Chat GPT has in a short time become popular within the world of schools and culture as well as business and the public sector. The service is also able to respond to users in Swedish. So is a Swedish competitor really needed?
— Now, as I said, GPT SW3 is not a service but the underlying model. But yes, I would say that. We have been completely transparent with what data we used, how we built the model and what considerations we had, says Magnus Sahlgren, research leader at AI Sweden which is Sweden’s national AI center and continues:
— It differs markedly from the commercial models where we do not know at all what data it was trained on.
GPT SW3, as the Swedish AI model is called, is a so-called language model. In other words, it is good at text, specifically at generating text, especially in Swedish. But it is not a direct competitor to the popular Chat GPT – which can do the same thing.
— GPT SW3 is only a pre-trained model. It knows the Swedish language. Chat GPT is a product, which is built by an American company. At the core of that product is a GPT-type language model. Then they built a lot of other things on top of that model, says Magnus Sahlgren.
Democracy issue
The question of data, the information that makes an AI service powerful, and where it comes from is something that Sahlgren points out is important. For several reasons.
— It is to some extent the Swedish people who contributed the data to these models. Then we consider it a matter of democratization to give the model back to the users.
— It’s all of us who have contributed to the power of Chat GPT. But then we have to pay to use it. I think that’s completely crazy. It should be free if we have contributed data to it.
AI tax
The transparency and the “label of origin” in terms of the underlying data and the working method are two advantages that Sahlgren highlights, in addition to “buying Swedish”.
— One result of the global development we have is that we will all pay a kind of AI tax to Open AI or Microsoft. It’s a situation I don’t think is super successful. Partly because all the data we have flows out of the country. Data is knowledge. And so we pay taxes to use this kind of AI as well.
GPT SW3 is launched publicly in different sizes, the smallest can be “run at home on the laptop”, and the hope is that both companies, organizations and authorities use the model as an engine to build their own services on top. Sahlgren points out, however, that the bigger the better principle applies in this context.
— The larger the model, the better the quality.