People are increasingly turning to the powerful capabilities of current large language models (LLMs) for AI applications in science, driven by their improving accuracy and methodologies. But what exactly is the intelligence?

Even without formal theoretical analysis, I believe true intelligence would be characterized by performing tasks like humans — continuous self-learning and improvement in any environment. It is similar to human beings, trained on general tasks that already contain fundamental reasoning and memory. It can also perform well in specific tasks in professional disciplines with a little adaption. That’s why the undeniable fact is that for a long time, computer scientists have prioritised general tasks over the diverse abilities that expect an agent to outperform in specific or professional scenarios as well.

AI for Science is not a novel area yet more challenging in perspective. Compared to those 99% abilities to support an AI model well-act in general scenarios, that 1%, importantly, for some reason may lead to suffering results in the end, and it’s often the 1% of tasks that lead to significant challenges. These times, people are saying they will solve protein folding or materials designing problems, as they believe the frontiers for sure with 1% are in front of the eye. Notably, there is also an obvious way to let AI behave like humans with specific abilities that are always across disciplines — anybody can tune the model and then directly run queries with prompts upon LLMs, or at a higher level we design some algorithms to let it avoid errors. It is promising that we are all in LLMs for pursuing an AGI.

The crux of the issue lies in understanding how these models can mimic the human ability to comprehend and interact with the physical world. At the point of comprehension, it is physicochemical mechanisms that define the world, while at the interaction level, rationality shows us perfect cognition.

However, it remains a blank area for people to understand the world from a pure physicochemical mechanisms level, instead, they stand on the statistical results and act as a king of Arther. You probably agree that LLMs excel at processing and generating language, but they lack the inherent understanding of the physical world that humans possess. This gap is significant when it comes to scientific applications, where the interplay of physical laws and chemical processes is crucial. For instance, predicting molecular interactions or understanding the behaviour of complex systems requires more than just linguistic proficiency; it demands a deep integration of scientific knowledge and reasoning.

To bridge this gap, it’s essential to develop models that can not only process language but also simulate and reason about physicochemical phenomena. This requires a multidisciplinary approach, combining advances in AI with insights from physics, chemistry, and biology. By embedding these principles into the training and architecture of LLMs, we can enhance their ability to perform scientific tasks more effectively.

Moreover, the integration of sensorimotor capabilities with LLMs can lead to more holistic AI systems. By enabling agents to interact with their environment and gather empirical data, we can foster a more profound understanding of physical processes. This approach mirrors the way humans learn and adapt, through continuous interaction and feedback from the world around them. But anyway, it is the well-defined physicochemical principles that promote the basic understanding of the world, something like causal, or facts with rule-like structures.

Stand on the frontier of current AI, I believe we may need more progress towards the essence of AI for Science — that is, leveraging those fundamental principles in the real world to make AI powerful, and assessing their abilities of comprehension in professional knowledge.

To go back to the title, I think we even don’t know what the physicochemical mechanisms, simple principles, or great laws are, in the context of AI for Science. But one more unassured guess, if there is one more try to leverage these most advanced technologies, one may realize that the true mechanisms of the world are not all in the sampled data — what machines learned from existing data is only the differences among perspectives of interpreting it with randomness.

<hr><p>Are Physicochemical Mechanisms Important to Agent for Science? was originally published in AI Mind on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>