Generative AI is a popular term in the business world, and the underlying technology, Foundation Models (FMs for short), is gaining traction due to its widespread use and adoption. The technology is of great value to organisations around the world as it enables them to gain a competitive advantage and accelerate business operations. However, this area is fraught with a lot of jargon as terms are often mixed up, confused and misused. Therefore, this blog post serves as a one-stop solution to clarify the basics of FM and determine its applications for your business.

What are foundation models?

Foundation Models are large AI models that have been trained on extensive data sets and are used for a variety of tasks in different domains can. They are a basic building block of modern AI development on which more specialised applications and services can be built, revolutionising the way AI systems are developed and deployed.

The term foundation model became particularly important with the emergence of large language models such as GPT-3 (Generative Pre-trained Transformer 3), developed by OpenAI and used in ChatGPT. These models have a deep understanding of natural language and are able to generate human-like texts in response to user input. For example, they can hold conversations on the topic of "Tips: Travelling to Istanbul" as well as "Importance of EBIT for companies".

The success of large language models such as GPT-3 demonstrates the potential of foundation models to serve as versatile AI systems. In general, the more general the task, the better the foundation models are in their immediate application. For more specific applications, it is possible to fine-tune the pre-trained models using additional (customised) data. This approach represents a break with the traditional paradigm of developing and training a completely new, specialised model for each specific task. This saves development time and reduces the risk of in-house development.

Training your own foundation model is difficult due to the large amount of data required and the high number of parametersusually complex and costly. Especially for the known language modelse.g. ChatGPT, require large servers with high computing capacity. They are therefore usually developed by the large US tech companies, Microsoft, Alphabet, Facebook, etc., and their AI research companies. Some of these companies make them available on the internet free of charge after development (open source). Developers can thus access foundation models and use them directly or, if necessary, retrain them for specific tasks. Complete in-house development is therefore no longer necessary.

Large language models are transforming interaction with technology and expanding its application from content creation to customer service. Our overview presents 14 relevant representatives in detail:

The 14 Top Large Language Models: A Comprehensive Guide

Properties of Foundation Models

Domain-unspecific

Foundation models are by definition large-scale and universal. They are trained on huge data sets, some of which span multiple domains, which enables them to develop broad knowledge and diverse capabilities. In contrast to traditional models, which are narrowly specialised for a specific task, foundation models can be adapted and fine-tuned for various downstream applications. This makes them extremely versatile and efficient.

Self-Supervised Learning

One of the key factors for foundation models is self-supervised learning, a Machine learning technologywhich enables models to learn from data that has not previously been elaborately processed by humans. By utilising the inherent patterns and structures within large datasets, these models can learn rich representations and extract meaningful information without having to rely on (many) costly classifications or descriptions (annotations) added manually by humans.

Transfer learning and fine-tuning

Foundation models utilise the concept of transfer learning, where knowledge gained through pre-training on a large, general data set is transferred and fine-tuned to specific tasks or domains. This approach reduces the need for extensive, task-specific training. Training data and enables rapid adaptation to new applications.

Multimodality

Multimodality is the future of foundation models. Some foundation models are already characterised by their multimodality, which means that they are able to process and understand different types of data simultaneously. This includes text, images, videos and audio data in particular, but also time series or IoT data, for example. This ability enables the models to combine information from different sources and derive more comprehensive and accurate results. Multimodal Foundation Models can thus better grasp complex relationships and support more diverse applications, which further increases their applicability and efficiency. The use of models that have been trained with texts also enables user-friendly interaction in natural language, which makes AI applications accessible to a broad user group.

Learn how large language models such as ChatGPT are improved through the use of Reinforcement Learning from Human Feedback (RLHF).

Reinforcement Learning from Human Feedback in the Field of Large Language Models

Examples of foundation models

There are different FMs for different applications. To make it easier to understand, below you will find the origins, functions and areas of application of some of the most common Foundation Models:

• GPT-3Generative Pre-Trained Transformer (GPT) models are language models pre-trained on huge amounts of text (mainly from the Internet). They can generate texts on all kinds of topics in natural language on the basis of a user's question. They can also summarise texts. GPT models can be used particularly effectively by companies to generate initial drafts for new texts, automate repetitive tasks, generate code for software development and as Chatbot answer customer enquiries.
• LLaMALLaMA from Meta AI is also a large language model like the GPT models, but unlike the ChatGPT models from OpenAI, for example, it is available as open source code and therefore free of charge. Companies can use the LLaMA model for customer service, for example, law firms can search for specific legal information and e-commerce companies can use it to create product descriptions.
• CLIP: The CLIP (Contrastive Language-Image Pre-Training) model is trained on huge amounts of images and corresponding text descriptions. It enables the understanding and linking of images and texts. Text descriptions can be generated for images and, as in the DALL-E application, images can be regenerated from text input. In addition to the creation of images, DALL-E also enables the editing and refinement of existing images. The model is known for its applications in the creation of images for marketing campaigns, in the creation of storyboards for games and films, and in the generation of ideas for physical products.
• SeamlessM4TSeamlessM4T from Meta AI is a multilingual, multimodal AI model that can be used to translate spoken language (audio) into text and vice versa. Depending on the task, the model can perform translations for up to 100 languages.

Discover the impressive ability of new AI models to create realistic images from text that are almost indistinguishable from real works of art.

"Content is AI-NG - text-to-image generators at a glance" - Alexander Thamm GmbH

Choosing the right foundation model

The following is a suggested strategy for selecting FM for your organisation:

1. brainstorming about value and benefits

When deciding on Foundation Model for your organisation, consider the value it brings to your business operations. This will help you identify your rationale for choosing FM with benefits for your organisation. Below are some suggested motivations for choosing the right Foundation Model for your organisation:

1. Accelerating efficiencyIncrease the efficiency of your organisation by automating tasks and enabling your employees to invest time and energy in strategic decisions.
2. Improve the decision-making processImprove decision making by using FM to gain business insights.
3. Improve the customer experienceBetter services for customers through the use of FM to personalise content.
4. Developing better productsUtilise the possibilities of FM for the development and introduction of new products.

2. identification of business requirements and objectives

You need to determine what you need the Foundation Model for by assessing the resources and budget. Based on the current state and feasibility, you need to source and prepare accordingly. Below is a short list of key factors to consider when choosing the right Foundation Model for your organisation:

1. Technical requirementsEvaluate the current state of your technical infrastructure and capabilities. FMs require large amounts of computing resources and data infrastructure. You need to assess whether you can provide or procure this type of data storage and processing. If not, you need to consider possible partnerships and collaborations with companies that specialise in such services.
2. PersonnelIf you want a suitable FM for your company, you also need employees who can select, maintain and implement the FM. This could include Data Scientists, Data Engineers for machine learning or NLP specialists. Determine the employees within the company and hire more if necessary.
3. Costs: FM training itself is expensive, so access to Foundation Models is expensive for an organisation. As with any other technology, the number of features and applications that FMs offer determines their price and value. FMs with more features are expensive, but they are also more generalised. Therefore, it is best to use large FMs to validate your Minimum Viable Product (MVP). Once your MVP is validated, you can use smaller models that are cheaper and tailored to your specific business application with more ease and profitability.
4. Latency timeCompanies take a structured approach to product launches and often have limited time. FMs vary in the time they take to train and deliver the desired results. Choose a model that prioritises speed if you need a business application that requires quick responses.

3. define areas of application

You need to know what you need the Foundation Model for. While there are many uses for FMs, it's best to be clear about what you need the Foundation Model for so you can choose the right one for your organisation, depending on what purpose you want it to serve. Here are some possible uses for a Foundation Model:

1. Content creationFMs are a powerful technology for creating business content through persuasive marketing copy, writing product descriptions for e-commerce websites or creating business reports based on meeting summaries.
2. Provision of customer serviceFMs enhance the capabilities of chatbots by generating human-like responses, and with some fine-tuning, the model can improve sentiment analysis and provide empathetic responses for customers.
3. Product developmentFMs can accelerate product development by analysing customer reviews from websites, research findings and social media data to improve products and bring new products to market.
4. Research and developmentFMs improve data analysis by analysing huge amounts of data that can form the basis for scientific research.

The proposed strategy provides a guide for choosing the right Foundation Model for your organisation. Once you are clear about why you need FM and what you need it for, you can make an informed decision. Choosing the right FM for your organisation will lay the foundation for its value and help you gain a competitive advantage.

In our article, we analyse how the CLIP model performs after fine-tuning to specific datasets compared to traditional models such as ResNet50.

Fine-Tuning the CLIP Foundation Model for Image Classification

Basis for efficient business processes

Foundation Models are effective tools for generative AI, and their applications are transforming organisations and industries. By further fine-tuning FMs, companies can customise FM training data to automate repetitive tasks, develop products efficiently, improve customer support and create compelling business documents. Although both FMs and LLMs drive the business applications of generative AI, both contribute differently, as FMs are more general-purpose, while LLMs are specialised for text only. Foundation Models contribute significantly to efficient business processes. However, organisations need to be careful when using them by first developing a strategy for selecting the right foundation model for their use cases.