---

Currently, there is a lot of discussion around OpenAI's ChatGPT, and it is popular to disparage AI as well as conjure up the end of humanity. But what is ChatGPT really capable of? And what is it unable to do? Does ChatGPT bring us closer to artificial general intelligence, or is all this hype unwarranted? We spoke with AI expert Pieter Buteneers about ChatGPT, language models, the end of Google, and the problems with OpenAI.

devmio: Thank you for taking the time to speak with us. Could you briefly introduce yourself to our readers?

Pieter Buteneers: I used to be the Director of Engineering for ML and AI at Sinch. Recently, I was promoted to Director of Sinch Labs, the team responsible for launching out-of-the-box new products. I also used to be the CTO of Chatlayer, a small company that provided a platform for non-technical people to create chat and voice bots. After Sinch bought the company, I developed AskFrank, a solution to facilitate the creation of chatbots. AskFrank can search content in your knowledge base and find an answer to a question there. We are still working on making AskFrank able to continue a conversation so that you can ask follow-up questions.

devmio: What is ChatGPT, and how does it work?

Pieter Buteneers: ChatGPT is based on the GPT-3.5 model, which is very similar to GPT-3. It is made up of 175 billion numbers which contain all the information that you can ask it about, which is basically the majority of the data on the internet. It's insane how little memory is required to store so much information.

GPT-3, however, can only finish your sentences. When you give it a cue, it will continue whatever the cue is. At OpenAI, they figured it might be interesting to turn this model into a conversational one that can answer questions, and that’s how ChatGPT was created. They fine-tuned GPT-3 in such a way that it can give answers rather than simply continuing your cue.

It’s basically a conversational summary of the internet up to the end of 2021.

To fine-tune ChatGPT, they basically had a person start a conversation as a cue and then watched what GPT-3 produced as an answer. Since the model is not entirely deterministic, you could sample those answers a few times. They would allow it to generate four different answers before asking a group of human experts to rank them. They did this for a large amount of data and then trained another model to predict the human expert ranking, in order to automatically fine-tune ChatGPT based on the ranking model.

They would then let it have imaginary conversations about millions of topics and use that ranking to further fine-tune it and get better and better results. That is how ChatGPT came to be. It's basically a conversational summary of the internet up to the end of 2021.

devmio: So training is done using reinforcement learning based on user feedback, but there are different layers between the model and the actual human feedback? Is it just two machines training each other then?

Pieter Buteneers: Yes! They chose that approach because it is much easier to understand how humans would rank something than to understand why they would rank something. ChatGPT had to figure out the why in order to change for the better. You need reinforcement learning because you can just say “this is good” or “this is bad.” You don't tell the system why, it needs to figure that out on its own.

However, reinforcement learning is really slow because it takes insane amounts of data. Since it's so simple to figure out how people would rank the answers, it's faster to just train a model that can do that and run it on a lot of data.

devmio: Is ChatGPT as impressive as it is thought to be?

Pieter Buteneers: I said it many times, and I’ll say it again: it's insane. If you had asked me two years ago if this was possible within the next 10 years, I would have said that it was very unlikely. That's how impressive it is.

The amount of information and understanding it has about different contexts is unimaginable. I use it over Stack Overflow and sometimes even over Google search because it is so incredibly knowledgeable about so many different topics. And the answers it gives, even though they are sometimes incorrect or completely made-up, are grammatically so sound and convincing. I can only figure out whether I'm being fooled when I already know the answer. It's that good.

It's even harder not to trust it because you don't know when it's wrong. It's grammatically correct, and most of the time it's right. And we are only talking about a beta release of ChatGPT, right? This is not their official release yet.

I said it many times, and I’ll say it again: it's insane.

When I tried ChatGPT, there seemed to be something hidden in the AI, where it can search the internet as well, so that you can have information from after 2021. Whether that's a something weird that the model produced to fool us or whether that's real, I don't know. But it seems to be that there is something in there that they're working on to connect it to the internet. And then Google will be wiped out in an enormous fashion. The only words I have for it is that it's insane.

devmio: If we couldn't have predicted something like ChatGPT happening, what is it that happened that made it possible to develop such a model? What was it, in terms of technology?

Pieter Buteneers: The most annoying thing about this is that it isn't that much. GPT-3 has been out there for about two years, and it was already very impressive. You could give it some cues and it would produce a piece of text that was fairly logical. However, it was more wrong than right.

They just took that model, fed it with fresh data, possibly tweaked a few parameters, and trained it with human feedback, and it became this insanely good know-it-all "person" you can talk to. And if you're not happy with the answer, you can ask follow-up questions just like you would with a real person. How little has changed to accomplish this is absurd. It only took fine-tuning in the right place and lots of computing power, which must have cost insane amounts of money. Apart from that, in terms of technology, not that much happened.

devmio: But isn't that model still probabilistic? It doesn't “know” anything. It just knows what the probable answer is, according to the data it was trained on.

Pieter Buteneers: Well, the same applies to you: Are you knowledgeable about a particular topic or are you merely guessing based on information that is stored in your head? It is exactly the same. Does ChatGPT really know things? Yes and no. If you consider that you don't really know things, it's the same for ChatGPT. But if you are completely honest with yourself, you know that you make summaries of reality. And you can condense reality into very high-level abstract concepts. ChatGPT does exactly the same. But does it matter? It keeps up the illusion that it does "know" and it's good at that.

Google will be wiped out in an enormous fashion.

Many people tried using ChatGPT for coding or writing poems. My wife, however, recently expressed interest in learning how to make yogurt. So I asked ChatGPT: “How do you make your own yogurt?” It gave us a recipe, although it wasn't really clear what tools you would need. So I asked: “What tools do you need?” And it gave me a list of tools. I was missing one thing, though: a reliable thermometer. So my next question was: “What if you don't have a thermometer?” It came up with another recipe where you can make yogurt without a thermometer. You would need to run a lot more searches on Google to find the same answer.

Of course, I also tried using it for coding. If I wanted to learn more about a new technology or tool, I would ask questions, and it would answer. I would ask again if it was using terminology I didn't like, and it would clarify it for me, provide examples, and so on. If you were to go down the Google rabbit hole, how long would it take you to find the same information? And that's what is so impressive about ChatGPT. It can conduct a reasonably intelligent conversation.

devmio: When we talked about Dall-E and the stable diffusion models, I asked you f this was the end for designers, artists, and other creatives. You replied that the models are merely tools. Does the same apply to ChatGPT and writers?

Pieter Buteneers: In its current state, it can undoubtedly help you complete your work a lot faster. You could, for example, ask it to write an introduction to this interview. If you ask the proper questions, it will work. You can then request that it rewrites particular sections or omits certain things. You can go back and forth, and it will write a nice intro or summary. It can be a nice tool, and it can help you out with these kinds of things. But it's not at the level where it can replace human creativity. It can't come up with completely new concepts. It can apply what it has learnt on fairly similar tasks but it cannot produce completely new things. So yes, some people might need to up their game.

But that's always the case when new technology comes around. Although it has not yet reached the point where every copywriter can be replaced, it will make copywriters more efficient and the texts cheaper. With this kind of technology, you might need fewer copywriters, or you might wind up with the same number of people doing a lot more work. The people who can use this technology to their advantage will benefit greatly.

devmio: Are there things ChatGPT definitely cannot do?

Pieter Buteneers: Well, the answer is always "not yet." What it can't currently do is look up information online or in a knowledge base so that you have answers that are recent. There are hints that if you ask the right question, something inside ChatGPT might be at some point enabled, that allows you to search the internet. That's most likely the next move for OpenAI. Just let it search the internet, and Google and Stack Overflow will vanish. It makes sense to assume that will happen. Additionally, there are models like AskFrank that can already perform a mixed search, so we know that's technically possible.

It also lacks the ability to tell you whether or not it is telling you the truth or just fooling you, which is another crucial feature. Maybe there are ways to train it to figure out how certain something is. There is likely a lot more work to be done, including a lot more manual labelling.

You can ask it “how do you do X,” but you cannot ask it to actually do X for you.

What it also cannot do is connect to something else, like another service or an API. You can ask it “how do you do X,” but you cannot ask it to actually do X for you. You can ask it to write a snippet of code that does X, but you cannot request that it connects to an API and stores data in a database. It has no connection to the outside, and I don't think sending data to an API will be possible anytime soon. However, there are indications that web search tools are being developed, which would be a killer feature in and of itself.

devmio: And then ChatGPT will kill Google?

Pieter Buteneers: In addition to killing Google, it will also destroy Stack Overflow. You should think of it as your personal assistant because it will be able to do a Google search on your behalf and only display relevant information.

devmio: Where are we heading with AI?

Pieter Buteneers: Researchers are likely looking into pretty much everything that we humans think is something that only we are capable of. We have demonstrated that AI has mastered language with the introduction of ChatGPT. Someone who was administering the Turing Test fifty years ago and was unaware of ChatGPT would unquestionably have concluded, "This is a human." Anything related to language that ChatGPT hasn't cracked yet, consider that as done.

Stable Diffusion is basically the link between languages and images. That was the first thing that happened, and now you can produce an image from a piece of text. Of course, you can mix the two. This mix of image recognition and language is now completely out in the open, and everybody can have a stab at it and crack this code.

The really interesting part about this is that we frequently use language to convey our reasoning. And because we have cracked language, all of our linguistic reasoning is undoubtedly within our grasp. It's frightening how close this is to artificial general intelligence. Although I wouldn't say we are extremely close, we are moving in that direction.

devmio: What are you looking forward to in 2023?

Pieter Buteneers: I wasn't a big fan of GPT-3. What you could accomplish with so few cues was impressive, but after seeing how they turned GPT-3 into ChatGPT, I started to wonder what GPT-4 would be like. I used to think it would add maybe just a bit extra on top of GPT-3. But now, I believe it will go well beyond that and be much more sophisticated. And I still find it difficult to grasp what that might be. Like, what more can it do? That I don't know, but I’m looking forward to finding out.

At Google, there is probably still a lot of head-scratching going on about what to do with ChatGPT. I don’t think they saw this one coming, even though they have been working on something similar. A few years ago, they had a fairly impressive demo on Google I/O with a voice bot talking about Pluto. But they can't be close to something like ChatGPT. What they had back then was far less advanced. I am aware from contacts within Google Deep Mind that there was a minor crisis among the Deep Mind team when GPT-3 was released. They were scrambling to come up with an answer asap. And that was only GPT-3. But with ChatGPT, I believe the entire company is trembling, not just Google Deep Mind. At least those who understand what ChatGPT can do are.

And because we have cracked language, all of our linguistic reasoning is undoubtedly within our grasp. It's frightening how close this is to artificial general intelligence.

At Google, they've known about GPT-3 for a while. They have been working on the technology, but they didn't come up with it. It's OpenAI that came up with this, and they'll have to work hard to prepare a response. They might create something comparable, but I doubt they'll outperform OpenAI.

devmio: Is all that fuzz about ChatGPT justified then?

Pieter Buteneers: ChatGPT, in my opinion, is the closest we've ever come to artificial general intelligence. That's the only thing I really want to emphasise. Many people in the world cannot do what ChatGPT does. Of course, they have social skills and a better understanding of emotions. However, what this is capable of is absolutely insane. It can tell you the most convincing lie with arguments that will baffle you. We shouldn't underestimate how important a breakthrough this is on the path to artificial general intelligence.

Of course, we're still a long way from AGI, but what I saw with ChatGPT has brought my prediction of when it will be reached 20 years closer. My guess was always around 2060, some researchers suggest 2030. After seeing this, I believe 2040 is more likely, which is unsettling to think about.

devmio: What is the one component of AGI that is still lacking? For ChatGPT to be AGI?

Pieter Buteneers: Out-of-context reasoning, coming up with new concepts or ideas. Strangely enough, not all people are capable of that either. Everyone is capable of doing it in a small context, at least generally speaking. What we as humans are good at is interpreting emotions from facial expressions and from how people speak. Those are things that ChatGPT cannot do. However, given how far image recognition has come, it seems clear that reading and understanding facial expressions and providing appropriate responses are not that far off.

We’ve also made such enormous progress when it comes to language models, so there are fewer and fewer things that these systems cannot do. Look at OpenAIs speech-to-text AI Whisper, which performs better than Google Cloud APIs, Microsoft, and AWS. It is also open source, allowing you to run it on your own computer and have a speech-to-text engine that outperforms the big players.

In the science fiction concept of personal robot assistants, we are technologically ages behind. However, everything else and the linguistic component are rather close.

devmio: There are also a lot of AI models that are open source, right?

Pieter Buteneers: Good thing that you mentioned this, because ChatGPT is not open source. It's frustrating. OpenAI started many years ago with the idea of democratising and opening up AI through the publication of all of their research, the release of all of their models, and so on. As a result, no one would be able to outcompete the others because everyone would be moving toward AGI at the same speed. And what do you see happening today? Nobody is moving toward AGI as quickly as OpenAI.

Of course, the open-source community is trying really hard to catch up, and they're doing a wonderful job. Also, many people are investing millions of dollars in that research, and I appreciate it a lot. However, OpenAI's decision to keep everything they do closed source and not democratise it as they had promised frustrates me.

We shouldn't underestimate how important a breakthrough this is on the path to artificial general intelligence.

I can understand that they fine-tuned their models in a certain way, so they are less misogynistic, and I understand why they make it hard to, for example, ask ChatGPT how to make a bomb. But by doing this, they are also making moral judgment, a judgment on what's right and what's wrong. And they don’t represent the entire world population. They are just a small team deciding for us what's right and what's wrong. And they do that for moral reasons. They also do it for the code. We are expected to put our faith in their ability to make morally sound decisions. So if there is anything about OpenAI that frustrates me, it is that it has the wrong name. It's AI for sure, but it's not open. They publish papers and blog posts or give us hints on how their models work. But they are far from being open about it.

Fortunately, when it comes to Stable Diffusion, an alternative to Dall-E emerged reasonably quickly, and the Stable Diffusion community continues to develop more features to make it easier to use and perform more interesting tasks with it. But it's clear that OpenAI isn't really democratising this technology. They also claim that they are afraid that someone will take it and do something crazy with it. We don't know, but it's possible that they are the ones doing it.