Tilde Reaching New Heights in Artificial Intelligence: Developing a ChatGPT Alternative for Europe

Machine translation systems, virtual assistants that help people communicate with state and municipal institutions, voice recognition and voice synthesis – Latvian language technology company Tilde does it all. Now, one of its main areas of focus is developing an innovative large language model, similar to the popular ChatGPT, for Latvian and other European languages.

Last year, Tilde gained important recognition as it won the grand prize in the Large AI Grand Challenge, an artificial intelligence competition held by the European Commission. The prize consisted of not only EUR 250,000, but a year of access to LUMI, the most powerful supercomputer in Europe, on which very large language models can be developed.

Tilde took part in the competition with the aim of creating TildeLM – a large language model focused on Latvian and other European languages, which have so far been lacking in support in artificial intelligence (AI). According to the company, global AI developers mostly concentrate on English and the other most widespread languages of the world, therefore solutions like ChatGPT produce significantly better results in these languages as compared to less popular European languages. To promote the use of AI in Latvian and other European languages, the new model is being developed based on open science and open-source code principles, meaning that it will be freely available to researchers and developers of commercial solutions.

The prize received from the European Commission and the LUMI supercomputer capacity will allow Tilde to develop a particularly powerful language model which will facilitate the inclusion of Latvian and other European languages in AI solutions. 

The new model is based on neuron network technologies and contains over 30 billion parameters. To develop a model of this scale and its compatibility with solutions like ChatGPT, computing power of the highest level is needed, and this is provided by the LUMI supercomputer. Ranking number one in Europe and number five in the world in terms of power, the LUMI supercomputer is in Finland. It is powered by renewable resources and the heat it produces covers approximately a fifth of the region’s heating needs.

TildeLM, currently under development, includes all European languages with a special focus on Baltic and Eastern European languages which have hitherto been under-represented in AI technologies. This model will be useful in adapting state administration electronic solutions in various languages and will allow for widely used AI solutions to be developed, such as translation systems, virtual assistants, text generation tools, voice technologies etc.

There are currently ten researchers working on the TildeLM project, five of whom hold doctoral degrees in science. Collective memory institutions (libraries, archives and other organisations that maintain public collections of knowledge) have provided multi-language data for training the model. As of the end of last year, about 3.3 trillion words had been collected for the model.

The aim of the Large AI Grand Challenge is to support the most promising AI businesses in Europe and promote European competitiveness in this field. Out of 94 participants, Tilde was able to best demonstrate its ability to develop large language models that can compete with global solutions.

Tilde is among the leading language technology and localisation companies in Europe. By developing AI, the company is reinforcing the viability of languages in the digital age. Solutions developed by Tilde, including machine translation, voice recognition, voice synthesis and virtual assistants, drive forward the development of language technologies throughout Europe. The company is actively collaborating with the top European research centres and promoting research of AI.