Use of Generative AI & Foundation Models: What needs to be considered?

2024-03-01
Technology
Henning Jostes

Generative AI dominated the AI market in 2023. Associated terms such as Large Language Model (LLM) and Foundation Model were on everyone's lips and were perceived as a technological breakthrough.

Large language model refers to a specific category of models that is characterized by the fact that it contains billions of parameters and a great deal of knowledge. This knowledge always relates to an understanding of language, but can also contain any number of other fields of application. LLMs are often used to generate language (so-called generative AI), e.g. to create AI assistance such as ChatGPT. LLMs can be used as foundation models.
The term foundation model refers to the use of a model as a building block for the creation of a new, more specialized model. Any model (whether LLM or another category) can be used as a foundation model and specialized by fine-tuning. This process is known as 'transfer learning', as the knowledge from the foundation model is transferred to a new model and supplemented with new specialized knowledge. The concept behind foundation models has been around for longer than LLMs. However, the importance of foundation models has increased significantly with the emergence of LLMs, as foundation models and transfer learning offer the opportunity to build new solutions on the broad knowledge of LLMs.

In this blog article, we contrast Generative AI in the form of 3rd party LLMs and specialized, customized Foundation Models and discuss what this means in the context of the productive usability of AI technologies for intelligent document processing and workflow automation, especially in regulated markets.

Challenges in the usability of generative AI and 3rd party LLMs

Thanks to their ability to understand human language and content and to generate new content, these technologies have the potential to automate tasks in a business context, thereby increasing efficiency and reducing workloads in a variety of application areas and industries. While many companies recognize the immense opportunity that AI offers, they are also faced with the challenge that generative AI solutions for specific, productive application scenarios - such as document processing - are often too general and not readily usable. The idea that a single, universal AI model could automate all workflows and processes without specific adaptations is proving to be a fallacy.

A major problem is their unreliability in tasks that require consistent and precise results, such as analyzing documents. Their tendency to generate creative and varying solutions, which is certainly an advantage when writing texts, can be more of a hindrance than an advantage in such contexts. Highly specialized, complex tasks such as reading out specific document information that has a special technical meaning is usually still a challenge for generic open source LLMs.

The lack of guarantee for the quality and timeliness of the generated content harbors further risks. Although generative AI systems can provide extensive information on a wide range of topics, they do not provide transparent information about the quality of the results. This can lead to misinformation or even inappropriate and discriminatory responses.

Added to this is the uncertainty in the regulatory environment, particularly in Europe, where the European AI Act will have an impact on the use of generative AI technologies. This poses a significant risk for companies that rely on secure and compliant solutions. Transparency and a secure legal framework are essential, especially in officially regulated industries such as banking and insurance, which also work with a lot of sensitive, personal data.

The lack of transparency regarding the training data of the AI models reinforces these concerns. Open source AI solutions that have been trained with a large amount of data available on the internet do not always produce unbiased or appropriate results if, for example, biased or discriminatory content is part of the training data. It usually remains unclear on what basis decisions are made and content is generated.

In view of these challenges and risks, it is clear that the use of generative AI solutions is not a one-size-fits-all solution for fully exploiting the potential for automation and efficiency gains. For more specific uses of AI technologies - for example in document processing - other requirements apply in terms of transparency, quality of results, productive integration into workflows and political framework conditions.

An AI model alone is not enough. The following questions arise:

Who actually trains the model?
What about copyright?
Are the results accurate enough?
Where does all the data come from?
What do I have to consider under the EU AI Act?
How do I integrate and operate this?

Use of specialized foundation models in a platform environment

Although Large Language Models (LLMs) and Foundation Models offer an impressive basis for understanding and generating content, they only develop their full potential with additional, specifically adapted modules and services.

Buildsimple relies on a highly specialized foundation model that was designed for the application area of intelligent document processing and thus goes beyond the limits of general LLMs.

In order to further improve the user experience and the quality of the results of the AI solutions, the Buildsimple platform is already being expanded to include functions from the field of generative AI. This expands the range of services and functionalities of the platform and allows customers to benefit from the strengths of generative AI in addition to the already available foundation model for intelligent document processing.

Independent training of AI models and transparency in the process

One of the key benefits of the Buildsimple solution is that customers have the opportunity to train their train their AI solutions themselves. This not only ensures that the usage rights and the intellectual property remain with the customer, but also offers full control and transparency over which data is used to train the AI. This autonomy is particularly important in areas where specific, proprietary or sensitive data is used.

The opportunity, to validate AI solutions professionally and train them themselves also enables companies to achieve high quality results for specific use cases achieved. By tailoring AI to the individual needs and requirements of the company, tailor-made solutions can be developed that are precisely aligned with the respective challenges and goals.

Buildsimple attaches great importance to the easy integration of AI solutions into existing systems. Thanks to standardized APIs, companies can seamlessly integrate the AI functionalities into their workflows, which significantly reduces the implementation effort and increases user-friendliness. Other functionalities such as central access management and the option to integrate a 'human-in-the-loop' also contribute to secure integration into workflows.

Another important aspect is the individual support and advice by our experts. From personal onboarding and training for your individual AI solution to the option of booking professional services and support to help you with implementation and questions. This personal approach at Buildsimple ensures that customers receive support throughout the entire process - from conception to implementation and beyond.

Adherence to compliance, information security and data protection regulations is particularly important in highly regulated industries such as banking and insurance. Buildsimple has designed its solutions to meet the strict requirements of officially regulated companies such as banks and insurance companies, enabling them to be used in such environments - including the processing of health data.

Finally, the active cooperation of the Federal Expert Commission for Artificial Intelligence of the Economic Council and the Federal AI Association plays an important role. This commitment helps Buildsimple not only to understand and evaluate the latest developments, technologies and trends in the field of AI, but also to actively shape and integrate them into its solutions. This also ensures that the Buildsimple platform meets the needs of customers and is always at the cutting edge of rapidly evolving AI technology.

Conclusion and outlook

The development and integration of Generative AI is at a crucial point. In order to offer powerful, customized AI solutions, it is important to leverage the strengths of Gen AI, but to use them in a secure environment. This includes specific adaptations to individual use cases and integration options as well as adherence to strict compliance, security and data protection standards. The ability to train AI solutions themselves and have complete control over the data used offers companies transparency and security.

In the future, this holistic approach to AI solutions will be crucial in order to fully exploit the potential of artificial intelligence for companies in a wide range of industries and application scenarios.

If you have any questions, please contact us at
:

René Weseler

Senior Executive Manager, Buildsimple

Member of the Economic Council and the
Federal Commission for Artificial Intelligence