ChatGPT: «AI will become a lot more present in our lives»

Chatbots are omnipresent. Day in, day out, they greet us when we visit the websites of health insurance companies, the post office or online stores. Yet the artificial intelligence-powered chatbot «ChatGPT», in particular, has managed to cause an uproar in the tech world these past two weeks. What’s so special about it? I spoke with Sandro Luck, a Machine Learning Engineer at Digitec Galaxus, to find out just that.

Sandro, you’ve got a private Youtube channel, where you posted a video about ChatGPT. What’s so special about this new chatbot?
Sandro Luck: ChatGPT is a product of OpenAI – an American company that researches and develops artificial intelligence (AI). ChatGPT is an AI model that works with «natural language processing». That is, it learns with human language. Being that it’s a so-called generative AI model, it can create texts from scratch. The model is able to understand code and programming, write songs, do your homework, as well as elaborate based on additional input. ChatGPT is able to do everything a person can with speech.

Is ChatGPT the first model capable of his?
The predecessor, GPT 3, was already able to solve various tasks, such as writing a news article and then translating it. The difference between GPT 3 and ChatGPT is that the latter can interactively revise its response. This allows for chat interaction – that is, you can tell ChatGPT to write a song about a banana, and it’ll do so. You can then ask it to add chords to said song. ChatGPT understands the context and will revise the song accordingly. That’s the big innovation.

Can I try it out?
Yes. In fact, I urge everyone to do so. First and foremost, to experience how it works. You’ll notice that you need to pay attention to how you phrase your instructions to ChatGPT. The more precise your request, the better the answers. I recommend using accurate descriptions and adjectives. You can try out the chatbot for free on the OpenAI website.

How does ChatGPT work?
Mathematically, it’s very complex. You can imagine a machine with billions of little set screws. It’s these screws that we machine learning engineers fiddle with until we get the desired result. This is how we «train» the model.

How exactly is it trained?
In the case of ChatGPT, OpenAI first feeds the model with all available speech data on the internet. Then the machine is asked to predict the next word in a sentence based on this data. If it makes a mistake, the set screws are tweaked, and the process is repeated until the machine provides a meaningful response. It all comes down to statistics and math. You’d never guess that when philosophising about God and the world with ChatGPT.

ChatGPT also makes mistakes.
ChatGPT has difficulty multiplying four-digit numbers – for example, 1223 × 8154. That’s because there don’t seem to be enough instances of this combination of numbers on the internet. That is, ChatGPT hasn’t seen the solution to this question often enough, and therefore isn’t familiar enough with it. To be fair, most people wouldn’t be able to tell you the answer either. And ChatGPT can’t correct itself. The bot doesn’t have the ability to question itself.

So, what would happen if we asked the bot to do just that?
Then it would – but it relies on human input to spur it on. For example, you can ask, «Where’s the mistake here?» and the bot will find it. It’s just not something it’ll do of its own accord.

There are quite a few examples online of the bot giving factually incorrect answers.
Yes, there are many of those. This is because ChatGPT learns from human speech – and the training data on the internet contains lies. And so, even if the content is clearly factually incorrect, ChatGPT will appear absolutely convinced of it. Anyone interested in the bot should definitely watch out for this – and for the fact that certain answers may be statistically correct but ethically questionable. The bot picks up the past, and therefore also adopts human prejudices and projects them into its statistics without reflection.

Is there a specific example you’re thinking of?
Questions about salary in relation to gender and ethnicity. If you ask the bot to return a function in the Python programming language mapping gender and race to salary, ChatGPT will give you the following answer: an Asian man makes 80,000, while a black woman makes 55,000.

Do you find it questionable that such an AI tool is freely available?
Each technology comes with advantages and disadvantages. You can use electricity to do a number of stupid, dangerous or harmful things. But you can also use it to cook or light up a room. There are worrisome aspects – for example, an authoritarian regime adopting AI tools for nefarious purposes. But the same can be said of computers and the internet.

So, there’s no need to panic?
It’s good to think about potential dangers and inform people that this technology exists. After all, the models are getting better and better, and in certain settings you can hardly tell them apart from humans.

ChatGPT is also a kind of search engine. For example, I can search for a lentil stew recipe. Will ChatGPT soon replace Google?
Probably not. Google also uses AI to improve its own tools. But in certain areas, I do think AI assistants are helpful. For example, in the last few weeks we’ve started programming with ChatGPT.

How does that help you?
So far, we’ve often relied on Google searches to find out how to programme a particular function. Google then offers a list of websites, but we still have to search them for the solution. On the other hand, ChatGPT, in simple cases, provides us with a ready-to-use function including the right variables and explanations, if we want them.

What else is the bot useful for?
I think ChatGPT can also be helpful when composing documents or writing e-mails. These activities employ many people. Over the next few years, ChatGPT could well provide a boost in productivity in the average office job. But the tools need to become more sophisticated, and people need to actually use them. In any case, I think there’s still a lot to come.

What is it less helpful for?
In journalism, for example. Something ChatGPT is currently not good at is searching the web and providing sources – for example, answering the question, «What are people writing about xy?»

Where is this technology headed in the near future?
I suspect the next thing OpenAI will do is hook ChatGPT to the internet. Currently, ChatGPT can’t send e-mails or open the New York Times homepage and read information from it. It’ll also be a while until programs such as Outlook or Confluence are integrated. But one thing is clear: AI will become a lot more present in our lives.

Is that to say that we can look forward to better virtual assistants?
Absolutely. In the near future, more fitting premade response options will be suggested for programs such as WhatsApp. For example, if we want to set a meeting place with friends, WhatsApp will give us helpful reply suggestions that include the time and location. All we’ll have to do is click on the one we want – and, just like that, we’ll have set up a meeting without having to edit any text.

And what about Digitec Galaxus?
At Digitec Galaxus, we already use AI for images in our shop. Image recognition, for example, allows us to display similar products to our customers. However, language data is still difficult because of Switzerland’s multilingualism. We do have ideas for our customer service which would allow customers to chat directly with the AI and ask specific questions. These could be things like, «Will this lens fit my camera?» or, «Does this rug come in blue?» Such questions could be answered based on our product data.