The Banality of ChatGPT: A Year-Long Reflection on Generative AI
Introduction to the Article
In this post, originally written to introduce a presentation on Generative AI — which, in the end, I didn’t have the opportunity to use — I try to summarize my personal vision of this technology. This perspective has evolved throughout 2023, a year that allowed me to deepen my understanding and apply Generative AI.
I chose not to alter it significantly, maintaining its original form; therefore, you may notice some references to the ‘presentation’. Although it was not possible to use it as intended, I am satisfied with the outcome. I believe it can be useful, and for this reason, I wish to share it with my readers.
Disclaimer
The only disclaimer I include is related to the parallel with Hannah Arendt’s “the banality of evil”. The reference is intended only to capture the aspect of accountability, without any other historical reference or moral judgment.
Feel free to comment on aspects you disagree with or other perspectives you have.
The post was written on December 15th.
The banality of ChatGPT
The presentation will focus on Generative AI, with a special emphasis on a tool like ChatGPT, which few of us are currently utilizing and still striving to fully understand, both from the user and the researcher’s perspective.
The Past Year: A Developmental Milestone; there has been an attempt to harness this technology for the benefit of humanity. For now, the most intriguing practical case is that of RAG. In our presentation, we aim to go further, exploring how we’re channeling this technology to leverage the capabilities of an assistant at this level.
Making ChatGPT Familiar
However, the first step in presenting this tool is to make it familiar. It’s essential to conceptualize it correctly, giving it a form. Otherwise, we might be frightened by something unknown, unused, and only heard of in terms of its potential consequences. In this context, I recommend the book “The Psychology of Money.”
Then, let’s start with the reason behind the title of this presentation: “The Banality of ChatGPT”. The reference is explicitly drawn from Hannah Arendt’s famous book. Why this comparison? If any of you have read the book or studied it during high school or university, you might vaguely recall its central focus, exemplified by the title. To summarize briefly, the book discusses the trial of Eichmann, from which Arendt inferred
“that the evil perpetrated by Eichmann… was not due to a malevolent nature, deeply rooted in the soul (as she claimed in her ‘The Origins of Totalitarianism’), but rather to a complete unawareness of what his actions meant.”
Source: Wikipedia.
Reformulating this, centering on ChatGPT, it would sound like this:
“The text produced by ChatGPT… is not due to its own awareness, to its own ideas, deeply rooted in the software (in the network weights), but rather to a complete unawareness of what its actions signify”.
This is an excellent description of ChatGPT. What we use, as intelligent as it may appear, is nothing more than a tool. Firstly, created by man, and therefore with a whole range of nuances that we will soon explore. Secondly, it’s a tool that utilizes “only” advanced probabilistic computing functions.
The Beauty of Functionality
The beauty is that it works. And it works brilliantly. It functions both in terms of code and language. The aspect that generates the most astonishment and research is perhaps the emergent capabilities this tool has demonstrated. To simplify the concept of “Emergent Capabilities”, we can define them as the “ability to perform actions for which it was not programmed”.
Let’s Consider an Example.
How many of you have used ChatGPT for summarization and abstracts? It excels in these tasks because that’s what it has been trained for. How many have asked it to write code? It’s trained for that too. But what about using it to extract information and format it as desired? It might succeed on the first try if the task is simple, but for more complex tasks, especially those involving specific domain texts or entities that are not easily identified by rigid rules (like in the case of “regex” rules), you probably had to provide some examples.
Once provided with examples of what and how we want to extract information, the tool “understands the context” it operates in and adapts to the task. Scholars attribute this behavior to “Bayesian inference”. In simple terms, the tool alters its prior conception of the context, adapting to what we present as a challenge. Once it understands the context/terrain, changing its conception, it then acts accordingly.
This is done through a simple “prompt”, or a text input. Another interpretation of the prompt is its ability to “modify” the network weights, the same weights that determine the output produced.
Without delving into technicalities, let’s return to the concept of “garbage in” — “garbage out”. If we write a text input that touches certain elements, the tool will reflect back those same elements and concepts, often embellished, typically with pleonasms, but it won’t add anything new to the ongoing discussion. The same goes when we ask for an opinion; if we try to guide it, it’s very likely that the tool will follow that same path, rarely questioning it.
So, where does this discussion lead us? It’s not ChatGPT that offers us the best answer; it’s us who can unearth it. How? By presenting the elements, concepts, ideas, and paths that we think can be explored by a tool that has everything encoded in a multidimensional sphere (or hypersphere).
Visualizing ChatGPT as a Mosaic Artist
Imagine having an incomplete mosaic in front of you. Each tile we choose and place — through our questions, our prompts, and our data — represents the starting point, the initial ideas we give to ChatGPT. These tiles are crucial because they begin to outline the image we want to create, but on their own, they are not enough to form a complete picture.
This is where ChatGPT’s unique ability comes into play: like a skilled mosaic artist, this tool takes our initial tiles and integrates them with others, selected from its vast knowledge, to complete the image. It doesn’t alter the tiles we have given, but finds additional ones that best combine with ours, positioning them where they best fit to form a comprehensive and meaningful image.
In this way, ChatGPT works to complete the mosaic, using our initial information as a base and adding what’s missing to provide an answer that not only reflects our initial request but enriches it, shaping something we couldn’t have realized alone.
Here lies the power of this tool: transforming a collection of scattered tiles into a complete mosaic, revealing an image that goes beyond initial expectations, while always respecting the original ideas.
Extraordinary? Absolutely, but let’s remember that it operates on a probabilistic basis, so choosing the right initial tiles (ideas) closer to the expected outcome allows us to tailor its response to the context in which we are operating and which only we fully understand.
However, this is also clearly a limitation. The tool will never intervene with its own cognition of what to do to achieve the goal. It won’t question what has been reasoned up to that point. Given the context, it will try to complete the request without a real general understanding of what it’s doing.
Perfect, I would say we now have a vision of what the tool is, and how can it be useful for us? So far, we have said that it operates on a probabilistic basis, exploiting connections between ideas and concepts to develop a result capable of providing, according to the highest degree of probability, the answer to our question.
The Evolution of Technology and the Role of ChatGPT
Not bad at all.
Leveraging ideas and concepts has always been a human ability. Until now, technology has primarily focused on other aspects that extend human capabilities.
Machines were invented to allow humans to travel distances greater than what their legs could cover. The telephone was created to communicate with people beyond the reach of our voice.
ChatGPT, however, was invented to emulate the mind, and in various aspects, it succeeds. In its own way, as much as a software currently can.
I define this new tool as the first extension that humans have of their own mind, allowing, for the first time, to enhance our reasoning capabilities.
And see, the tool allows for the enhancement of reasoning capabilities, for finding inspiration and comparison, but only if used appropriately.
It’s not difficult, but it does require getting familiar with a tool that, if understood, can be of enormous help. If not conceived correctly, however, it may bring no value.
Well. Now that we have conceptualized and become more familiar with it, we can not only start using it correctly but also understand how and where it can be applied…
