Part 3: Crafting AI – Building chatbots

This post is part of the AI Apprenticeship series:

• Part 1: AI Apprenticeship 2024 @ DiploFoundation
• Part 2: Getting introduced to the invisible apprentice – AI
• Part 2.5: AI reinforcement learning vs human governance
• Part 3: Crafting AI – Building chatbots
• Part 4: Demystifying AI  
• Part 5: Is AI really that simple?
• Part 6: What string theory reveals about AI chat models

By Dr Anita Lamprecht, supported by Diplo AI and Gemini

In week 3 of the AI Apprenticeship online course, each participant began to build their own functioning chatbot. For this purpose, we had to use a topic we are familiar with to ensure that we can later verify the accuracy of our chatbots’ output. We also needed to collect and upload datasets, compose system prompts, and choose a large language model (LLM) for each bot.

All these steps need to be taken with care. As creators of the chatbots, our choices serve as boundaries for our chatbots, influencing their level of unpredictability, often referred to as ‘hallucinations’, though I prefer the term ‘randomness’ as it seems more precise. Randomness more accurately reflects the way these unexpected outputs arise from the complex interplay of data and algorithms (I have explored the concept of randomness in more detail in part 2.5 of my blog series). But let’s return to our task.

Purpose is personal

Nature generously designs everything for us, but in technology, we must carefully craft our creations. Designing our bots starts with giving them a purpose. I am creating two bots, and their purpose, quite frankly, is personal. They will both serve my personal aims, but I hope the final result will also be useful to others. 

So, what are their purposes? My bots are designed to help me overcome my human limitations. Rephrased more positively, their purpose is to extend my capabilities. More concretely, they will enable me to keep pace with the speed and scope of publications in my field of research. 

How? By enhancing my naturally limited processing capabilities and memory. One bot will be dedicated to preserving and expanding my acquired knowledge of the topic ‘child safety in the metaverse’, whilst the other will help me sift through the flood of documents recently published by ITU as part of the UN Virtual Worlds Initiative. To put it succinctly: these bots will be an extension of me. They are, after all, artefacts. 

Let us begin.

The relation between system prompts and LLMs

The existence of our bots begins with providing system prompts. In the Geppetto metaphor I used in my first post, these prompts would be equal to taking a piece of wood and starting to chisel the outlines of Pinocchio. Experienced carpenters know that each piece of wood comes with its own character, expressed through its grain and other qualities. Our system prompts serve as the foundational blueprint for our chatbots, defining their core purpose and guiding their development. They are like Geppetto’s initial vision of Pinocchio as a puppet with the potential to become a boy. LLMs are where the magic happens.

Think of an LLM as the force that animates Pinocchio by enabling him to speak in a human-like way; it provides the foundational language patterns and understanding, drawn from a vast forest of information. In our case, it’s about statistics and pattern recognition, not magic. 

Data and the forest of information

In Part 2 of this blog series, we explored how AI interprets data by identifying patterns and correlations, similar to how flags convey meaning through colours and symbols.  Now, imagine the internet as a vast forest filled with different kinds of wood, each with its own texture and strength. This is the raw material from which our chatbot will be carved.

Our carefully curated dataset is like a selection of high-quality wood blocks, chosen for their specific properties. This ensures that the chatbot learns from reliable and relevant information. The LLM, in turn, helps the chatbot ‘see the forest for the trees’, revealing hidden connections within this vast amount of data and assembling them into a meaningful structure. It guides the chatbot to navigate the forest and uncover the insights most relevant to our needs.

DiploAI as our workshop

While datasets ensure high quality and relevance for effective machine learning, data from the internet provides a broad range of language patterns. However, this data requires significant filtering to remove noise and inaccuracies. DiploAI provides the workshop, the tools, and the environment for shaping this material. And we, as sculptors, use prompts as our chisels, actively shaping the wood (data) and carving out the desired form and features.

But what is this ‘workshop’ exactly? Well, DiploAI is an AI platform developed by Diplo’s AI and Data Lab. It focuses on exploring the possibilities of machine learning, neural networks, and natural language processing algorithms. DiploAI is actively involved in finding new datasets, testing different machine-learning models, and sharing its newly acquired knowledge through a weekly diary. It plays a significant role in enhancing the understanding of AI and data science in the context of diplomacy and global governance. (If you want to learn more about DiploAI and its activities, you can visit the following link: Diplo AI.)

AI learns through experience

Just as Pinocchio encounters various characters who influence his path – Honest John, Geppetto, or the Blue Fairy – users play a crucial role in shaping the chatbot’s development. Their interactions, feedback, and even misunderstandings provide valuable insights for refining the chatbot’s system prompts and datasets.

For example, frequent misunderstandings or irrelevant responses signal the need for adjustments in the chatbot’s system prompts or datasets. Diverse feedback from users can also reveal new functionalities for the bot. In addition, users can benefit from learning more about the skills they need to acquire to communicate with and use emerging technologies.

Tailor-made AI for diplomacy and global governance

What makes Diplo’s approach to AI so outstanding is that the workshop is independent of the individual LLM. We can make our Pinocchio from oak, mahogany, or pine at the click of a mouse. We could say it’s tailor-made, not off-the-shelf. This also means we retain ownership and control over our AI, unlike relying on pre-built solutions from big tech companies. While I’ve initially chosen a particular underlying LLM to bring this blueprint to life, the bot itself is not bound to this specific model. 

As the participants of the AI Apprenticeship course include those in diplomacy, global governance, and international relations, DiploAI provides a few important elements:

• DiploAI domain specificity: DiploAI is designed to understand and analyse texts specific to diplomacy, global governance, and related fields. This domain specificity allows it to provide more accurate and relevant insights tailored to the needs of diplomats, policymakers, and researchers in these areas.

• Customisable AI assistant: DiploAI enables the creation of customisable AI assistants that can be fine-tuned to address the specific requirements of diplomatic training, research, and courses. This customisation ensures that the AI assistants can effectively support and enhance the learning and decision-making processes in the realm of diplomacy and global governance.

• Retrieval-augmented generation (RAG) technique: DiploAI uses advanced techniques to combine the strengths of curated datasets and LLMs, allowing it to generate more accurate and relevant responses in diplomatic settings.

• Enhanced learning and insights: By being tailor-made for diplomatic and policy contexts, DiploAI can provide deeper insights, analysis, and recommendations that are specifically relevant to the challenges and opportunities faced in the field of diplomacy and global governance.

Overall, the specialisation and customisation of DiploAI make it a valuable tool for supporting diplomatic activities, enhancing research in global governance, and facilitating informed decision-making in the complex world of international relations.

Like Pinocchio, the essence of my chatbot lies in its system prompts, which allow it to be adapted and re-implemented with different LLMs as technology evolves. This ensures its longevity, adaptability, and the much-needed independence in the ever-changing landscape of AI. 

Elements in designing a chatbot

Let us look at the various roles that shape our relationship with AI:

1. The apprentice (Us): The creator, craftsperson, sculptor, or Geppetto. We are the driving force behind the chatbot’s creation. We have the vision (purpose), provide the guidance, and actively shape the bot’s capabilities using our prompts.

2. The system prompts: The sculptor’s chisel or blueprint. This is our primary tool for shaping the LLM. It defines the bot’s purpose, guides its behaviour, and sets the boundaries for its actions. It’s a dynamic set of instructions that evolves through interaction and feedback.

3. The large language model (LLM): Provides the foundational patterns and understanding of language that enable the bot to function and communicate in a human-like way.

4. The data: This is the knowledge that shapes the chatbot’s response. In our metaphor, data presents the trees within the forest of information, either carefully selected as a dataset or drawn from the vastness of the internet. 

5. DiploAI: The workshop for designing out chatbots. This is the environment where we build and refine our bot. It provides the tools, resources, and infrastructure for interacting with the LLM and shaping its behaviour through our prompts.

6. The chatbot: The workpiece or Pinocchio. The bot is the product, the embodiment of our vision brought to life through the interaction of our prompts with the LLM within the DiploAI environment. It is an entity with specific capabilities and a defined purpose, designed to be an extension of our own abilities. However, it is not a static creation; it evolves and adapts as we refine our prompts, experiment with different LLMs, and curate the data it learns from. This continuous refinement process allows the bot to become increasingly sophisticated and aligned with our goals.

7. The users: The chatbot’s partner in the conversation. Humans are part of the dynamic feedback loop, where their interactions with the bot not only refine its knowledge but also deepen their comprehension of AI’s potential.

We can just build a chatbot to achieve a quick win, but if we – as legal professionals and experts in governance – eventually want to master the challenges of AI governance, we need to be able to comprehend our mutual roles profoundly. For me, this apprenticeship is about developing a common sense of the abilities and limitations of both the technological and human aspects within our socio-technological system. 

This is particularly challenging when dealing with the intangible nature of AI, where code and algorithms replace wood and chisels. Even with the helpful metaphor of Pinocchio, bridging the gap between the tangible and the intangible requires a shift in thinking. 

The AI Apprenticeship online course is part of the Diplo AI Campus programme.