A flying start into artificial intelligence

In this article you will read,

• what possibilities exist in the cloud for developing artificial intelligence,
• how pre-trained basic models can support you in the process
• and how you can access a large language model via Github.

Generative AI (GenAI) was on everyone's lips in 2023 – and sparked people's imagination (especially that of non-experts). Many companies are fantasizing about various application scenarios for artificial intelligence (AI) and are painting a science fiction world in which super-smart assistants make decisions largely autonomously and work side by side with humans.

Many current use cases are not actually all that spectacular. Nonetheless, they bring a considerable increase in productivity, relieving employees of tedious tasks that tie up productive time. These include summaries of documents (collections) or meetings, searching for relevant information in the depths of company filing systems, pre-filling forms, or creating program code. And companies can already leverage this efficiency potential today – even without extensive in-house AI expertise. 

According to PAC analysts, a lack of expertise is currently the main hurdle on the path to the world of AI. But even companies that have the relevant expertise are facing challenges such as: Where do we get the necessary infrastructure resources to develop and train our own models? This is not the only area where the cloud can make a decisive contribution. Developers are provided with urgently needed resources such as GPUs at short notice and in line with demand. Training AI models is one of the most sensible use cases for the cloud.

The starting point for the use of AI therefore varies from company to company, although the boundaries are blurred. There is now a comprehensive portfolio of AI offerings, particularly basic models such as large language models (LLMs) that companies can use.

SaaS offerings provide standardized services that can be adapted to the specifics of each company, and IaaS offerings are aimed at in-house developers who want to minimize the effort involved in managing infrastructure resources. Somewhere in the middle are PaaS services that offer frameworks for AI development, pre-trained basic models or specific AI solution modules. 

Together with its partners, the Open Telekom Cloud illustrated approaches to the topic of AI at various events. The focus was on how users can quickly achieve functional AI with the help of existing (open source) services. Infrastructure resources from the cloud always play a role here – but they are hidden “under the hood”. What is often overlooked: not only does the development of a basic AI model require large GPU resources, but the operation of an AI service also needs a powerful basis (also handled with GPUs). 

“The AI strategy must fit the company. The services that are obtained from a cloud should be planned accordingly,” explains Holger Schultheiß, Chief Product Owner of the Open Telekom Cloud. With the IaaS approach, only infrastructure resources are obtained, e.g., to develop or fine-tune an own highly specialized AI model. However, this is only worthwhile if there is specific outstanding added value and sufficient suitable data is available. In any case, advice from an experienced partner is important for planning and implementation.

This also applies to a less differentiating “85% solution”. “As a rule, you don't need a specially trained LLM for this, but a well-developed system around an (existing) LLM,” explains Dr. Tim Delbrügger, AI lead at iits Consulting, “We usually clarify whether this is sufficient in initial workshops, where the customer's needs, the data situation and technical implementation options are evaluated.” “The greatest added value is usually generated added value through the rapid, direct integration of pre-trained AI into existing business processes,” summarizes the AI expert. 

The SaaS approach is the opposite. It is aimed at pragmatists who want to quickly utilize the potential of existing standard solutions for their company. One example of this is an AI-supported enterprise search (such as that from amberSearch). It is essentially based on two components: a pre-trained LLM (Large Language Model) with a basic ability to chat is linked to a RAG (Retrieved Augmented Generation) functionality. When the LLM “chats” with a user, it also accesses internal company data for this discussion and thus allows a discussion in the specific company context – without the internal data being included in the model. Such a service, e.g., for finding information within the company, can be implemented in just a few hours.

“PaaS offerings are the all-weather tire of AI,” says Schultheiß. Based on pre-developed and pre-trained basic modules (e.g. LLMs), customized AI solutions can be developed that are specifically adapted to the company's requirements, e.g., for confidentiality, without any compromises. In addition, seamless integration into business processes is possible by connecting various third-party systems, while an LLM hub gives developers access to basic models that can be easily addressed via API and added to infrastructures via Terraform. 

“By using our ready-made Terraform modules, a company can get started very easily, select an LLM and, for example, talk to a chatbot with its own data in its own infrastructure,” summarizes Delbrügger, AI architect.

“Ultimately,” says Schultheiß, “everything currently revolves around the correct use of the existing basic models. They can be further trained, fine-tuned, or enriched via RAG. There are various deployment models that are geared towards the security requirements of companies. For example, if a company wants to develop code using an LLM, a private installation or access to a hardened version in a sovereign cloud such as the Open Telekom Cloud is worthwhile. This ensures that the data used remains in-house and that the results generated remain the legally secure intellectual property of the user.