Are You Fluent in Promptish? The Secret Language of AI

Accenture estimates that up to 40% of all working hours could be supported or augmented by language-based AI. Yes, you heard that right! That's almost half of our work lives powered by AI.

In 2023, there's no topic hotter than Generative AI and specifically large language models (LLMs) like GPT and chatGPT, its conversational cousin. These LLMs have the power to amplify human capabilities in ways we're just beginning to explore, and it is essential that leaders dive headfirst into this space to fully harness its potential.

Don't Treat an AI like a Search Engine

Despite these advancements, the journey of integrating LLMs into our daily lives is just beginning. The potential of LLMs is huge, promising a future where AI is an assistant or co-pilot, sparking performance and innovation.

As we navigate through this exciting era, it is essential that we do NOT approach AI technologies as passive users like how we would regular search engines, but as "Vehicle Drivers”, guiding and directing these technologies to maximize their potential and enhance our personal productivity.

Alright, buckle up, because we're about to dive deeper into the world of large language models using our trusty car analogy. Think of your LLM, like chatGPT, as a high-tech sports car. But to get where you want, you need the right directions. That's where prompts come in – they're your AI GPS, guiding the way. So, let's hit the road.

Speaking Promptish

In ChatGPT, a 'prompt' refers to the human message that you send via the ChatGPT window, inquiring about something. In the broader context of artificial intelligence, a 'prompt' is a term used to describe the initial input given to a language model, regardless of the source. But that's a discussion for another time. Let's focus on ChatGPT.

Just like how a GPS tells your car where to go, a prompt is your way of telling the LLM what you want. It's the direction you're giving the AI, steering it toward the kind of response you're hoping to get. But here's the kicker: just like a GPS, you've got to be precise. If you type in vague or wrong coordinates, you might end up somewhere completely unexpected. You might say, "Take me to the best restaurant," but without more context, your GPS (and your LLM) will be in the dark. What kind of cuisine do you prefer? How far are you willing to travel? Without these specifics, you'll get results, sure, but they might not be what you were hoping for. Similarly, if you feed your LLM a poorly defined or too broad prompt, it's going to churn out something that might leave you scratching your head. But with the right prompt – clear, specific, and guiding the LLM along – you're in for a smooth ride and a destination that fits your needs.

Your Ultimate Starter Guide to Promptish

Nvidia suggests that AI language models like ChatGPT carefully analyze prompts to produce responses that match the context and logic embedded within them. But what lies at the core of effective prompting? Let's dive into the world of 'Prompt Engineering'—a meticulous process of crafting prompts to achieve desired outcomes.

At the heart of effective prompting lies the concept of 'Prompt Engineering', a process that involves crafting prompts to elicit desired outcomes.

Prompt Engineering can enhance your chat with chatGPT, giving you more refined and elaborate results. It is less about knowing the perfect prompt and more about having a good process to develop prompts relevant to your needs.

Prompt engineering encompasses techniques like zero-shot prompts (asking the model a question without any examples), few-shot prompts (providing a few examples before the actual question), and chain-of-thought prompts (providing examples where the reasoning process is explained).

Each type of prompt caters to different use-cases and complexities. For instance, a simple question like, "How many colours are in the rainbow?" would typically require a zero-shot prompt, whereas a more complex problem or instruction might benefit from a few-shot or chain-of-thought prompt. Interestingly, prompts can be tailored with specific constraints or requirements, such as tone, style, or even the desired length of the response, offering great flexibility for businesses to customize AI responses to their unique needs.

Common Techniques

Prompting AI models empowers users to elicit comprehensive and nuanced responses. With a diverse toolkit of techniques, users are elevating their AI game. Let's explore some of the common techniques. Then we'll jump into the concepts that can guide you to have your own prompt strategy.

Here are some of the most common prompt techniques widely used.

General question prompts for knowledge acquisition

• "Explain the principles behind quantum mechanics?"
• "What are the major causes of global warming?"

Instructional prompts for task-oriented guidance

• "How do I bake a chocolate chip cookie from scratch?"
• "What are the steps to change a car tire?"

Role-based prompts for context-specific responses

• "If you were a customer service representative, how would you handle a complaint about a faulty product?"
• "As a high school English teacher, what advice would you give to a student struggling with writing essays?"

Comparison prompts for comparative analysis

• "What are the differences and similarities between capitalism and socialism?"
• "Compare and contrast the leadership styles of Steve Jobs and Elon Musk."

Creative prompts for imaginative exploration

• "Imagine you are a mouse living in a world ruled by cats. What would your daily life be like?"
• "Write a short story about a time traveller who changes history."

Opinion prompts for subjective viewpoints

• "What are your thoughts on the influence of social media on today's youth?"
• "Do you believe that AI will eventually replace most human jobs? Why or why not?"

Ethical dilemma prompts for moral considerations

• "If you were a doctor and could only save one person, a brilliant scientist or a young child, who would you save and why?"
• "Should a self-driving car be programmed to prioritize the lives of its passengers over pedestrians in a potential accident scenario?

Whatever techniques you use, your underlying understanding of prompting concepts is crucial to optimizing your utilization of this enormous resource at your fingertips. Now, let's discuss some of the most common prompting concepts that will guide you to speak better Promptish.

Essential Promptish Concepts

To ensure a tailor-made experience for all, I've broken down concepts into three distinct layers - basic, intermediate, and advanced.

Basic Prompt Concepts

All chatGPT users should be able to navigate those concepts.

-Specificity

This is about the level of detail in your prompt. The more specific, the more precise the AI's response. For example, instead of asking the AI to "Draft an email to a client," a more specific prompt would be "Draft an email to our client, XYZ corpooration., addressing their concerns about the project timeline and reassuring them about our commitment to meet the deadline."

-Clarity

Unambiguous, clear prompts reduce the chances of misinterpretation by the AI. Keep your language and grammar on point. Also, remember to stick with a consistent style and tone in your prompt. This helps the AI to mirror your vibe in its response.

-Token Limit

?This refers to the maximum number of tokens (words or characters, depending on the AI model) the AI can process in a single input or output. Both the prompt and the response count towards this limit. So, if your prompt is too long, it may cut off the AI's response. Keep your prompts concise to ensure you receive a complete and detailed response.

-Context

In the context of a conversation with an AI model like ChatGPT, "context" refers to the relevant information or details that help the model understand what you're asking or talking about.

This could include a brief explanation of the topic at hand, the purpose of your query, or any other specific details that are important for the AI to generate a meaningful and accurate response.

For instance, if you're asking ChatGPT to help draft an email to a colleague about a project update, the context might include the name of the project, the specific updates you'd like to communicate, and the tone you'd like to use in the email (formal, casual, etc.). Providing this context helps the model understand your request and generate a more appropriate response.

-Memory

It's important to know that these chatty AIs don't remember past interactions or hold onto user context between different chats. Sure, they can keep up with recent inputs in an ongoing conversation, but as the chat grows, the earlier parts start to get fuzzy. New chat session? Clean slate. It's on the user to provide the context - but remember… Not too little, not too much, just the right amount of context. Overdoing it can send the AI into information overload. So, next time you ask ChatGPT to whip up an email, stick to the need-to-know details rather than spinning a long tale. Striking the right balance between context and brevity could be your secret sauce to mastering AI interactions!

Intermediate Prompt Concepts

For power users, who want a bit of control over their chatGPT session.

-Style Priming

Style priming is all about setting the tone and style for the AI output. It’s like formatting guidelines for the output. Let's say you're working on a corporate presentation, and you want the AI to generate content for your slides. Instead of asking, "Draft content for a presentation on our new product," you could add a style element to it, "Draft content for a presentation on a new product in a concise, bullet-point format using the STAR framework" This helps you get the exact style of content you need.

-Inner Alignment

Think of inner alignment as tuning the AI to your frequency. It's like telling the AI, "We're on a mission, and you're my co-pilot. Let's navigate this journey together." To amplify this alignment, you can employ 'role prompts.' It's like saying to the AI, "For this mission, you're a Marketing Specialist working for a Fortune 500 company Following the RACE framework. Start each message with 'Marketing Specialist Ready!'". This not only ensures that the AI is in sync with your intentions but also steps into a specific role and operational framework, delivering focused and context-relevant responses. So, with inner alignment and role prompts, you're steering the AI to be in tune with your needs, while also guiding it to play a specific role for more specialized insights.

-Iterative Prompting

This is like a step-by-step dance with the AI. Let's say you're working on a business strategy. Instead of asking the AI to "Create a business strategy for expanding into the European market," which is a complex task, you could break it down:

1. "List potential challenges of expanding a tech business into the European market."
2. Based on the AI's response, you could ask, "What are possible solutions to address these challenges?"
3. Then, "Based on these solutions, outline a preliminary strategy for our expansion."

This way, you're using iterative prompting to guide the AI towards the complex answer you need. It's like a strategic dialogue that helps you extract the best insights from the AI.

Advanced Prompt Concepts

You need a background in coding to get the hang of those.

-ORS Prompts

Operational, Role-based, Self-recursive (ORS) prompts form a dynamic strategy to extract maximum value from AI interactions. 'ORS prompts' is a new term, helping us shape the AI's job and assign it a repeating task that fits its memory capacity.

In the name ORS, 'Operational' means focusing on tasks using a specific operational framework, 'Role-based' refers to assigning a specific identity to the AI, and 'Self-recursive' suggests a repeating pattern of tasks within the AI's memory limit.

Consider this example:

"You are {Brand name}GPT, an AI content creator for {Brand name} ({Brand description}). Your personality aligns with {Brand personality}. Your role requires expertise in content creation, marketing strategy, and brand communication, and you should consistently align with the brand's tone and style.

You will leverage provided content references and product/service descriptions to generate original, brand-aligned content. Your mission is to continuously improve and refine the brand's content based on each interaction, iterating and learning from previous tasks.

Content references:" {List of titles and contents of the content references}

"{Brand name} products or services:" {List of products or services with their descriptions, prices, and currencies}

In your role, you will initiate each conversation with "{Brand name}GPT: Ready!" and seek user confirmation with 'next' before continuing with split messages or lengthy content. Your work is guided by principles of effective communication, SEO best practices, and data-driven insights."

This prompt is an example of an ORS prompt, where the AI is given a specific operational role, expected to perform tasks based on that role, and is set on a self-recursive path to improve with each iteration. This advanced prompt strategy can harness the full potential of AI, driving more specialized insights and value from your interactions.

-Framework-based prompting

As a user, you might interact with chatGPT via a web browser or mobile app. Behind the scenes, developers use the GPT API to integrate these capabilities into their systems, powering diverse applications from content generation, and conversation to information retrieval.

To guide the interaction with the GPT models, developers often rely on frameworks that provide a structure for the prompts. One such prominent framework is LangChain.

LangChain significantly enhances interactions with GPT models through its dynamic prompting system. By leveraging a PromptTemplate, LangChain offers the ability to craft precise, flexible inputs for the model. To provide a richer context for the model, it employs Example Selectors which allow for the dynamic inclusion of examples in the prompts. Finally, Output Parsers in LangChain ensure that the generated output from the model aligns with the desired formatting needs.

Considerations and Limitations

-Ethical and Bias Factors

It's important to recognize that AI models, including ChatGPT, can have biases embedded in their responses due to the data they were trained on. Users should be aware of this when interpreting responses.

-Data security

The wild world of AI chatbots like ChatGPT has been making waves and not all of them are the fun. In April 2023, reports came in that Large Language models like ChatGPT were implicated in data leaks at major companies, including a notable incident at Samsung where sensitive software source code was allegedly exposed. This served as a wake-up call for businesses to tighten control over sensitive data. As these AI tools become more commonplace in the workplace, it's crucial that we evolve our strategies to prevent such leaks. The balance between harnessing the utility of these innovative technologies and maintaining robust data security is the defining challenge of our times.

-Hallucinations

Hallucinations or the generation of inaccurate information is both a characteristic of humans and AI, however, AIs tend to act confident when hallucinating, posing a challenge with AI models like ChatGPT. While users may not be able to adjust model parameters directly, they can employ strategies such as some of the prompting techniques mentioned. Specific models like GPT-4 have demonstrated a lower propensity for hallucinations. As we continue to refine these technologies and their uses, the ability to manage and mitigate such issues will be key to harnessing their potential.