Peter Gentsch:
"The New Game Changer for Businesses is Generative AI"

How to Choose the Most Appropriate Generative AI:
LLMs like GPT-4, Llama and Falcon – in Search of the Holy Grail

One of the most important questions in the discussion about generative AI is the search for the “right” Large Language Model (LLM). The complexity of this task quickly becomes clear due to the different criteria and dimensions: open source vs. proprietary, public LLM vs. own LLM, on-prem vs. SaaS. Furthermore, stability, costs of training and use, vendor login, controllability and data protection are other important decision parameters.

The Foundation Circle discussed this question intensively with experts and companies developing and applying LLM. Of course, there is no single answer. Depending on the objective, budget and general conditions, the appropriate LLM has to be selected.

The illustration “On the way towards the right LLM” tries to give some guidance in the selection process (based on Ran Ding, 2023):

The two extremes open up the solution space: From the unfiltered use of the well-known frontier models such as ChatGPT to the development of one’s own language model from scratch. The cost gap underlines the extremes – while existing models such as GPT can be used free of charge or at comparatively low cost, the development of own language models usually requires an amount in the double-digit million range. There is a continuum of solutions in between. Successful projects show that existing open source models such as Llama, BERT or Falcon can be adapted to specific companies and domains via various optimization phases (prompting, embeddings, fine-tuning) to achieve a high degree of quality and maturity. These adapted models do not cover the breadth of general world knowledge and tasks as the “Swiss Army Knife GPT-4”, but they are much more focused and tailored, more stable and more controllable (in terms of quality and costs) for specific tasks (“shifting from Swiss Army Knife to set of scalpels”).

The yellow solution point represents a good compromise for entrepreneurial practice. Via prompting and additional vectorized knowledge, smaller open source models (low/medium number of parameters) with focus on specific tasks and domains are developed via transfer learning. The great advantage of these models is the comparatively low cost of training and use with acceptable quality. In addition, these open source models can be safely run and owned by companies.

Having this said, we will see different LLMs exist in the upcoming future. The large frontier model for broad and public use, adapted models for focused and efficient use in companies – whether fed by innovative open source models or OpenAI models. With the dramatically fast development and innovation cycles, companies are well advised to proceed largely model agnostic. Stable pipes consisting of clearly defined tasks that build flexibly on the respective LLM are increasingly becoming a success factor.