All the hype around AI is due to the surprise that computers have started talking to us. I’ll try to summarize for beginners what the magic of new tools like ChatGPT is.

When Did the Change Happen?

AI development teams have been working on AI for decades. A breakthrough in the development of text processing tools was made by computer scientist Ashish Vaswani, who, along with colleagues on the Google Brain team, published the study “Attention is All You Need” in 2017. In it, they introduced a new “Transformer Architecture” – which enables advanced natural language processing.

The quality of natural language processing tools depends on the quality and scope of the source data from which the language model is learned, as well as the learning method used. The newly introduced method has just radically accelerated the learning process of textual data. What previously took weeks of computational power has been reduced to days by developers. The team accomplished this by better analyzing the relationships between words and phrases in queries and answers – in other words, the new way to analyze text is not to brute force the entire text, but to understand its meaning in the context of the query.

Will We Depend on a Single AI Service Provider? 

We won’t be. Development teams are racing to see who can train their model better, with more data from different sources. In particular, access to sources of information is key to the quality of a given model; since the methods are freely available, anyone can use them to train their model. Even in Atollon, we made our first model, which serves as the basis for the system’s conversation with the user.

The key models that handle natural language processing tasks, built on top of the new Transformer architecture, are GPT and BERT. Meanwhile, GPT is the basis of the currently most popular service from the OpenAI consortium – ChatGPT.

OpenAI’s GPT is powerful in that it has been trained with a truly extensive library of documents, which include web pages, blogs, user forums, books, wikis, scientific publications, and other resources. Other resources may include Twitter conversations that have become the subject of a lawsuit between Elon Musk (currently the owner of Twitter) and the OpenAI consortium (which, ironically, Musk co-founded as an investor and is now working on its models, e.g., TruthGPT). The majority investor in OpenAI is currently Microsoft.

The most popular alternative to GPT is the BERT model. This model is mainly trained with data from Wikipedia and the Google document library.

How does AI generate responses and reactions?

GPT and BERT models generate responses and reactions by predicting the text that follows the input sentence or question. Basic text models can generate contextually relevant text based on the given input. More advanced text models, which are enhanced by understanding instructions, can better predict the user’s intent, allowing them to generate more meaningful responses. To achieve this, text models must be trained to understand language, sentences, and their meanings. In cases where the quality depends on human judgment, human trainers help improve the model.

The quality of the output provided by AI is directly dependent on the quality of the input query. To obtain the best possible output, it’s advisable to learn how to interact with AI and provide explicit instructions. Ideally, be as clear as possible in the first sentence about your specific request, and then elaborate on the details and context in which the question or command occurs.

AI query optimization is a challenge for both application developers and users who want to extract the most value from AI.

What Are the Pitfalls of Using These Models?

• Bias – when training the models, the quality of the output matches what the creators of the model wanted to achieve. It makes a difference whether they fed it with children’s stories or scientific literature.
• Hallucination – it is not the main goal of text models to tell people the truth. The degree of hallucination of the model (i.e., a kind of machine creativity) is, for example, configurable in GPT using the “Temperature” parameter – from the degree of conservatism (Temperature = 1), which is suitable when dealing with facts, to the degree of total creativity (Temperature = 0), which is suitable when dealing with creative activities, e.g., in marketing or writing engaging texts.

How Big a Change Does the Advent of AI Represent for Humanity? 

Quite fundamental. Text models alone are not something anyone should worry about. The concern of AI experts and entrepreneurs (Elon Musk among them) who signed a call to suspend training of new models on 22 March 2023 is that society is unprepared for the changes that lie ahead and, in particular, the way they will be used by humans.

Humans have a historically proven self-defeating ability to turn a good idea into a weapon. For example, social media can be a modern weapon, as has been confirmed in the Brexit decision in the UK or the various elections of semi-totalitarian leaders. Free elections and democracy are the first in line to be threatened by the easy manipulation of eligible voters.

However, AI will most certainly contribute to the further flourishing of the global economy and the individual flourishing of people and companies who know how to take advantage of it. The concentration of the knowledge of all humanity, which spreads unrestricted to millions of households at a fraction of the cost and time of a student studying at MIT or Stanford, will certainly accelerate the development of many disciplines. From research to application in people’s daily lives.

How can I use AI tools today? 

Various solutions can be implemented and used to automate routine activities and potentially replace some previously outsourced professions:

• Customer support
• Content creation
• Market research
• Automation of routine tasks
• Decision-making processes 

Utilizing AI now will enhance your ability to reach targeted customers while reducing the cost of creating marketing campaigns or serving customers.