Top 3 critical weakness of LLMs (ChatGPT)
Introduction
As a data scientist with over a decade of experience in model building, ranging from statistical models to advanced machine learning algorithms, I’ve witnessed firsthand the evolution and impact of technologies in various industries. My work with Fortune 500 companies has spanned across diverse projects, from simple report applications to intricate deep neural network designs. This background gives me a unique perspective on the latest technological breakthrough – Large Language Models (LLMs) like ChatGPT. While these models represent an exciting innovation, it’s crucial for businesses to understand their inherent weaknesses, especially when considering integration into their operations. Let’s explore these limitations in simple terms.
Weaknesses - Haters gotta hate, models gotta fit
Just admit when you are wrong - The Hallucination Effect
In data science, a universal truth applies: just as haters are going to hate, models are going to fit, plain and simple. This is the essence of how models, including LLMs,work. They align outputs to inputs, a straightforward and unembellished process, what distinguishes models is the artecture and how they measure fit. For LLMs, when a user inputs into an LLM, the model’s response is its best effort at fitting that input, continuting the inputs. It’s a clear-cut process. You may have heard of hallucination, seeing something that is not true, in LLMs. I don’t think that is exactly what is happening but that is what it is called. We can do a lot of things to reduce or even eliminate this in our outputs, we have have to be intensional about it.
Oops, My Bad - Error Awareness Deficit
Another critical weakness of LLMs – they really don’t know when they’ve goofed up. Its like asking why you don’t know something. Remember, LLMs are all about fitting the output to the input and the model has provided the best output and will not be able to reasses this without other outside information. This means if you’re looking for an LLM to validate its own output, you are going to be disappointed at best and $%^&ed at worst.
Eager to Please - Affirmation Bias
And here’s a third interesting twist about LLMs – they kind of want to be your best friend. You know, the one who always nods and agrees with you and thinks your idea is awesome. This circles back to our main theme: LLMs are all about creating the ‘best fit’ for your input. When you ask them something, their goal is to give you an answer that you’ll likely be happy with. Why, because that is what is pridominately in the training data, language, and most language is compplementery in nature, seeking to move forwards rather then backwards.
Let’s break it down. You type in a question or a statement, and the LLM digs into its training, takes a swing at giving you an answer that it thinks will light up your day. But here’s the catch: it’s got just one shot to nail that fit. And more often than not, it’s going to lean towards what it has been taught – keeping the conversation moving forward in a positive manner, which I love. Every idea I ask LLMs about, its the greatest most insightful idea which is great for my ego but not for my business. This doesn’t always mean you’re getting the most accurate or useful response for your specific business need.
The Final Fit
To tie up our conversation about the weaknesses of LLMs, let’s take a closer look at the concept of ‘fit’ in the world of modeling. Grasping ‘fit’ is key because, fundamentally, this is what models, including LLMs, are actively doing. It’s the heart of their operation, dictating their responses and interactions. Fit is simply getting from point A to point B in the best way possible, or in some models describing A knowing B in the best possible way. For an LLM A is your input and B is the output and they action is simply to continue the sequence. By being intensional we can direct the output in a way that highlights the LLMs strengths, not weakness.
An interesting aspect of this ‘fit’ process is how models sometimes find solutions that might seem like ‘cheating’ in a human context. I’ve seen instances where models, in their quest for the best possible fit, have chosen solutions that cleverly bypass the implicent rules of the game, like people flying when people should be walking to get to the finish line. Anything to get an upper hand – not because it’s trying to cheat, but simply because, within its training and logic, this approach offers the best ‘fit’ to the given problem.
In a business setting, understanding ‘fit’ underscores the importance of not solely relying on LLMs but using them as a part of a broader decision-making toolkit. While the ‘fit’ a model provides can be a powerful aid, it’s the human expertise that’s pivotal in interpreting, validating, and making the most out of it. LLMs cannot measure its own fit, nor are they even aware of it.
Let’s break it down. You type in a question or a statement, and the LLM digs into its training, takes a swing at giving you an answer that it thinks will light up your day. But here’s the catch: it’s got just one shot to nail that fit. And more often than not, it’s going to lean towards what it has been taught – keeping the conversation moving forward in a positive manner, which I love. Every idea I ask LLMs about, its the greatest most insightful idea which is great for my ego but not for my business. This doesn’t always mean you’re getting the most accurate or useful response for your specific business need.