PhariaAI: Aleph Alpha presents AI operating system

Aleph Alpha founder Jonas Andrulis said a few months ago that the development of a new large language model was not worthwhile. Now comes the announcement of an entire AI operating system. It will enable companies to integrate AI applications into their own structures. Compliance, in particular, will be a top priority. According to Aleph Alpha, anyone using PhariaAI is complying with the European AI Act, among other things.

The AI operating system should be easy to integrate into existing IT environments. PhariaAI can then be used to develop, control, train and more AI systems. Aleph Alpha’s main customers are companies and the public sector. There is already a first customer there: The state of Baden-Württemberg wants to support its employees with the AI solution. The AI system is called f13 and is designed to help analyze documents, automatically process applications and evaluate data, for example. The data protection standards are correspondingly high, and as a German start-up, Aleph Alpha promises to comply with them.

Two new LLMs for PhariaAI

Of course, the AI platform is also based on large language models (LLMs). These can be selected in each case. There is a proprietary LLM, called Pharia-1-LLM, as well as various open source models from other providers, such as Meta or Anthropic. You can then feed them data from your own company. PhariaAI also comes with ready-made components: “fine-tuning and evaluation, debugging capabilities, unique proprietary innovations from Aleph Alpha for the explainability and control of language models and for the traceability and auditability of AI systems.” Pharia-1 is available in two versions, Pharia-1-7B-control and Pharia-1-7B-control-aligned, with control designed for handling sensitive data. Both models are optimized for seven languages, including English, German, French and Spanish.

PhariaAI also runs both in its own data centers and in the cloud. According to Aleph Alpha, this should enable computing power to be outsourced to the cloud during peak times, for example –, but not critical and sensitive tasks. Jonas Andrulis writes on the blog post: “PhariaAI is the result of our strategy for innovation and sovereignty and represents a turning point for companies and administrations that want to use AI without jeopardizing their compliance or losing their own sovereignty.”

In the blog post, the Heidelberg-based company also assures that both LLMs have been developed and trained in “compliance with applicable EU and national laws, including copyright and data protection laws”. In addition to the model with the so-called weights, Aleph Alpha also publishes the source code for training. Everything is freely available for educational purposes.

(emw)