Between Artificial General Intelligence and Artificial Superintelligence lie many questions and speculations.
For more clarity, I wanted to write a short contribution in simple, explanatory words.
Artificial General Intelligence, scientifically called AGI, is defined and characterized mainly by its cognitive ability to adapt and to learn every intellectual domain without necessarily having been trained on the data related to those domains — it is, in reality, a rational simulation of the human brain’s adaptability when facing a new domain.
Thus, to reach this technology, we will not only need to expand the scale of data centers and servers, but also, and even more importantly, “invent or develop a new algorithm” beyond the Transformer architecture (LLMs, the technology behind services like ChatGPT), because research has shown that over time these recent large models tend to reach a plateau — a cognitive inability to go beyond the statistical patterns that drive their functioning. They become “experts,” yet remain surprisingly weak at learning the more basic things the human brain can easily acquire. [1]
Artificial Superintelligence (ASI), just like Artificial General Intelligence — which already presents us with a significant ceiling of difficulties to overcome before we can even speak of “intelligence” — is far more than a definition known by the majority. It is a research subject, a challenge, a frontier, a hypothesis.
(My opinion) What is very interesting is the fact that if we keep our eyes fixed on this theory—hoping to prove it scientifically and computationally—we increase our chances of achieving AGI. Because to even begin talking about superintelligence, we must first be theoretically capable of matching the core abilities of the human brain.
In simplified terms,
ASI = f(AGI) + λ
(Explanation) An ASI is the result of a function of “realized AGI,” normalized, and augmented (hence the lambda λ, a standardized representation of the new algorithmic dimension we would need to make this dream possible). This formula is written as an explanatory and synthetic expression of the thesis I support here, and the broader theory is indeed being discussed in current artificial intelligence research circles. [2]
References
———
[1] Buchanan, Scott. “Meta AI Chief Yann LeCun Notes Limits of Large Language Models and Path Towards Artificial General Intelligence.” Economist Writing Every Day, 22 July 2025.
[2] Alvarez-Teleña, S., & Díez-Fernández, M. (2024, September 15). Advances in Artificial Super Intelligence: Calm is All You Need [Working paper]. SSRN. https://doi.org/10.2139/ssrn.4924496
0 Comment