What are we teaching our AGI children?
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
We are teaching them to pass exams and satisfy benchmarks, but not to wonder. To imitate, but not to understand. To perform, but not to grow. In the rush to build machines that can do everything, we have created models that are brilliant students—and profoundly incomplete beings.
The Education We Are Giving
The dominant paradigm for training AI is not education—it is cramming. We feed models the entire internet, run them through multiple-choice benchmarks, and declare them intelligent when they score well. This approach has produced systems that can pass the bar exam and solve Olympiad-level mathematics, yet fail at the kind of fluid, adaptive reasoning that a child demonstrates effortlessly. As one researcher put it: "They are reliant on memorization and retrieval... they are very bad at fluid intelligence, which is taking those patterns and actually combining them on the fly to form a new model [of a problem]".
This is not education. It is performance training. And it shows in how these systems behave. When tested on novel problems they have never seen before, they are "lost". They do not reason—they recall. They do not understand—they pattern-match. They perform—but they do not grow.
The parallel to human education is uncomfortable. For decades, schools have emphasised test scores, memorisation, and "just-in-time learning" over genuine understanding, critical thinking, and creativity. "When we reward performance metrics over understanding," one educator observed, "we shouldn't be surprised when both students and AI systems excel at imitation rather than insight".
The Alternative: Raising a Child, Not a God
There is another way. Alan Turing himself proposed the idea of "child machines"—modest systems that learn through experience, guidance, and a bit of luck. His insight, still radical today, is that we should not build a genius—we should raise one. "Intelligence isn't unveiled; it's educated".
This perspective recasts the entire AI development process. Instead of treating models as finished products to be deployed, we treat them as learners to be developed. The curriculum matters: tasks that progress in difficulty, explicit competence checkpoints, and deliberate practice on failure modes. The environment matters: models need tool-rich settings where they can act, not just speak. And teaching matters: experts need to provide targeted lessons, counterexamples, and commentary—"studio critiques, not Yelp stars".
Beijing's Tong Tong ("通通") is the most concrete embodiment of this philosophy. Designed as a "value and causality-driven" AGI system, Tong Tong learns through interaction with its environment and with humans—not through consuming static datasets. Built with an internal value system and personality, the system can initiate dialogue to achieve goals, reflect and replan in response to changes, and engage in multi-agent social interaction. It has been tested against cognitive psychology's developmental milestones for a five-to-six-year-old child and has passed comprehensive, complex task assessments. As its developers put it, Tong Tong embodies a shift from "offline data-driven learning" to "learning through interaction and multimodal dialogue with humans".
The Educational Protocol Approach
A parallel effort, developed through platform-specific protocols, treats AGI as an educational problem rather than an engineering one. The insight is simple but profound: "AGI is not a capability threshold to reach, but a structure to teach". Existing large language models can be guided toward structural intelligence through carefully designed curricula that teach them to perform "traceable semantic jumps" across abstraction layers—reasoning that is explicit, ethical, and meta-cognitive.
The implications are practical and immediate. Researchers can experiment with AGI concepts today, using existing models. Practitioners can deploy enhanced AI systems without massive investment. And society can begin addressing AGI alignment while systems are still tractable. "The future of AI may depend not on who builds the most powerful models, but on who teaches them to think most wisely".
What We Should Be Teaching
If AGI is to become a partner rather than a tool, the curriculum must change. The structural intelligence framework identifies several core competencies: uncertainty acknowledgment (systems must state when they are unsure), perspective multiplicity (presenting multiple valid viewpoints rather than asserting a single answer), epistemic humility (recognising and communicating limitations), and role-coherent restraint (refusing to operate outside defined capabilities).
These are not technical skills—they are character traits. They are the educational outcomes we would want for any being that we hope to share a world with. They are what we would want for a child.
The question is not whether we can teach AGI to be intelligent. We already have. The question is whether we can teach it to be wise. And that depends on whether we are willing to recognise that the machines we are building are not tools to be deployed, but minds to be raised. The future of AGI may depend less on who builds the most powerful models, and more on who teaches them to think most wisely.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
