ChatGPT likely created the title of this post. I confess I’ve been highly intrigued by how one can integrate Large Language Models (LLMs) into workflows. I see lots of TikToks and Youtube videos using large language models to create net-new content for business, but I’ve found they are really helpful for me in transforming content. In essence, I can write an article (like this one) and then, through a series of automated steps, create the title, description, and social media posts automatically. They have also been very handy in the video side of work for clients, specifically taking in a transcript of a video and then transforming it into another type of content, like a blog post, social media posts, etc.
As NVIDIA, one of the pioneers in the computing that powers LLMs explains,
Large language models largely represent a class of deep learning architectures called transformer networks. A transformer model is a neural network that learns context and meaning by tracking relationships in sequential data, like the words in this sentence.1
So essentially, these LLMs work by predicting the next word based on what it’s prompted with, and it keeps predicting the next word until it comes to the point of completion as defined by the prompter. They predict the next word by ingesting lots and lots of examples, training data as it’s called. Essentially, the model is fed with many inputs of preferably high quality and various examples of text to learn patterns or heuristics.
There’s no shortage of think pieces and good writing on this process and its social implications, and I’ll link to some of my favorite ones at the bottom, but for today, I’m interested in thinking through some of the underlying assumptions that guide us to build LLMs in the first place. Neil Postman is not around today to tell us what he thinks about LLMs, but I think he and other media theorists provide an excellent way to approach the subject.
LLMs, and our embrace of technologies like Bard and ChatGPT, point out our implicit acceptance of statistical thinking as a good way of making decisions. We don’t usually say it that way, but the acceptance of that fact is implied in how we behave. Postman, in Technopoly, calls out such tools as invisible technologies, and he uses the example of a question itself as an example of one such invisible technology:
Once upon a time, in a village in what is now Lithuania, there arose an unusual problem. A curious disease afflicted many of the townspeople. It was mostly fatal (though not always), and its onset was signaled by the victim's lapsing into a deathlike coma. Medical science not being quite so advanced as it is now, there was no definite way of knowing if the victim was actually dead when burial appeared seemly. As a result, the townspeople feared that several of their relatives had already been buried alive and that a similar fate might await them. How to overcome this uncertainty was their dilemma.
One group of people suggested that the coffins be well stocked with water and food and that a small air vent be drilled into them, just in case one of the "dead" happened to be alive. This was expensive to do but seemed more than worth the trouble. A second group, however, came up with a less expensive and more efficient idea. Each coffin would have a twelve-inch stake affixed to the inside of the coffin lid, exactly at the level of the heart. Then, when the coffin was closed, all uncertainty would cease.
The story does not indicate which solution was chosen, but for my purposes the choice is irrelevant. What is important to note is that different solutions were generated by different questions. The first solution was an answer to the question, How can we make sure that we do not bury people who are still alive? The second was an answer to the question, How can we make sure that everyone we bury is dead?
Questions, then, are like computers or television or stethoscopes or lie detectors, in that they are mechanisms that give direction to our thoughts, generate new ideas, venerate old ones, expose facts, or hide them.2
Postman’s story goes a long way in helping us understand the situation here. The form of the question (“How can we make sure that we do not bury people who are still alive?” vs. “How can we make sure that everyone we bury is dead?”) dictates the response powerfully. We see this in our everyday experiences all the time. But we don’t explore how the mechanisms we employ, like ChatGPT, give structure to specific kinds of outputs. In the case of ChatGPT, it will give rise to the most statistically probable result by definition.
So, is this a good thing or a bad thing? Of course, it’s context-dependent. As Postman points out:
…a more practical question is, To what extent has statistics been allowed entry to places where it does not belong? Technoploy, by definition, grants free reign to any technology, and we would expect that no limits have been placed on the use of statistics.3
So, by society creating and using LLM-powered content generation tools, we are conceding a point: that statistical probability is an acceptable way of creating content that can form human minds. As my article introduction shows, I have mixed feelings about this question.
On the one hand, there’s a lot of work taking content from one context and transforming it into content that can be used appropriately in another context. I think these tools are beneficial. I hate writing social media posts from long-form content, but an LLM is happy to ingest my article and suggest three options I can choose. Even as I write this article, Grammarly is popping up with its red and blue underlines, suggesting how to re-phrase things to make my points more straightforward (I initially said “more clear,” but Grammarly thought clearer was better).
GPT-4 might even do a better job at getting right to the point. When I asked it: “Write a paragraph summary about what you think Neil Postman would think about Large Language Models,” it told me:
GPT-4 diagnoses the problem pretty well there, but that’s just it. It does pretty well. It’s literally a very average response, a statistical prediction that arises from the model having ingested plenty of the work of Neil Postman. There’s no stand-out insight there, and that’s alright in most circumstances. But it doesn’t have its human perspective. We read that lack of humanness as objectivity, but Postman himself reminds us that’s not the case. The way the prompt was worded, the statistical methods used to build the language model, and the language included in the model's input are all subjective. We transfer the “statistical = objective” mindset onto the LLM itself. Still, we must never forget that if the medium is the message, and all messages come from a particular perspective, there are no neutral mediums.
Postman has much more to say about this and related subjects in Technopoly, but I think his quote sums it up nicely:
“When statistics and computers are joined, volumes of garbage are generated in public discourse.”4
Thinking about society that’s my biggest concern. While LLMs stand to make a lot of facets of work more accessible: the ease with which we let technologies like this infiltrate all aspects of life without critical thinking will lead to these technologies being used in ways that are negative toward society. That’s a fact.
Does that mean we shouldn't use them at all? To be honest, until I’ve been writing this article, literally this sentence, I’ve not considered it too much. I use it (and as my creative partner Diana can tell you, I always look for new uses for it) because it's a new tool presented to me that I can use to save a lot of effort. That’s the power of invisible technologies. We adopt them and their underlying presuppositions without thinking. But becoming aware of the choices we are making is half the battle. The other half is maintaining awareness over time as new advances blind us to these truths.
So I can’t tell you if you should use them or not. I will probably keep using them but will do so by working to be mindful that, regardless of how innocuous these things seem, all technologies have trade-offs.
With that, thanks for reading, and see you again soon.
Resources:
I’ve really enjoyed the Hard Fork podcast in general on this subject. Here are a few episodes that have formed my thinking:
I have talked about Technopoly before, but in re-reading Chapter 8 to write this article, I can’t recommend it more strongly.
Social image by: Photo by Mohamed Nohassi on Unsplash
Technopoly, 126
Technopoly, 129
Technopoly, 137