible to academic institutions and government agencies, including those active in health, defense, and security sectors.
The goal of GPT-NL is to strengthen the digital autonomy of the Netherlands by deploying transparent, fair, and verifiable AI. Selmar Smit, chief scientist at TNO, emphasizes that GPT-NL will not only contribute substantively to Dutch organizations but also support the decision-making process in line with Dutch and European values, using technology and data.
What is the investment?
Approximately 13.5 million euros will be allocated for the development of the model. Compared to American tech companies, this budget is modest. OpenAI's valuation, for instance, is estimated at 90 billion dollars.
Competition
GPT-NL developer Selmar Smit indicates that they do not intend to compete with AI models from billion-dollar companies. Smit mentions that they aim to develop a much fairer and more responsible model. Transparency is crucial, and people will have the right to object if they disagree with the use of certain data.
How will privacy be ensured?
GPT-NL will be trained based on publicly available data, acknowledging that this data might contain personal information. To safeguard citizens' privacy, the consortium has agreed to exclude non-anonymized data from the training process. While a 100% exclusion cannot be guaranteed, the consortium is exploring solutions, such as an 'opt-out' option, to ensure citizens' data does not end up in GPT-NL's datasets.
The Netherlands Forensic Institute (NFI) highlights the importance of transparency in the use of models for forensic research. The consortium aims to ensure that all parties in the criminal justice process receive a clear explanation of the model's workings and training. These efforts are directed at minimizing the risk of subjectivity and unconscious bias in NFI's research conclusions.
The first version of GPT-NL is expected to be ready by the end of this year (2024).
