By Hanno Brink, Machine Learning Engineer at Synthesis
LLMs have come onto the scene and really shaken up the technology world over the last couple of months, and it is very clear that this technology unlocks a lot of exciting new possibilities. However, I feel like there seems to be a lot of fear and hype around the technology, and I think that it might be a bit premature. LLMs, as they exist today, are very good at two things: generating text and understanding natural language. Although we have seen a lot of hype around the potential impact of this technology, I do not think that either one of these innovations will change the world too much. I am also acutely aware of the danger of making predictions like this about technology.
To test the extent to which current text generation capabilities will impact the world, I have tried to use it to write an article, write a manual for a board game, write relatively complex code, and finally to document code.
When writing an article for publication, AI can provide you with a summary of the current thinking and definitions, as well as edit the text well. At the moment, we still have the problem that the technology often makes false statements, but even if we solve this problem completely, it only echoes current thinking and formats it nicely. Those things, however, do not really make for an article worth reading. What really makes an article worth reading are the brand-new ideas that get captured on the page, things that extend or go against current thinking.
When writing a manual for a board game, the AI does not know the rules, and explaining the rules up-front is the equivalent of writing a manual. AI is great at editing the manual, but the rules need the human element, which means that there is no real change in the way we work when creating brand new documents.
Code is interesting because the code AI generates for very small problems is mostly good enough. However, as soon as the problems reach a certain complexity, AI falls flat. This means you spend your time debugging unfamiliar code rather than thinking about code upfront and typing code. You might be thinking that this is the problem with LLMs, but I expect it would be able to do this flawlessly soon. The real problem is that code always solves real-world problems and lives in complicated domains. Code captures and formalizes real-world business processes. Processes that are so unique and convoluted where a language model won’t be able to capture all the context effectively. So, it might save you a few visits to Stack Overflow and help you type out code more quickly, but a developer would ultimately still need to understand the problem domain and come up with a good solution.
Finally, when documenting code, I have been very impressed by LLMs’ ability to identify how the code works, but what we capture in our documentation is often not how it works, but rather the real-world problem that it solves. This is something that AI cannot give you without a large amount of context about the problem domain.
From these experiences, I do not think that text generation will significantly impact most of what humans do on a day-to-day basis. I expect that the days of text that has not been edited by AI are pretty much over, but I also predict that with text becoming so cheap, we will soon value brevity and directness a lot more. That does not mean that we will not see any disruptive use-cases of text generation. Chatbots, video game NPCs, and interview coaching are just a few of the exciting new innovations we will probably see more of.
However, LLMs are also very good at understanding natural language, and this means that we can now interface with existing technologies using natural language. This should make tools with steep learning curves much more accessible to the layperson. I can imagine telling Photoshop to remove the lamppost from the photo of you and your friend on holiday and it magically happening, but as soon as the requirements become more complex, then using natural language will become quite cumbersome or even completely inadequate. This is because the way that these tools have evolved over the previous decades has led to interfaces that are optimized to be as simple, fast, direct, and unambiguous as possible. These are all properties that natural language does not really possess. Natural language is very messy, verbose, and ambiguous, and I suspect that anyone with some competency in using the software tools would much prefer the more precise control that the traditional tools give you.
In conclusion, LLMs have certainly already had some impact on how we write text and will soon start impacting how we interact with technology. There might also be odd cases where understanding and generating large amounts of text could lead to significant productivity increases or brand new innovations that change specific industries. However, from my experience with these tools, I think these use-cases are not nearly as common as the hype train would have us believe, and I reckon that our jobs would, for the most part, stay pretty much the same. Ultimately, it seems that the human element remains critical for generating truly innovative ideas, solving complex problems, and providing precise control over software tools.
Now, off to ChatGPT to see what it thinks of my writing…