Have you also been swept up by the hype of ChatGPT? Maybe you’ve had long discussions with colleagues, clients and even management on topics such as:
- Do I need to be worried about my position?
- Is ChatGPT really that revolutionary?
- How can we catch the opportunities of ChatGPT?
This is an NLP insider’s view on ChatGPT and how it will affect the world… of media. By giving my view, I hope to answer some of your questions and alleviate your worries.
To set the frame for this I want to mention two groups in the ongoing discussion. By now, most of you know that the internet is filled with various claims regarding ChatGPT and how it will affect the world. As always, some new people are convinced that AI is finally going to affect our everyday lives, as if that wasn’t already happening. This is the first group.
Other people, often NLP insiders such as myself, are acknowledging ChatGPT for its impressive features, but are not surprised since this is part of a journey of similar progress in the last few years since BERT and the rise of Transformer models. While ChatGPT is impressive there are certainly limitations and flaws such as factual accuracy to consider, as well as competing models. This is the second group and my target audience for this article is probably somewhere between these two groups. Now that we have straightened that out, let’s move on.
How will ChatGPT change the world of media?
I want to start by taking a step back. A few months back I found a journal article that despite being a few years old had an interesting image foretelling Natural Language Processing (NLP) research evolving into Natural Language Understanding (NLU) based on three overlapping evolution curves. The curves are described in a couple of ways and I find the Bag-of terminology to be the most intuitive. In this terminology the three curves are called Bag-of-Words, Bag-of-Concepts and Bag-of-Narratives.
Looking at this now and in terms of the capabilities of NLP technology and models today, we have come quite far in many areas, but we still have a long way to go until machines actually reach NLU. This means that we, as humans, still hold a massive advantage over AI/machines in certain areas and will continue to do so for a long time.
We can relate this back to Do I need to be worried about my position?
I would say, as long as your work involves fact checking, a degree of uniqueness/creativity or other complex tasks that require NLU, then you should be safe. However, the competition may very well augment their capabilities using ChatGPT or similar tools and that may force you to do the same in order to keep up. In general this means that most journalists should be able to keep their jobs.
Narrowing down to ChatGPT and its capabilities, I want to start by saying that it is impressive in the way that it can seemingly tackle multiple different problems and the quality of its output in terms of grammar and flow is really high. It can even compete with us humans when it comes to writing interesting text. However, there are issues and limitations. It is limited to what it has been trained on and it often can’t separate right from wrong.
Talking about ChatGPT’s flaws leads us to another group of people. They have made it their mission to find flaws in ChatGPT. Sometimes taking it as far as focusing on issues that are already widely known. For example, some devalue its technological capabilities because it isn’t factually correct and find lots of examples of this. The thing is, OpenAI is very open about the issues regarding ChatGPT’s factual accuracy. We know it isn’t perfect, like pretty much everything else in this world. However, somehow, many people seem to work around this and make practical use of it. The world simply isn’t black and white.
If we loop back to the question, Is ChatGPT really that revolutionary?
In general, the transformer technology itself has had a much greater impact on the world of NLP, however considering PR and exposure outside of the inner sphere of technology puts this in an entirely different perspective. I think ChatGPT has made the biggest AI publicity splash in recent years and we shouldn’t underestimate the effects of such a vastly successful PR campaign. In short, ChatGPT is obviously an attractive example of how to apply AI to generally applicable problems. I think that is pretty darn revolutionary as I, working in an AI company, still find it difficult to explain what I do to my family.
Moving forward to the third question, How can we catch the opportunities of ChatGPT?
We at iMatrics are always exploring new and interesting technologies. Since ChatGPT was met with a strong reaction, we started testing it early on and have given it a bit of extra attention. In short, it is a really fun tool and we have already managed to make some use of it. Below I have listed some pros and cons based on our experience that is applicable for anyone working with articles. This certainly applies to our friends in the news media.
👍🏻Pros: Successful use cases.
- Generating quick drafts/outlines.
- Suggesting rewrites when you aren’t feeling inspired, or just can’t get that paragraph to feel right.
- Reformatting an existing article for other channels.
- Code completion/generation.
👎🏻Cons: Identified issues.
- Fact-checking (Obviously).
- Original content creation, as part of your unique voice.
💡Considerations: Key actions to mitigate issues/risks.
- Careful fact and source checking.
- Verifying that the message and tone are in line with your goals.
- Don’t overindulge in quantity over quality.
Now you may wonder, is code completion/generation really relevant for people writing articles? Yes, maybe not for everyone, but certainly in the broader sense. There are plenty of situations where having an AI coding assistant to help you crunch some numbers can turn a mediocre story into a great one. Especially when coupled with some savvy visualizations, which can also be created using a script supplied by the aforementioned assistant. Imagine having your own data scientist.
iMatrics ❤️ ChatGPT?
To wrap things up, I want to relate ChatGPT to the business we do here at iMatrics. Today, we do not see a reason to add ChatGPT to our core product, the auto-tagging service. That is not what ChatGPT was built for. However, we are evaluating other transformer-based models from time to time as they hold promise in the auto-tagging domain. So far, they are significantly slower to train than our current technologies. The operating costs are also much higher and the overall quality improvement is not worth it.
Now you may think, are you guys just trying to skimp on costs to improve your margin? No, the discrepancy between the quality improvement and calculation speed is so vast that it doesn’t make sense. It would either slow down our service a lot or cost a lot more to run. On top of this, our clients probably wouldn’t even notice the quality boost. Considering things from an environmental stance is even worse. It is neither responsible nor sustainable for us to increase the load so much for a barely noticeable improvement to our product.
This doesn’t mean that ChatGPT or similar models aren’t interesting for us. We see that there are possibilities like adding text generation and/or rewrite tools in many places to help our clients’ writers become more efficient, but it needs to be done in the right way so that the output is always verified by a human. This type of process optimization resonates quite well with our existing business model and this leads us back to the question: iMatrics ❤️ ChatGPT?
We think that it is an impressive tool and love to play around with it. However, we still don’t know if it should become something more. I would love to discuss worries, opportunities and project ideas together with you, our future and current clients and partners. Please comment and/or reach out directly to me with your thoughts on the matter!
I tried asking ChatGPT for a few starting hook suggestions to this post.
Not too bad, but it isn’t quite what I had in mind. I will explain why for the first one. It looks like an overly long headline of an article with a different focus.
However, this actually inspired me to describe myself as an NLP insider. Besides, if the first try doesn’t work you can just ask for a bunch of suggestions. Odds are, that you either get an appropriate answer eventually or you get enough inspiration to come up with your own that matches your unique vision.