The Structure and Implementation of AI Tools
In the rapidly evolving landscape of artificial intelligence, understanding the intricacies of AI tools and their underlying structures is crucial. Whether you’re a seasoned design professional or just embarking on your AI journey, grasping the foundations of AI systems is essential for creating innovative, user-centric solutions. This article aims to showcase the inner workings of AI models and their structures, accompanied by practical visual examples. It serves as an initial guide for any designer looking to engage with AI, enabling them to contribute through all stages of the process to the end-user experience.
What is Generative AI?
Artificial Intelligence (AI) is the broad field of creating machines that mimic human tasks, from decision-making to pattern recognition. Generative AI is a niche within AI, specifically designed to produce new content, such as text or images, by learning from existing data. These used trained large language models (LLMs) such as GPT 3.5 Turbo, the brain behind ChatGPT. While AI replicates human intelligence, generative AI showcases the creative side of machines.
Generative AI has made significant strides in recent years. These models can now produce diverse and coherent content, including stories, news summaries, poems, lyrics, paintings, images, and even programs. The revolutionary potential of generative AI transcends numerous industries, from entertainment and journalism to healthcare and engineering, reshaping content creation, artistic expression, and problem-solving.
The Structure of AI Tools
The implementation and structure of all tools will depend on the tool or app type and the layers between the foundation model, such as a LLM, and the application itself. Some applications introduce additional filters or context requirements to ensure more precise answers, limit certain content, and reduce bias. The user might provide feedback in-app that helps the foundation LLM filter out the answers and pick a more appropriate answer for future interactions. However, this doesn’t happen in all models.
In the illustration above, users engage with an AI-powered application like ChatGPT. When users send a prompt to the app, it first passes through context or filters. These filters are a set of parameters established by developers to refine the prompt. These layers also tokenize the message to make it understandable to the LLM. After this step, the prompt is sent to the foundational LLM for processing. Once processed, the response is usually relayed straight back to the application. These applications can encompass various formats like text-to-text, text-to-image, text-to-audio, and text-to-video; users can receive outputs in any of these formats by simply inputting natural language commands into the app.
Integrating AI Into Existing Systems
We’ve discussed the structure involved when building a new app or tool. But what happens when we need to integrate AI into existing systems? While there are established guidelines and best practices to expedite this process, the fundamental structure remains consistent with what we’ve previously explained.
Let’s consider the example of integrating an AI-powered chatbot into a domain hosting company. Users will send requests in the form of text, and the system will respond in various formats depending on how the AI tool is configured. These responses could include text, images, CTAs (links), audio, and more. The key difference lies in how the models are trained or fine-tuned, including aspects such as tone of voice, chatbot expertise, and limitations. These factors are defined in the layers that provide context, enabling the chatbot to help users with specific company-related tasks.
The Importance of Foundational Models
Foundational models, often referred to as the ‘AI brain,’ are the core machine learning models that serve as the basis or starting point for various AI applications. These models are often highly sophisticated and general-purpose, like LLMs, which are trained on extensive datasets to understand and generate human language. These datasets serve as sources of training data that facilitate the formation of neural connections within the model.
Prominent companies like OpenAI offer fully trained and refined LLMs, such as GPT-3.5 Turbo, which serve as a fundamental building block for more specialised AI applications. When it comes to constructing a foundational model, there are two primary approaches. One can either train a model from scratch using bespoke datasets, a costly and time-intensive process, or opt for fine-tuning existing models by feeding it with custom data sets. Foundational models are typically highly versatile but may require additional layers or filters in the application to tailor their output to specific needs, maintain accuracy, and reduce potential bias.
Designing for AI Systems and Tools
As designers, we can take on multiple roles in the product development process. Maintaining a human-centred approach from the early stages is essential to ensure the final user experience is engaging, intuitive, and closely aligned with user needs. While proven best practices guide us in delivering seamless experiences, there is still substantial room for exploration and innovation, especially when fine-tuning or setting up the layers between the model and the app.
In AI-infused app development, designers generally focus on the end-user experience, which can be divided into different phases. Microsoft has compiled guidelines for AI human-centred design to aid in this process. The image below illustrates the four main stages of user interaction with AI. While not all guidelines apply to every scenario, this provides a sneak peek of the overall structure.
Implementing AI Design at XAM
At XAM, when our product uses an established LLM from a reputable company, which is the case 90% of the time, our primary responsibility centres around designing the user interface and ensuring smooth interactions. It’s not just about the mechanics; it’s about the overall feel of the product, its behaviour, and how it communicates with users. On the other hand, when we embark on building an LLM from scratch or adapting an open-source model, our role becomes more encompassing. This deeper involvement allows us to better address edge cases and refine responses, enhancing the user experience. It’s worth noting that the extent of our involvement can differ based on project specifics. For example, when the AI struggles to generate a response, we might design specific templates to guide user interactions.
The Future of AI and Design
As AI continues to reshape the digital landscape, designers stand at the forefront of this evolution, bridging the gap between intricate technology and user-centric experiences. The fusion of AI and design offers both challenges and opportunities. It’s not just about aesthetics anymore; it’s about crafting meaningful interactions in a world increasingly influenced by automation. Staying informed, prioritising the end user, and fostering collaboration will guide designers in navigating this AI-driven era.
- Generative AI encompasses various domains and can be part of existing systems and brand-new tools and apps.
- AI tools have intricate structures that influence output quality and, therefore, the user experience.
- AI app development demands teamwork between designers, developers, and data scientists.
- Designers must ensure fairness and transparency in AI tools.