on pseudoism in the era of llms
large language models have become dramatically more intelligent over the past few years. this progress has come from several breakthroughs: reinforcement learning, scaling laws, improved post-training techniques, and better architectural decisions.
while the improvement of large language models is remarkable, there is a broader consequence that deserves attention. much of the intellectual struggle that helps human beings become more thoughtful is beginning to disappear.
this is happening because we are increasingly delegating tasks that once required reading, writing, and sustained thinking to language models. in return, the clarity of thought that naturally develops through those activities is slowly being lost.
by "being lost," i mean that the models are increasingly doing the difficult cognitive work that humans themselves should experience. it is often through this struggle that better reasoning emerges. wrestling with difficult ideas, writing imperfect drafts, and reading carefully all contribute to the development of deeper cognition. without that process, we risk replacing meaningful intellectual growth with convenience.
the rise of llms cannot be reversed. the economic incentives are too strong, and they are already reshaping the way knowledge work is done. the delegation of core cognitive skills is steadily shifting toward large language models.
using llms as helpers that enable critical thinking—not as replacements for it—should be the goal.
if you encounter a problem whose solution is relatively straightforward, and an llm could solve it in a single prompt, it is still worth spending 20 to 30 minutes attempting it yourself before asking the model. the objective is not efficiency at every moment, but the cultivation of the ability to think.
in that sense, we should continue exercising our minds. we should improve our reasoning while also avoiding unnecessary burnout. patience becomes an essential part of this process.
reading and writing may gradually become scarcer habits. in an age dominated by llms, those who continue to practice them deeply may be viewed as unusual by some, or even outdated by others. yet these activities remain among the most effective ways of developing genuine understanding.
the central idea here is emergence.
a person who reads many books slowly, carefully, and with genuine understanding accumulates far more than isolated facts. over time, this long chain of learning creates emergence—a moment where understanding suddenly expands, where seemingly disconnected ideas become connected, and higher levels of insight appear almost unexpectedly.
it is crucial to think deeply, reason carefully, and develop opinions of your own. not opinions in the sense of political identity, but opinions about history, mathematics, physics, science, philosophy, and the world in general.
when an llm generates an answer, the opinions it presents inevitably depend, to some degree, on the biases, assumptions, and training distribution of the model. if people accept those outputs without independent reasoning, we risk creating a generation of pseudo-intellectuals—people who possess fluent answers without genuine understanding.
that could mark the decline of deep thinking and the rise of intellectual mindlessness.
at the same time, the technology is genuinely extraordinary. the concern is not with the capability of the models themselves, but with human nature. convenience often wins over originality, and delegation often replaces development.
there will undoubtedly be tools designed to preserve and strengthen human thinking alongside ai. even so, there remains something fundamentally different about arriving at understanding through your own effort. some experiences are simply best lived directly rather than outsourced.
perhaps ai can eventually become a better intellectual partner—one that disagrees, challenges assumptions, and forces deeper reasoning instead of merely producing agreeable answers. however, that runs against the primary objective of current alignment, which generally aims to make models helpful, cooperative, and safe rather than intellectually adversarial.
the future should not be about humans thinking less because ai thinks more. it should be about ai making human thinking deeper, not replacing the very process through which intelligence emerges.