The artificial intelligence models we know today—such as GPT, Claude, Gemini, Grok, Llama, and Qwen—represent the “brains” behind modern AI applications. They are the systems that understand language, generate text, write code, analyze images, and solve an impressive range of tasks. The applications users interact with every day, such as ChatGPT, Claude, or Gemini, are merely the interfaces through which people communicate with these models. Understanding how these artificial “brains” are built is the first step toward understanding how modern artificial intelligence actually works.

Why Not Google?!

A decisive moment in the history of artificial intelligence came in 2017, with the publication of the scientific paper Attention Is All You Need, in which Google researchers introduced the Transformer architecture. It fundamentally changed the way models process language. Unlike previous architectures, the Transformer can analyze all the words in a text simultaneously and identify the relationships between them, even when they are separated by hundreds or thousands of other words. This approach made it possible to build models that were much larger, faster, and more capable than anything that had existed before, becoming the foundation of almost all modern large language models.

Building on this architecture, OpenAI launched GPT-1 (Generative Pre-trained Transformer 1) in 2018, the first model in the GPT family and the direct precursor to ChatGPT. Although GPT-1 was modest by today’s standards and remained known primarily within the research community, it demonstrated that a model pre-trained on enormous amounts of text could acquire general knowledge about language and be successfully adapted to a wide range of different tasks.

Although Google was perhaps the world’s most advanced company in artificial intelligence, had developed the Transformer architecture, and possessed exceptional technical resources, the company also faced extremely high expectations regarding the quality and reliability of every new product it launched. Integrating a conversational system capable of producing incorrect or unpredictable answers into its portfolio risked undermining the trust built around mature services such as Google Search, Gmail, YouTube, and Google Maps. OpenAI, being a much younger company with no established ecosystem to protect, could take such risks more easily. It chose to release the technology publicly at a moment when it had become useful enough, even if it was not yet perfect, turning ChatGPT into the first conversational AI system to achieve global-scale adoption.

Within just a few months of its launch, OpenAI attracted tens of millions of users who used ChatGPT in real-world situations, providing the company with an unprecedented volume of feedback about model errors, desired features, and user needs. This gave OpenAI an important technical advantage, allowing it to rapidly improve subsequent versions and develop new capabilities. Financially, ChatGPT’s success generated substantial revenue and attracted massive investment, enabling OpenAI to remain in the race to develop frontier models alongside the industry’s giants. Those companies, in turn, invested enormous sums to close the gap, meaning that OpenAI’s early advantage proved highly significant—but not impossible to overcome.

OpenAI & Co.

At the end of 2020, eight future co-founders of Anthropic, reportedly joined by several other colleagues, left OpenAI around Dario Amodei, the company’s former vice president of research and one of the people involved in the development of GPT-2 and GPT-3. Among those who left were researchers and leaders who had an insider’s understanding of both OpenAI’s technology and strategy. The split was not a mass resignation, but rather the departure of an elite team that effectively crossed the street and built its own company. The stated reason was a difference in vision over how increasingly powerful models should be developed and controlled. A few years later, Claude had become a direct competitor to the GPT models created by the company from which Anthropic’s founders had departed.

Elon Musk and Sam Altman were colleagues and co-founders of OpenAI. In 2018, Musk left the organization, and in the years that followed, the relationship between the two visibly deteriorated, turning into one of the most closely watched rivalries in the artificial intelligence industry. The two have repeatedly criticized each other publicly and have even become involved in legal disputes over OpenAI’s direction and mission. In 2023, Musk founded xAI, the company behind the Grok family of models, which competes directly with GPT.

AI engineers and researchers capable of building frontier models have become something of a new class of technology-industry stars. At major American labs, total annual compensation can frequently reach several hundred thousand dollars, while packages for elite specialists can exceed $1 million through a combination of salary, bonuses, and equity. In exceptional cases, the competition has gone much further: Google DeepMind reportedly offered some researchers packages worth as much as $20 million per year, while Sam Altman claimed that Meta attempted to recruit OpenAI employees with signing bonuses of $100 million. Beyond money, these specialists are attracted by access to enormous computing power, research freedom, and the opportunity to work on models that could become the foundation for thousands of products.

MADE IN CHINA

If the Americans impressed with their enormous investments in artificial intelligence development, the Chinese entered the competition with something their industry has been exceptionally good at for decades: drastically reducing costs. DeepSeek demonstrated that a model capable of competing with market leaders could be built and offered at costs far lower than previously thought possible, forcing the entire industry to ask the same question that any healthy economy should ask: can we achieve similar results with less money?

The good news is that this competition does not benefit companies alone—it benefits users as well. As models become more efficient and cheaper, artificial intelligence can more easily make its way into schools, hospitals, small businesses, and millions of personal computers. The American response was not to ignore the challenge, but to accelerate the race toward more efficient models, lowering prices and investing heavily in optimizing their own technologies.

EUROPE, THE WELL-REGULATED ONE

Europe has chosen a different strategy from the United States and China. Instead of entering the race for massive investments in the development of the most powerful AI models, the European Union has focused on establishing clear rules regarding transparency, safety, and responsible use.

At this stage of the market’s development, this approach can represent a disadvantage for European startups, because every additional regulatory requirement means costs, time, and resources that early-stage companies feel most acutely. In the medium and long term, however, this strategy could turn into a competitive advantage. As AI models become increasingly powerful and widespread, user trust, predictability, and compliance with high safety standards may become just as important as technical performance.