4 Using Generative AI
Introduction to LLMs and Generative AI
What is Generative AI?
Generative AI refers to a subset of artificial intelligence systems designed to generate new, original content based on patterns and information learned from existing data. Unlike traditional AI models that focus on classification, prediction, or detection, generative AI models create data that mimics the characteristics of the training data. This can include generating text, images, music, and other forms of media. The role of generative AI in AI is to enable machines to produce creative outputs, making them capable of tasks that require creativity and innovation.
What are LLMs?
Large Language Models (LLMs) are a type ofgenerative AI designed to understand and generate human-like text based on the data they have been trained on. These models can process vast amounts of text data to perform a variety of tasks, such as translation, summarization, question-answering, and more. They play a crucial role in AI by enabling machines to interact with human language in a way that is increasingly indistinguishable from human interactions.
How Generative AI Works
Generative AI models function through sophisticated machine learning techniques that enable them to learn from and generate new data. Here’s a high-level overview of how generative AI models work:
- Training Process: Generative AI systems learn through exposure to vast datasets relevant to their intended output. For instance, a system designed to generate marketing copy could study millions of existing marketing materials, advertisements, and related content to understand effective communication patterns and styles.
- Processing Framework: The system operates through an interconnected network that processes information systematically. This network analyzes and interprets patterns within the training data, developing an understanding of structure, style, and context. The system continuously refines its understanding through this analysis, improving its ability to generate appropriate content.
Different generative AI systems are engineered for specific purposes, each optimized for particular types of content creation. Examples include:
-
Text Generation Systems process and produce written content by analyzing language patterns, context, and communication styles. These systems can create various forms of text, from creative writing to technical documentation.
-
Image Generation Systems learn visual patterns and artistic principles to create new images. They understand elements like composition, color relationships, and visual coherence to produce original visual content.
Generative AI has evolved rapidly, with each new development building on the successes and limitations of previous models, leading to increasingly sophisticated and capable systems.
Capabilities and Limitations
Capabilities:
- Content Creation: Generative AI models can create new content across various media, including text, images, music, and video. This capability is valuable for creative industries, marketing, and entertainment.
- Personalization: These models can generate personalized content tailored to individual preferences, enhancing user experiences in applications like recommendation systems and targeted advertising.
- Data Augmentation: Generative AI can create synthetic data to augment training datasets, improving the performance of other AI models and addressing data scarcity issues.
Limitations:
- Bias: Generative AI models can inherit and amplify biases present in their training data, leading to biased or unfair outputs. Mitigating bias remains a significant challenge.
- Data Dependency: The quality and diversity of the generated content are heavily dependent on the training data. Poor or unrepresentative data can limit the model’s effectiveness.
- Resource Intensive: Training generative AI models requires substantial computational resources, making them expensive and environmentally impactful.
- Quality Control: While generative AI can produce high-quality content, it may also generate outputs that are nonsensical or factually incorrect, necessitating human oversight and quality control.
Does Generative AI think?
The characterization of generative AI systems as novel forms of intelligence or artificial cognition has become prevalent in both popular and academic discourse. This perspective, while intuitively appealing, may lead to misconceptions about the nature and capabilities of these systems. Vallor (2024) uses the metaphor of a mathematical mirror, which reflects patterns of human-generated data, to provide a more nuanced and accurate conceptualization.
This reflective model provides a framework for understanding how these systems operate without attributing to them qualities of cognition or understanding that they do not possess. It allows us to appreciate their capabilities while maintaining a clear distinction between their functioning and human-like intelligence.
The Reflective Model: A Conceptual Framework
The analogy of a physical mirror serves as an accessible entry point to understanding the more complex mathematical processes at work in generative AI. Just as a mirror’s physical properties determine how it reflects light, the mathematical properties of AI algorithms determine how they “reflect” patterns in data.
Key aspects of this analogy include:
- Reflection vs. Generation: Like a mirror that doesn’t create light but reflects it, these AI systems don’t create new information but reflect patterns from their training data.
- Fidelity and Distortion: Just as mirrors can provide clear reflections or distort images, AI systems can produce outputs of varying accuracy or relevance.
- Amplification: Some mirrors can magnify reflections; similarly, AI systems can amplify certain patterns or biases present in their training data.
Operational Mechanism of the AI Reflective Model
The operational mechanism can be broken down into three main phases:
- Data Input (Training Phase): This phase involves exposing the AI system to vast quantities of human-generated data. This data serves as the “light” in our mirror analogy. It typically includes:
- Textual data from books, articles, websites, and social media
- Conversational data from online forums and chat logs
- Structured data from databases and knowledge bases
- Pattern Recognition and Replication: During this phase, the AI system processes the input data to identify statistical regularities. This involves:
- Analyzing the frequency and co-occurrence of words and phrases
- Identifying grammatical structures and linguistic patterns
- Recognizing contextual relationships between different pieces of information
Importantly, this process does not involve semantic understanding. The system is not comprehending the meaning of the text, but rather identifying mathematical relationships within the data.
- Output Generation: When prompted, the AI generates responses by:
- Identifying patterns in the prompt that match patterns in its training data
- Combining and extrapolating from these patterns to produce new text
- Using statistical probabilities to determine the most likely sequence of words or ideas
Functional Capabilities and Limitations
Understanding the reflective nature of these systems is crucial for accurately assessing their capabilities and limitations:
- Lack of True Cognition: While these systems can produce outputs that appear to demonstrate reasoning or emotional awareness, they are not engaging in cognitive processes as humans do. They are performing complex statistical analyses based on patterns in their training data.
- Absence of Understanding: The AI does not comprehend the content it generates in any meaningful sense. It cannot reason about the implications of its outputs or verify their factual accuracy beyond pattern matching.
- Contextual Limitations: The AI’s responses are limited to reflecting patterns present in its training data. It may struggle with novel contexts or information outside its training distribution.
- Impressive but Bounded Capabilities: These systems can generate highly sophisticated and contextually appropriate text, solve certain types of problems, and even engage in creative-seeming tasks. However, these capabilities are bounded by the patterns present in their training data and do not represent open-ended intelligence.
By conceptualizing generative AI through this reflective model, we can better understand its functioning, appreciate its capabilities, and recognize its limitations. This framework provides a more accurate basis for discussing and researching these systems, avoiding anthropomorphization while still acknowledging their significant technological achievements.
How to Give Clear and Effective Instructions to Generative AI Models
Understanding Prompt Engineering
Generative AI models like those from OpenAI are designed to respond optimally to certain prompt formats due to the way they are trained. This process, known as prompt engineering, involves crafting inputs to get the best possible outputs from these models.
To optimize prompt engineering with generative AI models, follow these best practices:
- Use the Latest Model: Newer models provide better performance.
- Clear Instructions: Place instructions at the beginning of the prompt, separated by delimiters.
- Specificity: Be detailed about the desired outcome, length, format, and style.
- Examples: Provide output format examples to guide the model.
- Step-by-Step: Start with zero-shot, then few-shot, and fine-tune if necessary.
- Precision: Avoid vague descriptions.
- Positive Guidance: Instruct on what to do, not just what to avoid.
- Code Generation: Use “leading words” to nudge the model towards desired patterns.
For detailed examples and more information, visit the Prompt Engineering Guide.
Additionally, experimenting with different prompt structures can also yield beneficial results. Applying these strategies will help you communicate more clearly with generative AI models, ensuring you receive the most accurate and useful responses.
Examples
Information Extraction: Extracts specific details from text, like identifying a product mentioned in a research paper.
Question Answering: Answers questions based on given context, for instance, identifying the source of a molecule.
Text Classification: Categorizes text based on sentiment, like determining if a statement is positive, negative, or neutral.
Conversation: Adapts tone and complexity of responses, such as explaining black holes to different audiences.
Code Generation: Generates code snippets based on specific instructions, from simple greetings to complex database queries.
Reasoning: Solves arithmetic and logic problems by breaking them into steps.
Four important prompting techniques
Effective prompting techniques are essential for maximizing the potential of language models in various applications. Among the most powerful methods are zero-shot prompting, few-shot prompting, chain prompting, and the tree of thoughts approach. Each technique offers unique advantages and can be tailored to specific tasks to enhance the performance and reliability of AI models. Let’s explore these techniques in detail to understand how they can be applied in different scenarios. Please visit the links below and review the full examples. You can also find additional prompting techniques at: promptingguide.ai.
Zero-shot prompting involves providing the model with a task it has never seen before without any examples. The model uses its pre-trained knowledge to generate a response based on the prompt alone. This technique is useful for tasks where you want the model to generalize from its training data to new, unseen situations.
Few-shot prompting provides the model with a few examples of the task at hand within the prompt. By seeing these examples, the model can better understand the task requirements and generate more accurate responses. This approach is beneficial when you need the model to perform a task that might be slightly outside its usual scope, but still within the realm of its training data.
Prompt chaining involves breaking down a complex task into a series of smaller, manageable prompts. Each prompt builds upon the previous one, guiding the model step-by-step through the problem-solving process. This method helps improve the accuracy and coherence of the model’s responses, particularly for tasks that require multi-step reasoning or detailed analysis.
The tree of thoughts technique is a structured approach where the model generates multiple potential solutions or thoughts at each step of the problem-solving process. These thoughts branch out like a tree, allowing for the exploration of various possibilities before converging on the best solution. This approach is particularly effective for tasks that involve creative problem-solving or scenarios where multiple pathways need to be considered before arriving at a final answer.
From prompt engineering to problem formulation
Managing Generative AI as a Skilled Yet Inexperienced Team Member
When working with generative AI models, it’s helpful to imagine them as an employee who is always confident in their answers but doesn’t always know the best way to accomplish a task. This analogy can guide you in providing clear and precise instructions, just as you would for an eager yet inexperienced team member.
Generative AI models excel at generating responses based on patterns in the data they were trained on, but they may struggle with tasks that require specific knowledge or nuanced understanding. Therefore, it’s crucial to give detailed, step-by-step instructions to ensure they perform correctly. Much like you wouldn’t expect an employee to guess what you want without proper guidance, you shouldn’t expect a generative AI to produce perfect results without clear and explicit prompts.
By treating generative AI like a confident but sometimes uncertain employee, you’ll be more mindful of how you craft your prompts. Start with clear instructions at the beginning, use specific and detailed language, and provide examples to illustrate the desired outcome. This approach will help you get the most accurate and useful responses from your AI, just as clear and effective communication ensures the best performance from your human team members.
Improving Problem Formulation Skills
This is a summary of Acar (2023)
To enhance problem formulation skills, we can draw insights from past research and practical experience, particularly from crowdsourcing platforms where organizational challenges are routinely articulated and addressed. Four key components of effective problem formulation are problem diagnosis, decomposition, reframing, and constraint design.
Problem Diagnosis
Problem diagnosis involves identifying the core issue for AI to address. This means pinpointing the main objective you want generative AI to achieve. Some problems are straightforward, like gathering information on human resource management strategies for employee compensation. Others, like finding innovative solutions, are more complex.
A notable example is InnoCentive (now Wazoku Crowd), which has helped clients formulate over 2,500 problems with an impressive success rate exceeding 80%. Their success stems from their ability to identify the fundamental issue underlying a problem. They often use the “Five Whys” technique—repeating the question “why?” five times, each time directing the current “why” to the answer of the previous “why”—to differentiate root causes from symptoms. For instance, in tackling the Exxon Valdez oil spill, they identified that the core issue was the viscosity of the crude oil in subarctic waters, leading to a solution that involved vibrating the oil to keep it in a liquid state.
Problem Decomposition
Problem decomposition involves breaking down complex problems into smaller, more manageable subproblems. This is crucial when dealing with multifaceted issues that are too convoluted to solve directly.
For example, in the InnoCentive challenge for amyotrophic lateral sclerosis (ALS), instead of seeking a broad solution for the disease, the focus was narrowed to detecting and monitoring its progress. This led to the development of a noninvasive and cost-efficient ALS biomarker.
Testing how AI improves with problem decomposition, a challenge like implementing a robust cybersecurity framework can be broken down into subproblems such as security policies, vulnerability assessments, authentication protocols, and employee training. This approach leads to more practical and effective solutions.
Problem Reframing
Problem reframing involves viewing a problem from different perspectives to find alternative interpretations and solutions. By reframing, you can guide AI to explore a wider range of potential solutions, overcoming creative blockages.
An illustrative example is Doug Dietz of GE HealthCare. When he observed the fear children experienced during MRI scans, he reframed the problem to ask, “How can we turn the daunting MRI experience into an exciting adventure for kids?” This led to the creation of the GE Adventure Series, significantly reducing pediatric sedation rates and improving patient satisfaction.
Another example could be reframing an office parking problem. Instead of focusing solely on increasing parking spaces, reframing the issue to consider employees’ stress and commuting options can lead to solutions like promoting alternative transportation and remote work options.
Problem Constraint Design
Problem constraint design involves setting the boundaries for problem-solving by defining input, process, and output restrictions. Constraints can guide AI to generate solutions that are valuable for the task at hand. For productivity-oriented tasks, specific and strict constraints are effective. For creativity-oriented tasks, experimenting with constraints can help discover novel perspectives.
For instance, brand managers use AI tools like Lately or Jasper to produce social media content, setting precise constraints on length, format, tone, or target audience to ensure consistency with the brand image. Conversely, when seeking originality, constraints can be minimized. A great example is GoFundMe’s Help Changes Everything campaign, where unorthodox constraints led to the creation of a visually striking and emotionally impactful year-in-review video using AI-generated street mural-style art.
Real-World Examples: Generative AI and LLMs in Marketing
Understanding the theoretical aspects of generative AI and Large Language Models (LLMs) is crucial, but seeing their practical applications in the real world can make these concepts more tangible and relevant. Here are some specific case studies and examples of how generative AI and LLMs are being used in marketing today.
Case Study 1: Personalized Email Marketing with LLMs
Company: Persado Application: Personalized Email Campaigns
Persado, a leading AI company, utilizes LLMs to generate personalized email content for marketing campaigns. By analyzing customer data and preferences, Persado’s AI crafts personalized subject lines, body text, and calls-to-action that resonate with individual recipients. This approach has significantly improved open rates and conversion rates for their clients, demonstrating the power of personalized communication in marketing.
Impact:
- Increased email open rates by up to 68%.
- Enhanced engagement through tailored content.
- Higher conversion rates leading to increased sales.
Case Study 2: Dynamic Ad Generation with Generative AI
Company: The Grid Application: Automated Web Design and Ad Creation
The Grid is a web design platform that uses generative AI to create visually appealing and effective advertisements. By inputting basic information about a product or service, the AI generates customized ad designs that are optimized for various platforms, including social media, search engines, and display networks. This allows marketers to quickly and efficiently produce high-quality ads without the need for extensive design expertise.
Impact:
- Reduced time and cost for ad creation.
- Consistent brand aesthetics across multiple platforms.
- Improved ad performance through data-driven design adjustments.
Case Study 3: Chatbots and Customer Interaction
Company: H&M Application: Customer Service Chatbots
H&M, the global fashion retailer, employs LLM-powered chatbots to handle customer service inquiries on their website and mobile app. These chatbots are trained to understand and respond to a wide range of customer queries, from product information to order tracking and returns. By providing instant and accurate responses, the chatbots enhance the customer experience and free up human agents to handle more complex issues.
Impact:
- Faster response times for customer inquiries.
- Improved customer satisfaction and engagement.
- Reduced operational costs associated with customer service.
Case Study 4: Content Creation for Social Media
Company: Coca-Cola Application: Social Media Content Generation
Coca-Cola has experimented with using generative AI to create engaging content for their social media channels. By leveraging LLMs and generative models, the company can quickly produce a variety of posts, including captions, hashtags, and visual content ideas. This not only keeps their social media presence active and dynamic but also allows them to test different content strategies to see what resonates best with their audience.
Impact:
- Increased social media engagement through diverse content.
- Enhanced brand visibility and interaction with followers.
- Ability to rapidly adapt to trending topics and events.
Summary: Using Generative AI
Key Takeaways:
- LLMs are a specific type of AI focused on understanding and generating human-like text, while Generative AI is a broader category encompassing systems that create various types of content (text, images, music).
- Generative AI works through a training process that analyzes patterns in vast datasets, functioning more like a mathematical mirror reflecting patterns rather than truly “thinking” or understanding.
- Major capabilities include content creation across media types, personalization, and data augmentation, while limitations include bias inheritance, heavy resource requirements, and quality control needs.
- Real-world applications show successful implementation in marketing, such as Persado’s personalized email campaigns and H&M’s customer service chatbots, demonstrating tangible business impact.
- Effective prompt engineering requires clear instructions, specific details, and appropriate examples, with four key techniques: zero-shot prompting, few-shot prompting, chain prompting, and tree of thoughts.
AI in Action: Case studies demonstrate AI’s practical impact, like Coca-Cola’s social media content generation and The Grid’s automated web design platform, showing how theoretical capabilities translate into business value.
Think Deeper:
- How might the “mathematical mirror” concept of AI influence how we approach prompt engineering and problem formulation?
- Consider the balance between automated AI solutions and human oversight – what criteria should organizations use to determine the appropriate level of human involvement?
- How can organizations effectively decompose complex problems for AI while maintaining sight of the broader strategic objectives?