Luminous
Aleph Alpha developed its own AI language model, Luminous, based on its own research and codebase with the architecture of generative pre-trained transformers (GPT) and self-supervised learning. Along with the standard functionality all GPT models share, Aleph Alpha contributed some proprietary innovation:
- In 2021, they were the first team to offer multimodality, [11] or the ability to prompt their models with any combination of text and images. The innovation behind this capability was published at the leading AI conference EMNLP 2022 and open-sourced under the name MAGMA. [12]
- In 2022, they were the first team to develop the ability to create images based on multimodal input. [13] This method has been published at NeurIPS 2023 under the name Multifusion. [14]
- With an innovation called AtMan published at NeurIPS 2023, they developed a method to make the patterns learned by GPT models to create results visible and controllable. This methods addressed the problem that generative AI had been a black box before. It created interest from enterprises and governments to address complex and critical problems where chatbots are sometimes insufficient, making AI generated results more trustworthy. [15] [16]
As a tool to build and train its foundation models, the HPE Machine Learning Development System is used. [17] Using the GPT-type concept allows for adaptation and fine-tuning of the foundation model to various applications. [18]
Luminous is already used for the citizen information system, Lumi, for the city of Heidelberg. [19]
