ChatGPT and similar tools have caused a lot of furor in recent months. Not least in content creation, they can be a tremendous help. At the same time, you should be aware of the limitations of these tools and understand how to achieve particularly good results.

In this article, I'll explain why the new generation of AI offerings is so much better. It is the start of a series on the topic. The topic of "artificial intelligence" has been experiencing periods of exuberance and dejection for decades now.

Elon Musk's predictions in the area of autonomous vehicles are well known here: The field made rapid progress for a while. Computer-controlled cabs and buses seemed within reach. However, the curve of improvements soon flattened out, and in the meantime, forecasts in this area have become much more cautious. In this respect, I have become accustomed to a healthy skepticism when it comes to such hype topics. 

That, however, quickly flew out the window with ChatGPT. When I was able to try it out for the first time, I was as amazed as I was thrilled: this "chatbot" finally worked the way its many predecessors had only promised it would. It was almost uncanny.

ChatGPT responds to questions in an amazingly human way. And it seems to know an answer to every question - or several. It dynamically adjusts the type, length, and complexity of the output based on my input. It understands contexts of the conversation and can draw on topics and facts that have been raised before. It processes long and complex inputs with little delay. And it also understands and responds in German.

Amazingly powerful text generator

It quickly became clear that ChatGPT is not only a powerful AI assistant, but also an AI text generator.

My previous attempts with tools of this type had always been big disappointments. The products could never keep up with the marketing promises. They were perhaps suitable for stimulating ideas. One could extract set pieces from some of them. But the texts were rarely useful as a whole.

It's different now with ChatGPT and its variants and competitors: Used correctly, they can provide not only ideas, but also a comprehensive concept and at least a good first draft.

As someone who spends much of his livelihood writing and editing, I hate to say this: what ChatGPT delivers is often on the same level I've seen from human writers. Sometimes it's even better.

However, the limitations and special quirks of these new AI tools then also become clear after intensive testing. More on this below.

How did this progress become possible?

But how was this sudden leap in quality possible in the first place? Three points are usually mentioned:

1. Training data: Today's AI models learn from existing content (text, images, audio, code, etc.). The amount of data already available digitally has increased rapidly, which helps training enormously.
2. Computational power: Specialized computers and components have greatly accelerated training operations, enabling larger and more complex models.
3. algorithms: Last but not least, there have been important advances under the hood. The "transformer" method, for example, is considered a major reason why AI can understand and generate texts so much better today than it could a few years ago.

It is also impressive to look at the number of "parameters" of language models in recent years. More parameters allow the model to encode more knowledge and handle more complex tasks:

• 2019, GPT-2: 1.5 billion parameters
• 2020, GPT-3: 175 billion parameters
• 2022, PaLM: 540 billion parameters
• 2022, GPT-4: around 1 trillion parameters
• 2022. Claude: around 10 trillion parameters

These numbers are fascinating, but more complex models are not always automatically superior or the better choice. One current trend, for example, is to train models more specifically for certain tasks and to adjust them appropriately. It is therefore to be expected that, in addition to general tools, there will be more and more offerings that are intended for a clearly defined purpose. 

Models that are additionally trained with individual data are also exciting in this context: For example, companies can feed all their documents into such a system to get answers dynamically.

Another interesting metric that has recently come more into the spotlight is context length. The larger this value, the more content of a conversation the tool can include. More context thus helps the AI to conduct longer chats, but also to process larger inputs. 

Depending on the use case, this can make a significant difference when an offering like Anthropics Claude, for example, processes and analyzes entire books within seconds.

Context length is measured in "tokens", where a token is roughly equivalent to one word. Some examples:

• GPT-2: 1,024 tokens
• GPT-3: 2,048 tokens (in a new version up to 16,000 tokens)
• PaLM: 65,536 tokens
• GPT-4: up to 32,000 tokens
• Claude: probably around 100,000 tokens

A longer context requires correspondingly more computing power and memory. It is therefore a technical challenge to further increase these values.

Three options to use such tools

Anyone who wants to use such tools currently has three main options:

1. In the cloud. ChatGPT, Claude, but also image generators such as MidJourney or Stable Diffusion can be used as software-as-a-service. This means that the user's own data is processed on the providers' servers. Depending on the type of information, this can be quite problematic. At the same time, as a user, you have to be satisfied with the interface and the options of the offer. Companies like OpenAI, Microsoft, Google, and Anthropic have specialized, high-performance servers for this purpose. 
2. Via an API. OpenAI in particular actively offers its interfaces. Not all AI models are immediately available to everyone. Nevertheless, they can either be used to implement one's own applications or to use third-party apps. Data processing here still takes place on the servers of the AI companies. Where and how the offering can be used, however, is customizable in this case.
3. On your own computer or server. Not only specialized computers have become more powerful, but also commonly available laptops, tablets and even smartphones. With modern and appropriately equipped devices, this can be enough to use tools such as AI assistants directly on one's own computer. They're not as powerful as the high-end applications in the cloud. But that is not always necessary. Instead, the user's own data remains on the computer. In addition, the software and the model can be selected entirely according to one's own needs. One example is LM Studio for Windows and Macs, which allows you to use language models such as Meta's Llama family on your own PC.

In addition, there is currently another trend that I believe will become even more prevalent: AI assistants that are integrated into other offerings. Examples are "Copilot" in Microsoft 365, Adobe's "Firefly", Bing Chat or Google's experimental, AI-supported "Search Generative Experience" (SGE).

Limitations of AI tools

In further articles in this series, I will show you in more detail how I personally use such services to research topics, generate ideas and concepts, and create texts and images. 

For all the enthusiasm for the opportunities and possibilities that these new little helpers make possible, they have limitations that you should be aware of, and there are legitimate criticisms.

An offer like ChatGPT, for instance, has learned to give a linguistically correct and meaningful sounding answer. That is the focus. The validity of the facts and figures mentioned, on the other hand, is not. They can be true or they can be fictitious. So you should not take the statements unchecked.

In some tasks, these tools are also completely overwhelmed. For example, they often can't handle numbers and calculations well.

The providers are trying to counteract this. On the one hand, the AI assistants are to be educated to be more honest. If they don't know something exactly, they should make that clear. On the other hand, OpenAI, for example, has added plugins as an option: This allows ChatGPT to access specialized tools and information sources for certain topics and tasks. Bing Chat is another example: It substantiates the sources of its answers with links and makes it clear when it could not find a piece of information.

Moreover, the knowledge of an AI assistant like ChatGPT or Claude often only reaches up to a certain date. Everything that has happened since then is unknown. The training process of such an AI is so elaborate and lengthy that new information cannot simply be added. You have to be aware of this for some topics.

Furthermore, another problem is that an AI can spread and thus reinforce biases and misinformation that it has found in its training data. After all, the AI does not understand what it is doing there. It usually does not check or research either.

Another thing I always miss with AI assistants in everyday life: They don't get to know me and they don't learn from previous conversations. As described above, there is a certain context length per chat. But the context ends with the current chat in any case. If I start a new conversation, the AI assistant doesn't know anything about previous interactions. My hope is that these offerings will become more personalized in the future. SHO.AI promises something like that.

Criticism of the AI tools

A fundamental criticism of tools like ChatGPT for texts or Stable Diffusion for images is the training material. As already described, these data are indispensable for the learning process. However, the authors were often not asked whether they wanted to make their works available for this purpose or not. The fact that AI image generators can imitate artists' styles caused a stir. Is that automated copyright infringement? Or is it comparable to human works, which can also be inspired and influenced by the works of others? These are exciting questions that will be with us for years to come. 

The debate about this is sometimes heated. No wonder: Some artists see themselves as involuntary stapes holders for an AI that could make them superfluous in return. And the companies earn money with a product that has taken their work for free.

OpenAI meanwhile offers an option to at least block the content of its own website for such training purposes in the future.

This also raises the question of whether the results of such tools may be used at all. I talked to attorney Dr. Carsten Ulbricht about this. As is so often the case, the question cannot be answered with a clear yes or no.

Last but not least, the question of whether works enjoy copyright protection if they originate from an AI and who is considered the author in this case is completely open. According to some, the yardstick here is how much work the AI has done and how much the human being has done. 

Conclusion on content and AI

The AI world has been booming and hype over the past few months. As I hope I've been able to show in this post, the enthusiasm is not completely out of thin air. The progress is clearly noticeable. The tools can be used for everyday tasks and can be a great help.

With all that said, they are not perfect, they make mistakes, they react unexpectedly, or they fail at a task completely (and may even deny it). Moreover, there is legitimate criticism of how these tools work and how they acquired their skills.

With these points in mind, in the next part of the series, I'll show how I use various AI tools for creativity and productivity.