ChatGPT is the latest language model from the AI giant OpenAI. The latest version of the conversational AI is freely available to users who have an OpenAI account. We spoke with Christoph Henkelmann — an expert in AI, machine learning, and deep learning — about how ChatGPT works, what makes the model stand out, and whether the excitement about it is justified.

devmio: Thank you very much for the interview. Can you briefly introduce yourself to our readers? Who are you, and what exactly do you do?

Christoph Henkelmann: My name is Christoph Henkelmann, and I currently live in Cologne, Germany, where my company DIVISIO is also based. I co-founded DIVISIO, and I am its CTO. We focus on AI, machine learning (ML), and deep learning (DL), with an emphasis on practical feasibility. I got into AI during my “classic” computer science studies in Bonn, which included a lot of math and theory. Then I took a detour into the world of Java Enterprise development before I ended up in AI and ML professionally.

devmio: Then let's get straight to the topic. What is ChatGPT? How would you describe it?

Christoph Henkelmann: Formally, ChatGPT is, as the name suggests, a large language model based on GPT-3.5, fine-tuned for dialog flows. This may not sound very helpful at first, so let me elaborate a bit.

GPT stands for “Generalised Pretraining for Transformers.”GPT is a family of architectures and various resulting models. Many neural networks follow this architectural pattern. GPT models are trained to simply complete text. They perform the same functions as old Nokia phones with T9. GPT attempts to predict what the next text block will be as you type.

To train GPT, you collect vast amounts of text, for example, with web scraping. Then you let a neural network make predictions, sometimes for months. For instance, if you begin a sentence with “The snow in front of my house is,” it will most likely say "white" - or maybe "yellow." This is how GPT learns to continue or complete text, as well as how to answer questions. In these vast amounts of text, the model sees questions and answers.

If I prompt GPT with “Question: What is the capital of France? Answer: ...”, it will complete this text. It doesn’t even understand that it’s answering a question, it simply wants to complete or continue the text. This is also the problem. GPT will continue the text even if it has no answer. The model is not able to reflect that. It cannot say "I don't know" because there is always text that will be a statistically likely answer for GPT.

It’s not a system with a reflected consciousness like in science fiction movies. It’s a stochastic system that compiles responses based on previously seen data. ChatGPT is simply a GPT model that’s been adapted to make sure that continued texts always follow a dialog structure.

devmio: Can you give insight into how they made this adjustment and how it was fine-tuned?

Christoph Henkelmann: Fine-tuning is the process of adjusting an existing neural network that has been trained on a general task, to a specific task, such as the aforementioned text continuation. The focus is thereby shifted to a very specific problem with a much smaller data set.. OpenAI accomplished this through reinforcement learning with human feedback (RLHF). This technique isn’t normally used in the context of language models. In ChatGPT, RLHF was combined with other techniques to achieve the desired effect.

In reinforcement learning, a neural network can operate in a virtual "world" that gives it feedback. For example, this can mean solving physical tasks in a simulated 3D world. But a world can also be fairly abstract, like the “world of chat histories.” This is the world in which ChatGPT "acted" during its training. Normally, I’d like a world that gives feedback automatically, through simulated physics or something similar. But that’s not possible here, as the feedback needs to be provided by humans, as in this case, you need a world where GPT learns to have good chats. In this context, "good" means correct, ethical, moral, and non-toxic answers.

Since this is typically a task for humans, it relies on human input and data, which cannot be fully automated. The way the OpenAI team did this is very clever. First, they trained another neural network to evaluate the quality of an answer. Since GPT could already generate answers with the existing model, creating artificial chat histories was already possible. GPT then answered these questions multiple times, and the alternatives were rated by humans. These ratings were used to train an auxiliary teacher model that, like in figure skating, could score the answers. It attempted to mimic the taste and intuition of human reviewers.. This model also didn’t really understand what it was doing either. It only learned which answers appeared to be good, not which actually were good.

Finally, ChatGPT was permitted to conduct dialogues in the dialogue world, and the quality of the answers was evaluated using this teacher model. If ChatGPT gave a good answer, it received positive feedback. If it gave a poor answer, it received negative feedback. ChatGPT was automatically fine-tuned on chat discourses by going through the teacher model, and it can now independently conduct chats that look as realistic as possible.

ChatGPT is simply a GPT model that’s been adapted to make sure that continued texts always follow a dialog structure.

devmio: So there are some levels of abstraction between human feedback and the final product?

Christoph Henkelmann: Human feedback was indeed included, and a lot of it. But ultimately, a detour was taken by pouring human feedback into a teacher model that did the actual training. OpenAI trained a model, which was then used to train ChatGPT. Using a "teacher model" allowed ChatGPT to be trained on orders of magnitude more data than would have been possible with only human feedback. I don’t know the exact factor, but I assume you can generate an estimated 1,000 to 100,000 times more training data using the teacher model than with humans alone. This more than compensates for any potential loss caused by abstraction levels. Even though there is some loss, you get better results.

devmio: Does quantity beat quality here?

Christoph Henkelmann: Absolutely. That is (practically) always the case with deep learning. It's true that the most beautiful results come from using less data because it gets us closer to where we want to go in AI. But, in the end, that's not how it works right now. I once tried boxing and trained at the Bonner Box Club for a short time. "Endurance beats technique," said a large poster hanging in their gym. Unfortunately, this is still the case with deep learning. Before we can beat large amounts of data, we need to put a lot of brainpower into it.

devmio: Is there anything ChatGPT isn’t good at? Have you noticed anything?

Christoph Henkelmann: There is no such thing as 100 percent accuracy in ML and DL. Perhaps a very reliable model can generate 99.999 percent correct solutions, but maybe only 80%. It will always generate a certain number of errors. You must decide what problem you want to solve and what percentage of incorrect answers you can live with.

The error rate with language models varies greatly depending on the problem. We can, for example, solve simple grammatical text problems with incredible accuracy. This used to be incredibly difficult. However, ChatGPT fails at simple logical conclusions and more sophisticated tasks. For instance, I described a physical store to ChatGPT, including the opening hours. "On weekdays, the store is open from 8:00 to 18:00." "Please tell me: is the store open at 3:00 p.m. on Tuesdays?" Although the answer appears to be obvious, ChatGPT was unable to provide an accurate response.

But when I changed the wording, I got the correct answer right away. You can see how sensitive the system is and how it reacts to different phrasings. This is where the statistical nature of the model becomes apparent. The answer is the result of training on a lot of data, not of a conclusion. ChatGPT observed texts about opening hours and learned from them. The more closely my wording matched what it has "read", the better the answer.

devmio: The model is based on what it reads. So it doesn't “know” anything?

Christoph Henkelmann: It is true that very large models exhibit so-called emergent properties. Additional problems can be solved as the size increases, and sometimes even hints of reasoning begin to show. But the question remains whether this is simply due to the model's size, which can simply remember much more and thus has a better chance of memorising the correct answer, or if they are truly capable of drawing conclusions.

So, for the time being, I would not commit to a single interpretation. However, with each new model, it becomes clear that in order to progress, data and compute must grow by an order of magnitude. When you plot model size and performance, you get a curve that flattens out toward the end as the model size grows. Even though we don't know for sure, the general consensus is that GPT-4 will have trillions (plural!) of parameters. In comparison, GPT-3 has approximately 175 billion parameters. So there will be at least a tenfold increase. GPT-3 is also about ten times larger than its predecessor, GPT-2. Capabilities increase linearly, but resource consumption increases exponentially. Ironically, this is the polar opposite of "singularity!"

devmio: To outsiders, similar to DALL-E or Stable Diffusion, this looks like pure magic. Are the results still impressive to you? Or was it just the next logical step in foreseeable development?

Christoph Henkelmann: It was not the logical next step. I'm surprised by how quickly these things are happening, and I'm sure most experts are as well. What is happening continues to astound those who study the subject on a daily basis. Dall-E and Stable Diffusion also left an indelible impression on me. If you had asked me shortly before those models were released how long it would take to achieve that performance, I would have said "years." So, despite the issues and limitations I mentioned, ChatGPT impressed me greatly. I continue to be amazed.

devmio: What impact will these developments have on us as a society? What about texting and coding, what about art and culture?

Christoph Henkelmann: It will change things, without a question. How exactly? Such forecasts always strike me as extremely dangerous. I was very much surprised by Stable Diffusion and Dall-E, so I doubt I'll be able to predict where we'll be technologically in two years. However, I am aware that fundamental issues, such as logical inconsistencies, have yet to be resolved. I think there are at least one or two fundamental ideas still missing that we are not even researching. If they surface, then there might be another big leap forward towards “real” AI. But I have no idea what those might be or when they might come to fruition.

With the current state of technology, we will already see changes in art, school, among artists, cultural work, copywriters, and other creative professions. The technology does not require much further development; it is nearly complete. There will be new art forms and creative processes, and some jobs will become obsolete. Because of the invention of photography, fewer oil portraits were painted, causing some people to lose their jobs - we will similar effects with these new AI systems.

Certain “simple” artistic tasks will indeed be automated. If I only need a small logo for my startup, the AI solution will most likely suffice. There will simply be another artistic tool, based on AI, similar to Photoshop's magic wand or blur tool. Artists will not be replaced, but they will be able to create things that are far more complex and distinct in a much shorter amount of time. Good artists will continue to outperform laymen or standalone AI because they will be able to use these tools in ways that people like me cannot because I am not an artist.

So, if we begin to entrust these systems with social decisions like whether an applicant is the best fit for a job opening, evaluating legal decisions, or even autonomous driving too soon, we will cause significant damage.

devmio: Would you say there are legitimate concerns and fears about AI development?

Christoph Henkelmann: Fortunately, the discussion about realistic potential risks is slowly gaining momentum. The problem is not, and we need to be clear about this, the so-called singularity (an exponentially learning superintelligence). That, I believe, will never happen. I do believe that at some point, whether in five, fifty, or 500 years from now, we will be able to develop self-aware intelligence capable of keeping up with or even surpassing us in some areas, but it will not be an exponentially learning, god-like creature.

So the problem with AI is not Skynet or anything like that. The problem with AI is that many people do not yet have an intuitive understanding of how it works. No one is concerned about chess computers taking on a life of their own and conquering the world because we know roughly what their limits are. This type of AI has arrived in society, but laypeople are still unable to assess what current Deep Learning-based AI is capable of. This means that things an AI can't do are frequently overestimated or misclassified. We believe that because a Deep Learning system is so good at predicting protein folding, it must also be capable of determining whether a refund at a supermarket is justified — which is simple for us but much more difficult for it. So, if we begin to entrust these systems with social decisions like whether an applicant is the best fit for a job opening, evaluating legal decisions, or even autonomous driving too soon, we will cause significant damage.

Another problem is that many people believe an algorithm or machine is neutral. The Deep Learning algorithm that powers the system is, in fact, a mathematical construct that is completely neutral. The data used to train the systems, on the other hand, is not. It transfers all of the biases, discrimination, and problems present in the data into the system one-to-one. The system is unable to detect these issues. We all have biases, but we can reflect upon them and, hopefully, correct them over time. When we make social decisions that will significantly affect people’s lives, we have to weigh them very carefully. Certainly, you can use these systems. However, you must be aware of their problems and consider how high the error rate is in each case. Is it greater than that of a human? What is the consequence of an error? Is it necessary to involve a human in the decision-making process at some point? Or is the subject simply too delicate to entrust to an automated system?

A beautiful expression comes to mind: machine learning systems are a lever for the human mind. Just as we discovered mechanics thousands of years ago and used levers and pulleys to help our muscles, these systems are a lever for our minds. That can be both good and bad.

Returning to the creative industry, I recommend that every creative person, whether artist or copywriter, learn these techniques as soon as possible. These AIs will be available in the industry in a matter of months. We're not talking about how copywriting will change in five years. My company, DIVISIO, is currently developing a system called Sokratext that handles copywriting, and that system is already operational. A copywriter will have a competitive advantage if they can master a system like GPT and use it better than the competition. The same is true for designers who employ Stable Diffusion or something similar. This is most likely how it will work in the future.

So, if we begin to entrust these systems with social decisions like whether an applicant is the best fit for a job opening, evaluating legal decisions, or even autonomous driving too soon, we will cause significant damage.

devmio: Sometimes you get the impression that more research results are published on the topic of AI, ML, or DL than actual applications. Is there a research surplus?

Christoph Henkelmann: That's not just an impression; it's definitely the case. I'm also having trouble keeping up at the moment. I always make sure to actually read the most important papers. The time between scientific publications and commercial applications is also getting shorter and shorter. However, it is unfortunate and dangerous that true breakthroughs are always made by large corporations such as Google, Meta, or OpenAI. Many important papers have partial collaborations with university researchers, but just as many papers come purely from private companies. Google, in particular, writes incredibly good papers that demonstrate a focus on scientific quality. However, you must exercise caution so that this does not go out of hand.

There are always very good ideas from other research groups, little diamonds in the rough off the beaten path that don't look impressive at first but have a lot of potential. There are still undiscovered pearls in academia that have not yet made it into production and have not received the attention they deserve.

To be fair, I should also say that an incredible amount of knowledge is freely available. As a small German company, I have full access to Google's research results and can freely download many Google-trained models from Huggingface. That's relatively fair. The problem in Germany is that we're sitting on our hands, hoping that this stupid Internet and computer thing will go away and we'll be able to use our fax machines again. Maybe we'll have to get off our asses at some point.

Fortunately, there is some movement in this direction. For example, a new EU law now allows us, as a German company, to train my models using data crawled from the Internet. Previously, this was always a legal grey area. Things are finally moving forward, but at a relatively slow rate.

The technology is impressive, and it will undoubtedly alter the landscape. However, it will not overthrow society or render programmers obsolete.

devmio: What are you looking forward to in 2023? What do you wish for?

Christoph Henkelmann: I can't really anticipate anything concrete because all major new AI results are always surprises — so I have no idea what's coming. However, I'm not looking forward to GPT-4. I don't care unless it surprises me and crosses a magical threshold into true AI, as OpenAI claims. I don't believe so, but I'm not opposed to being surprised.

Above all, I'd like to see basic research into algorithms to improve model efficiency. I'd like to see us move away from simply making existing technology bigger and bigger, and instead see advancements in areas such as transformer efficiency, for instance. This would also give smaller market players an advantage.

I also hope to see advancements in the field of combining deep learning and symbolic AI. Symbolic AIs include old chess computers, knowledge bases, ontologies, and so on. There is still research in this area, but major breakthroughs are rare. Combining these technologies with DL would be a significant step forward in terms of practical applicability.

Personally, I am hoping that the new language model that we plan to train at DIVISIO next year is as good as I believe it will be. I'd also like to work on a technical problem. I want to make it possible for letters to be used as input for these text-based systems in an efficient manner. This is because the only input these systems currently understand is word snippets rather than letters. Misspellings are not properly handled with this method and can greatly confuse the system. Systems based on letters would be better at adapting to new problems and words on the fly. Vocabulary would be expanded more easily during training and would work directly at the letter level rather than at the word fragment level.

devmio: Thank you very much for taking the time for this interview. I’d like to leave space for a short closing word.

Christoph Henkelmann: The technology is impressive, and it will undoubtedly alter the landscape. However, it will not overthrow society or render programmers obsolete. It's also not completely useless, "stupid," or simply "some statistics." This technology is a fantastic new tool, and we must learn how to use it as professionals and as a society. It is possible to look forward to new and exciting things at every stage of development while maintaining a healthy scepticism about the risks.

While this year has been impressive in terms of AI development, I believe the next one will be even better. The pace of innovation and publications will remain high for quite some time. I'm afraid of that as well, because I have to read all of the research findings. So it could move a little slower for me.