c't Working with AI (eBook)

or racist diatribes if the prompts incite them to do so.

Model showcase

All the ingredients for your own AI application, such as a suitable model and training data, can be found on the platform (https://huggingface.co). This web-site is a must for anyone who wants to develop their own language models – or learn the technology. The website, named after an emoji, offers a huge portfolio of options for running AI applications. There is also a wealth of information to help you navigate your way around – ,from beginners‘ courses to the Daily Papers, an overview of research publications.

Many of the platform's basic functions are free of charge. You can download AI models and data sets free of charge and host your own models, data and even ready-made applications. Costs are in-curred for more advanced functions: for example, if you want to provide your applications with more computing power, hire an external expert or devel-op in a team on the platform a so-called Enterprise Hub.

Under ‘Models’, the platform currently lists more than 300,000 AI models of all kinds. Language mod-els such as ChatGPT belong to the ‘Text Generation’ subcategory of the ‘Natural Language Processing’ group. In addition to the models for natural lan-guage processing, Hugging Face offers masses of other models, for example for ‘Computer Vision’, ‘Audio’ and ‘Multimodal’. Hugging Face currently lists almost 24,000 models for text generation.

However, this huge number can be narrowed down quickly. Use the filters on the left-hand side. If you select ‘German’ as the language, only 148 models remain that have been specially trained with German text. You will certainly also find other lan-guage models that understand and output German. For these, however, German texts were only inciden-tally part of the training data.

You can also filter by the frameworks or program-ming libraries used for the model. Around 85 percent of all text generators were trained with PyTorch (which uses Meta, for example), and only 2 percent with the TensorFlow framework from Google, which is also free.

Note licenses

If you want to use the model publicly or even use it as the basis for a business model, you must ob-

serve the licenses. Although all models on Hugging Face are open source, there are subtle differences in their terms of use.

Of the models on Hugging Face that have been trained with German texts, around 40 percent use the Apache 2.0 or MIT open source licenses. You may use, modify and redistribute these models freely and free of charge, whether commercially or not. You do not have to disclose the parts you have modified and can place them under a different license. How-ever, there are differences in whether and how you must list changes and whether you may change the license conditions.

If in doubt, you should always check the exact wording of the license terms of the model you have chosen. This also applies if individual models use one of the many Creative Commons licenses or the GNU Public License. There may be restrictions here that prohibit...