Beyond Language Models
Problems in the physical world are complex. Social world problems reach a higher level of complexity.
[perhaps this needs to be a longer essay. The ideas seem clear in my head, but may only be sketched here.]
Think of three realms: the digital realm, the physical realm, and the social realm.
Impressive as they are, LLMs have only shown the ability to master the digital realm. The physical realm, with its many arrangements of particles and millennia of evolution, is more complex. The human social realm, with the rapidity of cultural evolution, has reached an even higher level of complexity.
When we hope for today’s models to cure cancer, we are expecting the masters of the digital realm to solve a problem outside of the digital realm. Cancer is a problem in the physical realm. And it may also have causes in the social realm.
When we hope for artificial intelligence to revolutionize education, we are asking it to solve a problem in the social realm. Those of us who follow the Alpha School by reading the Austin Scholar substack know how much she emphasizes the role of motivation strategies in addition to the AI learning tools.
Cybersecurity appears to be a problem in the digital realm. But it is actually a problem in the social realm. I want convenient access to certain computer files, such as my bank account, because of their relevance to my social world. But by the same token criminals would like to impersonate me in order to have access to such files. The easier it is for me to access a file, the easier it is to impersonate me. As systems have evolved in recent decades, the drive to make it harder to impersonate has made access less convenient. I waste much more time trying to authenticate myself than I did twenty years ago. Yet vulnerability remains. As Thomas Sowell says about problems in the social world, “There are no solutions, only trade-offs.”
Modern philosophers, following Descartes, tend to gloss over the social realm. They want to retreat to the room where he meditated and emerged with “I think, therefore I am.” But “I think” is not the best way to describe how we acquire knowledge. Thinking is social. We decide what to believe by deciding who to believe.
The digital realm is very important. The major LLM labs have achieved a great deal. I am grateful to be able to use their tools. But I take seriously the importance of the physical realm, which is more complex. I also take seriously the importance of the social realm, which is even more complex. So I doubt that genuine artificial intelligence is close at hand.
Although it is quite true that the human social realm is highly complex (and has been for a long time – recently finished the Payton and Mammitzch translation of Huizinga’s The Autumn of the Middle Ages which provides more than ample evidence should anyone doubt), that is not to say that human social institutions are particularly competent at the businesses they are ostensibly about and that absent certain sentimental attachments to the corrupt, random, and arbitrary, AI couldn’t readily replace a majority of humans in such enterprises as the courts, government bureaucracy, and the schools. And indeed each of these are lousy enough that there would be little lost in the unlikely event proved incapable of producing substantially better outcomes.
Dr. Kling mentions the human role in promoting motivation. While the research does demonstrate that certain techniques in the vein of self determination theory such as autonomy promotion can have demonstrably positive effects, these techniques are rarely taught in schools of education and few teachers have the wherewithal to practice them. Indeed observational studies find that the incidence of positive classroom environments in the US is rare. As most people are well aware:
https://www.pewresearch.org/short-reads/2024/04/04/about-half-of-americans-say-public-k-12-education-is-going-in-the-wrong-direction/
AI on the other hand can execute motivational strategies consistently and competently:
“Yes, research indicates that AI can significantly boost student motivation, particularly when tools are designed to support autonomy, competence, and relatedness. A 2025 meta-analysis of 71 studies found that generative AI produces moderate positive effects on motivation and engagement, with the impact being stronger when students use AI consistently over time rather than in isolated instances.
The motivational benefits are driven by several key mechanisms:
Personalization and Autonomy: AI adapts to individual learning paces and interests, allowing students to direct their own learning paths, which fosters intrinsic motivation and reduces anxiety.
Immediate Feedback: Real-time, personalized feedback helps students understand their progress, enhancing their sense of competence and self-efficacy.
Engagement: Interactive features, such as adaptive gamification and conversational agents, make learning more enjoyable and sustain interest.“
Traditional teacher led classrooms perhaps are best understood through the lens of Clastres’ work on torture rituals in primitive cultures. Clastres views both torture rituals and public education as forms of coercion designed to maintain social order, but they operate differently:
Rites measure personal endurance and mark the body to prevent the rise of separate political power; they are an immediate, physical imposition of the group's will on the individual.
Schooling is a more abstract, systemic imposition where the "severity of the law" is taught intellectually, creating a population that cannot claim exemption from legal authority due to lack of knowledge. Whether the latter form of torture is more repulsive is a matter of debate one might suppose. At any rate, AI offers a measurable outcome oriented alternative amenable to scientific testing.
And similarly with the courts. Encouragingly, countries like Canada and Brazil have tested AI-assisted Online Dispute Resolution for simple civil claims and small money disputes, where algorithms may propose outcomes or mediate, though final binding judgments often still involve human elements or require party acceptance of AI proposals. Its hard to see how failure to achieve total replacement of judges by AIs in the next 5 to 10 years will be able to be attributable to anything other than sentimentality. It appears that support among the public would be widespread: https://www.pew.org/en/research-and-analysis/issue-briefs/2025/08/experience-with-state-courts-highlights-areas-for-improvement
Achieving supplantation progress in the simplest and easiest situations for AI, that is the rule and procedure bound administrative processes of government, are ironically the most difficult politically given the administrative state’s foundation in patronage and clientelism. The sinecures will be the hardest to abolish given the corrupt slugs we call a Congress.
Perhaps an ideological framework is needed to spur progress. Turning to Clastres again, perhaps Society against the State (https://theanarchistlibrary.org/library/pierre-clastres-society-against-the-state ) is a framework compatible with humanitiy’s innate psychological needs: autonomy, competence, and relatedness. When these needs are satisfied by social environments, people experience greater intrinsic motivation, well-being, and personal growth. AI may prove to be ideal tool to promote such social environments, because it decentralizes access to power thereby opens the door to progress. Clastres posited that “Knowledge is the ability to obtain, process and use information, so that it benefits you as an asset and not a liability.” AI ought be able to make this possible so that we are no longer trapped in the manifestation of Tom Clancy’s law of information: “ If you can control information, you can control people."
Longer? No. Less is more.
As the late Howard Pattee wrote, "No amount of semiotic information, thought, or discourse alone can cause the body to move. It takes some physics."
The first function of language was to cause actions, not to make statements.