From the source
Aleph Alpha introduced a groundbreaking tokenizer-free (T-Free) LLM architecture that enables superior efficiency and effectiveness for fine-tuning and customization of AI across different languages, alphabets and specialized industries.
This innovation addresses the limitations of conventional LLMs and unlocks new possibilities for sovereign AI solutions for governments and enterprises.
The collaboration with AMD and Schwarz Digits strengthens Aleph Alpha’s new LLM architecture with high-performance computing and a sovereign cloud solution.
Aleph Alpha, a leading AI technology solutions provider headquartered in Germany, has announced a new architecture innovation for LLMs to address one of the most critical challenges in AI.
Teaching today’s popular closed- or open-source LLMs new languages or unique industry knowledge (often crucial for enterprises and governments) tends to produce underwhelming results and fine-tuning often proves ineffective.
A key reason for this is that the patterns these LLMs learn are based on the tokenized version of the text they were trained on.
If new text differs considerably from the original training data, it cannot be efficiently tokenized.
…
