By the time you read this, everything will probably be a lot better, faster, smarter and connected to other amazing services, systems and tools. That’s why I’ve been putting this article off for so long since its launch on November 20, 2022. 🙂
What is ChatGPT?
Okay, let’s get to it. What exactly is ChatGPT? Let’s try asking ChatGPT directly. The first answer will be something like this:
“I am ChatGPT, a language model created by OpenAI. I was trained on a large dataset with text using a deep learning algorithm to be able to understand and generate human-like responses to a wide range of inputs. I can answer questions, provide information on various topics, or have a conversation with you. As an AI language model, I have no emotions or personal beliefs, and my answers are based solely on the data I was trained on.”
Right… But the magic happens when I ask again, but this time – explain it to my grandmother:
“Hi, Grandma! ChatGPT is like a computer program that can talk to you in a way similar to how people talk. It can answer your questions and have a conversation with you. It works by using a lot of information that it has learned from reading many books, websites, and other things that people have written. This helps ChatGPT understand what you are saying and respond in a way that makes sense.”
And that’s the point. That’s why the whole world is freaking out. And so am I! At OpenAI, they simply took enough training data to create a large language model whose communication is of an excellent standard.
Why is ChatGPT lying to us?
Despite all of this, ChatGPT lies… very convincingly. In fact, it’s still good to keep in mind that this is not artificial intelligence as we know it from the movies – it’ll probably be a while before we create Holly from the Red Dwarf. If we really simplify it, the primary purpose of ChatGPT is to put words into sentences and sentences into paragraphs, in a way that sounds as natural as possible. The fact that it happens to answer your question is a bit of a side effect. Well duh, since it went through the whole Wikipedia. So ChatGPT is doing great, but it’s missing the fact-checking that the next generation should be able to deliver.
Having spent many years as a developer, I haven’t missed the early fascination with ChatGPT being able to find bugs in code or being able to explain or even generate code. But…! A couple of simple demos were no problem, but the story of StackOverflow, the best friend of every developer, showed the true colors of ChatGPT. They simply disabled it because users started using it to generate responses. It was very easy to do, but the veracity of such posts was simply too low and at the same time they sounded totally convincing. A lethal combination.
How we don’t use ChatGPT
At first glance, it would seem that chatbots are doomed. We simply deploy ChatGPT everywhere, throw internal resources at it, and it answers everything. We’ve been hearing a lot from our clients, and not only from them, in the last few weeks, “What do you think of ChatGPT, it’s kind of putting a spoke in your wheel, isn’t it?”. Well yes, and no.
There are a few catches to this, though. As has already been mentioned, it can go astray very quickly and in conversations, it would start to say things that our clients nor we would like. We like facts and conversation under control, and in that respect, ChatGPT so far has been a wild horse that either wins the “Velka Pardubicka” steeplechase or flies into the audience and knocks over a hot dog stand.
Another problem, that is, the fact that the oldest data used to train this model is from 2021, has been solved by Microsoft in the meantime in their Bing search engine. They have used the connection to its current database and can therefore include hot news in the conversation. However, even this has not been without significant problems. If you’d like to know more, check out this article.
How we actually use ChatGPT
Okay, okay, lots of problems, but there are already plenty of situations where you can already use the current version. How about generating a bot personality based on the existing text on the client’s site so that everything fits together nicely? Yes, that’s precisely where we’re already reaching for ChatGPT today and trying to give it all the context and background info that makes sense. In return, it might suggest a name and character for the bot and a few sample texts. Then our conversation designer takes over and fine-tunes everything to perfection. It rarely happens that the output is usable straight away. It needs to be gradually tweaked and tested. There are also cases when it simply creates more work for us and it’s better to build everything the old-fashioned way.
How about generating a set of possibilities for how can the site visitor ask a certain question and train the NLP model in the process? Because if the client doesn’t have a conversation history to tap off of, it’s up to the imagination and experience of the conversation designer. And sometimes, it physically hurts to sweat out enough variations because people like to ask really creative questions. As a result, the very first version of the NLP model deployed into production is again a bit better prepared.
Or maybe you could generate texts for each target group based on its data? I could go on, but then you probably wouldn’t read it. And yes I haven’t gotten into how it helps us in other departments that don’t deal with creating bots, but all the necessary sauce around it – the structure and wording of a newsletter or a business email, a webinar or a presentation… But I think you’ve already got a peek at this from a bunch of other articles and posts.
A word to conclude
ChatGPT is definitely revolutionary and will continue to surprise us for a while. In fact, OpenAI is gradually launching new versions of the models that we all helped improve by using ChatGPT.
Anyway, this is just the beginning of the ride we’re on. One thing is for sure, a whole lot of people are already using ChatGPT to do their daily work and make their lives easier. And that’s what it’s all about.