German start-up : Aleph Alpha introduces new generation of artificial intelligence

 

 

“The Heidelberg start-up Aleph Alpha announces an improved version of its AI models. This should not only provide better answers - but above all be able to explain the sources on which they are based.

The Heidelberg start-up Aleph Alpha is launching a new generation of its language model. A so-called "Control" version is now available for each of the company's models. These "Control" models are able to provide significantly better answers and, in initial evaluations, are neck and neck with the best models on the market, founder Jonas Andrulis announced to the F.A.Z. at. The models are optimized for better processing of natural language and solving "highly complex" tasks in computational linguistics.

Similar to Open AI, Aleph Alpha has developed ChatGPT, an application that can use artificial intelligence (AI) to answer a wide variety of queries.

  Unlike ChatGPT, however, the offer is aimed at industry and public administration and less at end consumers. For example, the company has developed an AI assistant for employees of the state administration in Baden-Württemberg.

Competitive advantage over American competition?

A step forward in the new model generation is the change from so-called "few-shot prompting" to "zero-shot prompting". Prompts are the inputs that applications like ChatGPT respond to. Until now, most large language models have had to be fed a series of examples in order to be able to solve tasks for which they were not actually trained. With zero-shot prompting, the models only need a single instruction to generate robust responses. As a result, the new models are optimized to solve tasks for which they were not explicitly trained, advertises Aleph Alpha.

In addition, it is the first model generation that integrates functions for source verification and the explainability of AI-generated content. In April, Aleph Alpha announced that its language model could now use an extension to indicate which text passage in a source the answers generated by the system are based on and which sources may contradict it.

Andrulis sees this, together with data protection, as a decisive competitive advantage over the competition from America, especially with regard to the forthcoming AI regulation of the European Union and its requirements for explainability of AI models. Source verification is necessary "so that people in demanding areas such as law, health and banking can continue to make responsible and understandable decisions," says Andrulis.

Aleph Alpha is regarded as a German beacon of hope for artificial intelligence. According to reports, negotiations are currently underway for a new round of financing of over 100 million euros. The company valuation should then be between 450 and 500 million euros. According to information from the F.A.Z. the software group SAP is also to negotiate with the start-up about becoming a strategic investor. The chip companies Nvidia and Intel should also show interest.”