Build Your own LLMs: Customization and Hosting of AI Models

IT After Work

At the recent IT After Work event in Pforzheim, I had the opportunity to share my insights on “LLMs in Eigenregie: Individualisierung & Hosting von KI-Modellen.” My presentation delved into the motivations for developing a proprietary Large Language Model (LLM). I emphasized the importance of data privacy and security, adaptation to internal company use cases, performance enhancement through access to internal data, complete control over the model, integration into other IT systems, and lower operational expenses.

I began by tracing the evolution of language models, starting from ELIZA, the first system for processing and generating natural language, to the sophisticated models of today like ChatGPT. This historical perspective underscored the transition from rule-based systems to neural language models and the advancements in instruction-tuned LLMs, particularly the significant improvements achieved by ChatGPT in following instructions reliably.

A critical part of my talk was highlighting the differences between proprietary and open-source models, as well as pretrained versus instruction-tuned versus alignment-tuned models. I discussed various factors influencing the choice of an LLM, including performance (understanding, reasoning, creativity), latency, privacy, costs, context size, dataset considerations (bias, fairness, transparency), and legal requirements.

An overview of known LLMs provided a comprehensive look at the landscape, featuring models like GPT-3.5-turbo, GPT-4, LLaMA 2, StableBeluga 2, Claude 2, Luminous, PaLM 2, and others from organizations such as OpenAI, Meta, StabilityAi, Anthropic, Aleph Alpha, Google, and Microsoft. This section helped the audience understand the diverse options available in terms of type, context window, size (parameters), and licensing.

The talk then transitioned to discussing the customization of LLMs for specific organizational needs and concluded with the concept of evolving LLMs into agents, underscoring their potential and future direction.

I thoroughly enjoyed giving this talk at the IT After Work event in Pforzheim and found it immensely rewarding to meet and engage with all the participants there. The exchange of ideas and networking opportunities provided a great platform for sharing knowledge and experiences in the field of AI and LLMs.