A Universal LLM Does Not Exist

If you follow TEK2day you know we believe that language models will proliferate and that there will never be a time when one model serves all use cases. There is no Universal LLM.

I’ll give you a personal example. The EPS Call transcript analytics service that we’ve rolled out (HERE, let us know if you want a one-off analysis), is built on top of Anthropic and OpenAI. If we decided to roll out a service that would allow for batch analytics where the application analyzed 100 or more transcripts simultaneously with an appreciation for all the nuances that are specific to earnings calls, neither OpenAI nor Anthropic could deliver. FinBERT’s language model could deliver, as it has been trained to analyze financial text (including earnings calls) to deliver sentiment analytics. FinBERT is a domain-specific language model and there will be many more such models across industry-verticals.

The knowledge base that models are trained on is a critical component to understanding a model’s capabilities, along with understanding how the language model processes information (inference).

80% of the models in production over the next 5 years will be small language models, trained with a specific use case(s) in mind. Many of those efforts will be open source models.

We have many related writings on LLMs. Check out Rise of the SLMs: HERE

---