Will artificial intelligence rebel? Expert: We have ways to hit the brakes early enough

The Infoshare 2023 conference, one of the patrons of which is Next.Gazeta.pl, is the largest celebration of technology and business in this part of Europe. Over 7,000 participants, over 220 technological startups and corporations, and nearly 200 experts from around the world came to Gdańsk to share how technology can help in business. We are currently witnessing a huge revolution related to the development of artificial intelligence, so it is not surprising that the entire conference revolved around this topic. Infoshare 2023 is also a great opportunity to learn more about the extremely rapid progress in the field of AI.

Bartłomiej Pawlak, Gazeta.pl: Where is the artificial intelligence? Does the average person already use it in practice?

Serg Masis, Syngenta: She is everywhere now. We may not even notice it, but she’s been all around us for a long time. For example, Netflix uses artificial intelligence in its recommendation system [filmów – red.] for some eight years. Facebook does the same with its feed [“tablicy”, na której wyświetlane są treści – red.]. When you ask Alexa something [asystenta głosowego firmy Amazon – red.], that’s where the AI ​​comes in too. Artificial intelligence also works at the airport, when you are scanned by facial recognition cameras or when you shop online, it is the artificial intelligence behind the recommendation system that tells you what you might want to buy next.

The thing is, it’s not something we’re directly involved in that we notice. Well, unless you’re writing an email and Google tells you the next word. However, we already take it for granted. In my opinion, what has captured the imagination of people is a type of AI that we call generative AI. Generative AI doesn’t predict something (e.g. weather), but creates new things based on what it knows. It is thanks to this that we can generate new photos, e.g. in DALL-E or Midjourney, or new text based on something. That’s what generative AI is all about, and that’s what has touched us more than anything I’ve witnessed in the last few years.

What exactly is this artificial intelligence? How it works?

It’s a broad concept. There are many ways and means of trying to emulate human intelligence, but our focus right now is machine learning, especially supervised learning and deep learning. So when we talk about ChatGPT, for example, we are talking about deep learning.

What is machine learning?

Machine learning is a bit like programming. You create a program, which in this case is called a model, and which is trained. So you have an algorithm that the computer follows to optimize predictions. Said ChatGPT has been trained to predict the next word. So you give him a piece of text and the word that comes next. The program starts connecting the dots. When you give him a whole lot of different examples, he starts to understand that this sentence always goes with this word.

For example, “You only live… (once)”?

Exactly. And then the program starts to understand how to accurately predict the next word. The way of learning, then, is to judge how accurate the match will be and find the right word.

So this is how artificial intelligence learns to somehow combine questions and answers?

As if. It’s trying to match inputs with outputs, which is training data with some label, in this case the word that comes last. Looking through many different examples, he finally starts to see the right connections. But we are talking about thousands or millions of examples here. In the case of ChataGPT, it was probably even billions.

After all, the possibilities of ChataGPT probably surprised even its creators?

Yes, and Sam Altman, CEO of OpenAI, admitted it himself. As I mentioned earlier, the model that ChatGPT is based on predicts the next word. It’s nothing revolutionary. But it has so much data that it can do it with pretty good accuracy. On this you build another model that can already answer questions. This is something that is a bit closer to a chatbot. But then you take these questions and answers [wygenerowane przez AI – red.] and you let people judge. Let them decide if they are correct or not. And that is the key to this development. And the result is a bit magical. They don’t even understand how they created things they couldn’t have foreseen.

Is the development of artificial intelligence something comparable to the industrial revolution?

Yes. But for me, it’s like the last revolution, when the internet started to count. I was a web developer back then and I felt the same as I do now. People suddenly wake up with such a novelty and try to use it everywhere. Of course, initially – at least in that case [internetu – red.] – it didn’t work for all industries. For example, then someone had the idea to sell pet food online, but no one wanted to buy it online. Now that has changed.

We can see the same things happening with AI. Many industries want to introduce it in different ways. Sometimes, however, people are not interested in this particular application, but perhaps they will like another. So we’re still going through that testing phase, figuring out what people want to use the technology for. For now [AI – red.] will not solve all problems.

And in the future?

Maybe. The thing is, it’s a general-purpose technology, just like the internet. Using it, you can solve many problems. But I still see a lot of work that needs to be done on privacy, trustworthiness, and more. If we can deal with it at all, we can improve our capabilities.

For example, I could talk to you in Polish now, if only I had a device that translates instantly. In fact, the appropriate technologies already exist, because we have, for example, Google Translate, we have text-to-speech and speech-to-text, but we need some kind of device that will combine them, adding, for example, body language recognition. So there are many ways in which we can still improve the technology, but this requires further work.

Can we improve artificial intelligence in such a way that it helps, for example, in medicine?

Of course. This is what I meant when I talked about increasing our capabilities. Nowadays, even with the best surgeon, there is still a certain amount of uncertainty in operations that are risky and require great precision. The technology could greatly simplify such operations, increasing the surgeon’s capabilities.

Should we be concerned about this rapid development of artificial intelligence? Do you think AI could potentially become self-aware?

I don’t think that is currently possible. Altman himself recently admitted in an interview that he wouldn’t be surprised if GPT-10 [obecnie najnowszym modelem jest GPT-4 – red.] has reached that level. However, I am inclined to think that with the current architecture it is rather impossible. I can’t imagine AI as anything more than a program that predicts your next word. Anticipation is actually part of human intelligence and all. But the missing piece is a unique “physical” consciousness that a machine simply cannot have. There is no awareness and ability to evaluate your own answers. He only knows after the fact [po przyznaniu oceny przez człowieka – red.].

By creating one model that evaluates the responses of the other model, we would be a bit closer. Though, of course, in reality it would be much more complicated. Perhaps then we would be able to reach some level [samoświadomości – red.]but not with the current architecture.

Maybe not today, but in 20 or 30 years?

It would be possible, but honestly, I don’t think we have anything to worry about. Because even if we were to create self-aware machines, they would still be just machines, and we could turn them off. Even if we had the technology to destroy ourselves, wouldn’t we be able to stop it? I’m inclined to think we’d rather handle it. We saw what happened during the pandemic and I think we learned a lot from that. We have many ways to hit the brakes early enough to prevent the AI ​​from going any further.

—