We have all received this message from ChatGPT:

"Something went wrong. If this issue persists, please contact us through our help center…"

But what do these error messages really mean and how can we communicate better? It’s not always what you think.

In this week’s edition of AI & YOU, we’re delving into the reasoning behind Large Language Models (LLMs) like ChatGPT, exploring how ‘failures’ are not always what they seem and how the right prompt can change the game. We then guide you through how to encode understanding through ChatGPT prompt engineering techniques.

As always, our team of expert AI practitioners is on standby to help your organization harness the power of AI effectively and efficiently. If you are building enterprise systems that utilize ChatGPT’s (or other LLM’s) API, unpredictable responses make your solutions less reliable. If your enterprise needs help incorporating such APIs into solutions, or building custom solutions to leverage LLMs to answer questions on your data and databases, book a call with me below.

Make sure to subscribe and share our content if you find it helpful!

What your LLM isn’t saying

n the realm of AI, Large Language Models (LLMs) have become revolutionary tools, reshaping the landscape of numerous industries and applications. From writing assistance to customer service, and from medical diagnosis to legal advisory, these models promise unprecedented potential.

Despite their robust capabilities, understanding LLMs and their behavior is not a straightforward process. While they may fail to accomplish a task, this ‘failure’ often hides a more complex scenario. Sometimes, when your LLM (such as the popular ChatGPT) seems to be at a loss, it isn’t because of its inability to perform, but due to other less obvious issues, like a ‘loop’ in the decision tree or a plug-in timeout.

Understanding and Overcoming Those Error Messages

When an LLM like ChatGPT encounters a problem and fails to execute a task as expected, it doesn’t typically communicate its struggle with words of defeat, but rather through error messages. These messages can often signal the presence of an internal technical issue that is causing an impediment rather than indicating a limitation of the model itself.

This could be a result of the model getting caught in a loop during its decision-making process decision tree, causing it to either repeat certain steps or halt altogether. This doesn’t mean that the model is incapable of completing the task, but rather that it has encountered a problem in its algorithm that needs to be addressed.

Similarly, a plug-in timeout can happen when a specific plug-in, which is an additional software component that extends the capabilities of the main software, takes too long to execute a task. Many LLMs weren’t originally designed for the fast-paced environment of web-based applications and might struggle to keep up with the demanding speed requirements, leading to plug-in timeouts.

Real-life Examples and Solutions

Consider an instance where an LLM, like ChatGPT, is being used for automated story generation. The task is to generate a short story based on a user-inputted prompt. However, the model gets stuck in a loop, continuously generating more and more content without reaching a conclusion. It appears to be a ‘failure’ as the model is not able to deliver a concise story as expected.

• The true issue: The model has gotten stuck in its decision-making loop, continuously extending the story instead of wrapping it up.

• The solution: A small tweak in the prompt or a subtle adjustment in the model’s parameters could steer the model out of the loop, enabling it to complete the task successfully.

*You can find more real-life examples and solutions in our blog.

Deciphering LLM’s Silent Messages

When an LLM encounters a problem, it isn’t necessarily a ‘failure’ in the conventional sense. Instead, it’s often a silent signal – an unspoken word – pointing towards a specific issue like a decision loop, a plug-in problem, or unexpected behavior that has interfered with the model’s task.

Understanding these silent messages from the LLM can allow us to adapt, optimize, and improve its performance. Therefore, the key lies not in focusing on the error message alone, but in unraveling the deeper, often hidden, meanings behind these messages.

Check out the full blog: “What Your ChatGPT Error Message Means”

How to Encode Understanding Through Prompt Engineering

Prompt engineering with large language models (LLMs) like ChatGPT and Google’s Bard is an essential, yet often overlooked aspect of these powerful AI tools. It is akin to setting the stage for an AI-powered dialogue, offering initial direction to the computational conversation. When you’re engaging with an LLM, your initial prompt is your first step into the vast landscape of possibilities these models offer. It’s your way of setting expectations, guiding the conversation, and most importantly, shaping the AI’s response.

The Power of Encoding a Typical Example

When we encode a typical example in our initial prompt, we’re providing the AI with a clear idea of what we want. This is especially valuable when it comes to handling complex requests or tasks. Let’s consider a scenario where we want our AI to help draft a business proposal. Instead of a vague instruction like “Draft a business proposal,” we can provide a typical example: “Draft a business proposal similar to the one we did for ABC Corp. last year.” Here, we’re encoding a typical example into the initial prompt, providing a clear direction to the AI.

Influencing the Way of Thinking: Guiding AI through Prompts

Through careful and thoughtful prompt engineering, we can influence the AI’s “way of thinking”, steering it towards generating responses that are closer to what we need or anticipate. However, it’s not merely about providing a clear command or a set of instructions. It’s about capturing the essence of a thought process or a reasoning path in the prompt.

For instance, let’s say we want the AI to solve a mathematical problem. Instead of directly asking for the solution, we could guide the AI to demonstrate the problem-solving steps. A prompt like “As if you were a math tutor, walk me through the steps to solve this equation…” can significantly influence the AI’s response, eliciting a step-by-step solution that mimics a tutor’s way of thinking.

The Initial Prompt as a User Guide: Setting the Stage for Interaction

In the realm of AI interaction, an initial prompt can serve a similar function to a user manual, giving the user guidance on what’s possible. It helps to condition the user, providing a roadmap for their interaction with the AI. It’s like a prelude, setting the tone for the ensuing conversation.

A well-crafted initial prompt might look something like this: “Imagine you’re a travel writer crafting an article about the best cafes in Paris. Begin your piece with a vivid description of a charming cafe by the Seine.” This not only directs the AI towards the desired task but also sets an expectation for the user about the kind of response that can be generated.

Encoding Expertise into AI

As we unravel the intricacies of large language models, it becomes clear that prompt engineering is not just a technical requirement—it’s a fundamental tool for encoding our way of thinking into artificial intelligence. Whether it’s a simple reminder or a comprehensive guide, the initial prompt serves as the cornerstone of human-AI interaction, defining the boundaries and possibilities of the conversation.

By effectively using the initial prompt, we can encode a typical example of how the AI should respond, shape the user’s way of thinking, and guide the AI’s responses.

Check out the full blog: “How to Encode Understanding Through Prompt Engineering”

Thank you for taking the time to read AI & YOU!

*Skim AI is a Machine Learning and Artificial Intelligence consultancy that educates executives, performs due-diligence, advises, architects, builds, deploys, maintains, updates and upgrades enterprise AI across language (NLP), vision (CV) and automation based solutions.

*Chat with me about Enterprise AI

*Follow Skim AI on LinkedIn