By Gabriel Cunha, Oracle AI expert
Artificial intelligence is undergoing a transition phase. After years in which the advancement of large language models, the so-called LLMs, was driven by increased computational power, data availability, and model growth, the industry is now facing a more complex question: how far these architectures can evolve and whether they will be sufficient to achieve General Artificial Intelligence (AGI).
Large language models are at the center of the transformation that brought generative AI to a new level. Systems like ChatGPT have demonstrated an increasingly greater ability to understand commands, generate text, write code, summarize information, and perform tasks that previously required human intervention. The same evolution has reached images, videos, and audio, significantly expanding the technology's application potential.
But performing complex tasks efficiently does not necessarily mean replicating broader human intelligence capabilities. Current models are capable of processing large volumes of information, identifying patterns, and establishing relationships based on the data used in their training. The challenge for the coming years will be to develop systems capable of learning in a more continuous manner, handling different types of information, understanding broader contexts, and adapting to new situations.
It is precisely at this point that one of the main discussions about the future of AI arises: how far can we go by simply scaling existing models? They have already demonstrated impressive capabilities, but intelligence is not just about predicting the next word or producing a coherent response. Achieving truly general systems requires understanding which capabilities are still missing and which new approaches may be necessary to develop them.
This discussion divides researchers and companies at the forefront of the technology. One of the views holds that there is still a lot of room for evolution of current architectures. Larger models, combined with more data, greater computational power, and increasingly sophisticated training methods could continue to expand system capabilities. From this perspective, artificial intelligence itself may also contribute to accelerating its development, assisting in programming, research, and the creation of more advanced models.
The growth of Anthropic illustrates the scale of this race. According to data published by TechCrunch, the company's annual revenue went from $9 billion at the end of 2025 to over $65 billion in July 2026. Therefore, the competition for the next generation of artificial intelligence is not just technological. It involves billions of investments, infrastructure, product development, and a race for prominence in one of the most strategic markets in the digital economy.
At the same time, researchers like Yann LeCun from AMI Labs question whether language models, even at ever-increasing scales, will be sufficient to represent broader capabilities of human intelligence. The very way people learn shows that intelligence does not depend solely on language. Perception, memory, interaction with the environment, reasoning, experience, and continuous learning are also part of this process.
This perspective opens the door to the search for new architectures capable of combining different modalities of information and building more comprehensive representations of the world. LLMs are not necessarily losing their importance in this scenario. On the contrary, they continue to be fundamental components, but integrated into more complex systems capable of gathering different forms of learning, memory, perception, and reasoning.
The issue is particularly relevant because scaling has been one of the main strategies responsible for recent advances in artificial intelligence. For years, expanding models and providing more data and computational power have produced significant performance gains. Language models may continue to evolve and overcome limitations that today seem structural, while new architectures may be developed to solve problems that are not easily addressed merely by adding more parameters, data, and processing. The two paths can also advance simultaneously and eventually converge.
The next phase of artificial intelligence, therefore, will be marked by a change in the very question that guides the industry. Instead of merely discussing how large and powerful models can become, it will be necessary to understand what still needs to be achieved for them to be able to learn, reason, perceive, and adapt more closely to general intelligence.
In the race for AGI, the real challenge will be to discover whether intelligence can be indefinitely expanded from the architectures we know, or whether the next major leap will depend on a new way of building machines capable of learning and understanding the world.
Gabriel Cunha wrote this article at the invitation of Olhar Digital. His opinions do not necessarily reflect the opinions of OD.
Bruno Capozzi is a journalist, master in Social Sciences, and executive editor of OD.
