Can AI systems invent new concepts rather than reuse trained ones?

The paper argues that current systems are strong at reasoning, coding, theorem proving and tool use, but all of these run inside a representational frame that is "typically fixed and supplied in advance." Open-ended innovation needs a second class of operation: creating, stabilizing and reusing new representational primitives, which "alter the space being searched rather than simply searching within it." It locates the distance to that capacity in two gaps. The vocabulary gap is "the difficulty of inventing and stabilizing new representational primitives rather than merely recombining existing ones." The verifier gap is "the difficulty of judging the value of a new primitive when its full payoff may be visible only after future reuse." Its examples, "number", "negative numbers" and "entropy", are concepts that became valuable by reorganizing many problems at once.

The mechanism is a minimum-description-length view of abstraction. A primitive earns its place by compressing a family of observations, and the paper's critical requirement is amortization: a primitive "must justify its representational value across a family of problems, not a single local instance." It frames intelligence as cognitive discrepancy reduction and separates intra-space transformations, which work inside a fixed frame, from generative transformations, which "may modify the frame itself." On the verifier side it observes that inside a fixed frame "evaluation is usually fast, cheap, and decisive": a program passes its tests or a proof is accepted by a checker. A newly invented primitive has no such checker yet, because its value shows only through later reuse.

This sharpens the nearest notes. Do foundation models learn world models or task-specific shortcuts? shows models that predict well without the structure a world model would give them. The paper makes the concept-level version of that point: current models form concepts, but "these are inherited from the training distribution rather than initiated autonomously." Its thought experiment, showing a model five apples and seven chairs and asking whether it would invent "number", is a test of initiation that the heuristics probe does not run. Can AI research itself without losing human oversight? is the nearest engineering counterpart, and it contrasts: its analyzer distills reusable insights, but its cognition base supplies human priors, which is the part the paper says systems do not originate. The "persistent memory architectures for invented primitives" the paper proposes are the step that note leaves open. Do language models fail at reasoning due to complexity or novelty? describes the same boundary from the evaluation side: search inside a covered frame fails at the novelty edge, and the paper's argument is that only frame change moves that edge.

What the excerpt does not establish: this is a position paper. It reports no experiments, and its claim that no current system autonomously creates a concept like "number" is asserted rather than tested. The "ladder of innovation autonomy" and the adaptive verifiers appear in the abstract but are not developed in the excerpt. The generative condition is introduced as holding "when both hold", and the excerpt does not list the two conditions. The implication is that the two gaps are a diagnosis with a testable prediction. A benchmark that rewards reuse of a newly introduced primitive across a family of tasks would show whether a system closes the vocabulary gap, and the paper does not supply one. At the strength the evidence allows, this is a conceptual framing, not an empirical finding about today's models.