AI can be closer to original scientific research than we think
Science has always been a human business, fueled by curiosity, creativity and an obstinate desire to wonder what others hold for granted. But what happens when artificial intelligence begins to do the same-not just to help human scientists, but to design independently of experiences, to analyze data and to train conclusions?
This question has become more than theoretical recently, when an AI system of Sakana IA of Japan has generated a hypothesis, designed experiences and wrote a revolutionary scientific article by peers on its conclusions, all without human intervention.
Title Regualization of the composition: unexpected obstacles to improve the generalization of the neural network,, paper was accepted as a projector paper at ICLR 2025, one of the most prestigious automatic learning gatherings on the ground. Silently, this event marked a threshold: AI was the author of the original research deemed worthy by its human peers.
The rise of the AI scientist
The system, called AI-V2 scientistis not only another language model. It is a fully autonomous research agent designed to automate the entire scientific process. The criticisms, ignoring that the newspaper was written AI, obtained it high enough for acceptance, placing it over almost half of all submissions by humans.
The implications are deep: a machine has not only included a field of research, but has formulated questions, carried out experiences, written of the code, analyzed data and clearly expressed its results.
The promise – and the problem
At first glance, the realization suggests that we can head towards an “intelligence explosion” – the point where AI helps not only with science, but motivates it, adding to human knowledge faster than humans cannot themselves. Some, like the former researcher Openai Léopold AschenbrennerBelieve that this tilting point could happen in 2027.
But not everyone agrees.
Yann LecunThe chief scientist of Meta AI and a winner of a Turing Prize, has long warned of confusing the tuning of models with real intelligence. Current AI models cannot form the type of mental models that underlie real reasoning or original discovery, he says.
In other words, the AI scientist-V2 may have “written” a research document, but if he Understood What he did – or simply sewn models of his training – is always an open question.
Sakana’s prudent breakthrough
On their credit, Sakana .i has been dealing with this experience as exactly that: experience. The company withdrew the document before the conference, recognizing the ethical gray area it occupies.
However, as AI systems become more capable, they will play more and more roles in scientific discovery. They already amplify the process, accelerate literature reviews, accelerate the development of the code and generate experimentation conceptions in a few minutes instead of months.
It is, according to Lecun, the future in the short term: AI as a powerful tool, not an autonomous genius.
Beyond imitation: towards original thought
So Regularization of the composition Search really original? In a narrow sense – yes. The document introduced a new experimental configuration, studied a new angle of generalization and was deemed worthy of presentation. However, its results were increasing, showing that its hypothesis failed. And in the broader philosophical sense, originality in science is not only a question of novelty; It is about intuition, asking questions and an ability to see beyond the data.
Lecun compares this to the difference between solving a mathematical problem and inventing a new branch of mathematics. The latter needs a system to understand the world, make predictions according to experience and plan actions based on abstract objectives. These capacities, he maintains, are always out of reach for AI.
However, the fact that a machine can imitate form of scientific thinking that this well is not trivial. It increases the bar for it. In the years to come, AI will probably generate hypotheses, automate laboratory work and perhaps a research day worthy of a Nobel co-author. But the prudence of Lecun reminds us: paternity does not imply understanding, and prediction is not the same thing as understanding.
The Road Ahead: Collaborative Intelligence
The foreground can reside in hybrid intelligence – where AI systems manage complexity and scale, while humans provide an overview, ethics and conceptual jumps. The stage of Sakana.ai is not the destination, but a waypoint on a longer journey towards the reshaping of the way in which we are doing science.
Sakana AI’s experiences rekindle the long -standing discussion if AI will soon optimize its own architecture, refine its own reasoning capacities and accelerate the discovery rate in a way that we cannot yet predict. The success of the AI scientist does not mean that we are at this inflection point, but that suggests that we can be closer than many people do not think so.
