There is an observation about human society that I have found particularly interesting. Looking at how civilization has evolved, one direction that seems consistently identifiable is our increasing ability to understand the universe around us. I do not mean to suggest that this is necessarily the ultimate purpose of human society, or that every action we take is consciously motivated by the pursuit of knowledge. Rather, I am making an observation about the collective trajectory of society. Even activities that seem to have entirely different motivations can, through various routes, contribute to the development of knowledge.
For instance, technological developments motivated by military or economic interests can eventually lead to advances in our understanding of physics, mathematics, and engineering. Even when some technology remains confidential, its existence may demonstrate what is possible and motivate others to develop similar or better technologies. Of course, this is only one example, and there may well be more fundamental explanations for the direction in which society evolves. My observation is simply that the accumulation and refinement of knowledge is one persistent direction that we can identify.
Now, an interesting consequence of this observation is that much of the technological development throughout human history can also be understood in terms of our attempts to communicate and preserve this knowledge.
Human thoughts are complex, and expressing them to others is not always straightforward. Even in ordinary conversation, we frequently take detours, revisit earlier statements, and struggle to find the right words for what we have in mind. Something that appears perfectly coherent to us might not be equally coherent to someone listening. And this is understandable, because the structure of our thoughts is not necessarily linear, whereas language, particularly spoken language, largely forces us to express them sequentially.
Writing was an important development in this regard. It allowed us to preserve information outside our own memory, communicate ideas across generations, and revisit previously expressed thoughts. But writing also comes with certain limitations. It is relatively slow, and deciding how to organize a complex thought into a sequence of sentences and paragraphs can itself be difficult. One can often imagine several different ways of writing the same idea, each emphasizing different aspects of it. The problem, then, is not merely to express a thought, but to find a representation that communicates it as faithfully as possible.
Computers brought another important development. Information could now be stored, searched, edited, reorganized, and reproduced with considerably greater ease. Unlike writing on paper, digital representations could be modified without having to reconstruct everything from the beginning. The development of the internet further transformed how information could be shared, making communication across geographical and institutional boundaries significantly easier.
I think the importance of these developments becomes particularly apparent when we consider scientific research.
For much of history, scientific communication itself presented a substantial obstacle to research. Consider, for instance, the independent development of calculus by Newton and Leibniz. Although scientific correspondence and publication networks certainly existed, the circulation of new ideas was much more restricted than it is today. Researchers working in different geographical regions might remain unaware of developments elsewhere, and even published results were not necessarily accessible to everyone who could benefit from them.
The development of scientific journals, libraries, and other institutions gradually improved this situation. But even with access to a well-stocked library, a researcher faced another difficulty: deciding which books and publications were actually relevant.
This might seem like a relatively minor problem, but I think it is worth considering how much effort it can require. Before deciding whether a book contains something useful, one might have to read a considerable portion of it. And even after identifying the right books, understanding them and relating their contents to other existing work remains a time-consuming process.
Computers and the internet greatly reduced some of these difficulties. Scientific publications became easier to discover and distribute, and researchers could access information that would previously have required considerable effort to obtain.
However, something interesting seems to have happened alongside these developments.
As scientific knowledge accumulated, scientific fields themselves became increasingly complex and specialized. The amount of existing knowledge that a researcher might need to understand before making a meaningful contribution also increased. This is particularly apparent in fields such as theoretical physics and mathematics, where contemporary research often builds on several layers of earlier developments.
Consequently, although accessing scientific information has become substantially easier, understanding enough of it to contribute to a particular research area can remain extraordinarily difficult.
This also helps explain why experienced researchers and supervisors are such important sources of information, particularly for students entering a field. A senior researcher may have spent years reading the relevant literature and understanding how different developments relate to one another. A younger researcher can benefit enormously from that accumulated knowledge, rather than having to reconstruct the same understanding independently.
In some sense, then, we encounter an interesting difficulty: the accumulation of scientific knowledge, which is itself central to scientific progress, can also make further scientific progress increasingly difficult for individuals attempting to enter a field.
This phenomenon is closely related to what is sometimes called the burden of knowledge: as our collective understanding expands, individuals must acquire an increasingly substantial body of existing knowledge before they can contribute something new.
And this is where I find the emergence of large language models particularly interesting.
We generally discuss LLMs in terms of artificial intelligence, reasoning capabilities, and their ability to perform tasks that previously required human intelligence. But I think there is another perspective through which we can understand their development.
LLMs can be viewed as an epistemic response to the burden of knowledge.
I do not mean that they were necessarily developed with this particular objective in mind, or that this interpretation captures everything about them. Rather, I find it useful to situate their emergence within the historical progression of technologies that help us acquire, express, communicate, and organize knowledge. As the body of existing knowledge grows, so does the difficulty of navigating and understanding it. Language models seem to address, at least partially, precisely this difficulty.
Consider something as simple as expressing one's own thoughts. Recently, I have started experimenting with dictating my thoughts in their raw form, without worrying too much about their organization or whether every sentence follows coherently from the previous one. The thoughts are often scattered, and I might jump between different aspects of the same argument before returning to something I mentioned earlier.
A language model can then take this material and organize it into a coherent explanation.
What I find significant about this is that I no longer have to devote as much attention to deciding how my thoughts should be expressed while I am still trying to articulate them. I can concentrate on developing and communicating the underlying ideas, while the model helps construct a suitable linguistic representation.
Of course, there is no guarantee that the resulting explanation will perfectly preserve the original thought. The model might misunderstand certain relationships or introduce connections that I did not intend. But the broader capability is nevertheless quite significant.
And I think something similar is happening in scientific research.
Suppose a researcher is trying to understand a particular topic and discovers several dozen potentially relevant papers. Search engines and electronic databases have made finding these papers relatively easy. But the researcher must still understand their contents, identify the relationships between them, and construct a coherent picture of the existing literature.
Language models can assist with precisely this process. Instead of approaching every publication independently, one can use them to help identify connections, explain unfamiliar concepts, compare different approaches, and organize the information into a more understandable form. Naturally, the resulting synthesis must still be checked against the actual literature, particularly when making scientific claims.
The distinction is that earlier technologies primarily improved our ability to preserve, distribute, and retrieve information, whereas language models can also assist in transforming that information into coherent explanations.
I think this is particularly consequential for younger researchers. The growing complexity of scientific fields creates substantial barriers to entry, and a considerable part of the difficulty involves acquiring and organizing existing knowledge. Language models can help reduce some of this burden, making it easier for individuals to explore unfamiliar subjects and begin developing their own understanding of them.
Viewed this way, LLMs seem much less like an isolated invention and more like a natural development along a trajectory that human society has been following for a very long time.
We have continually developed tools that allow us to preserve more information, communicate it more effectively, and make increasingly complex knowledge accessible to others. Each such development has changed how we interact with existing knowledge and, in many cases, what we are capable of discovering.
There is also an interesting relationship between these developments and the growth of knowledge itself. As our collective understanding expands, the amount of information that needs to be communicated and organized also increases. This creates new difficulties, which in turn motivate the development of better tools for handling them. These tools can then facilitate further scientific progress, producing even more knowledge.
Language models can be understood as part of this continuing process.
I also suspect that the progression will not end with LLMs. There may eventually be technologies that allow us to express and communicate complex thoughts much more directly, without requiring us to organize everything into linear sequences of natural language.
After all, a single idea can often be expressed in several different ways, and each representation may capture something that the others do not. It is conceivable that future technologies will allow us to work with richer representations of thoughts and their relationships, rather than relying predominantly on sentences and paragraphs.
I do not know what such technologies would look like, and I do not think we need to speculate too precisely about their eventual form.
But the broader direction seems interesting to me.
Perhaps the significance of language models is not merely that we have developed machines capable of producing human-like language. Perhaps they are also a development in a much longer history of human attempts to overcome the limitations of how we express, communicate, and collectively understand our thoughts.
And from this perspective, the emergence of LLMs seems less surprising. It becomes intelligible as a continuation of something that has been happening throughout the development of human civilization.