Learner prerequisites — not information abundance — can bottleneck AI-driven knowledge adoption: unless teaching reaches the prerequisite depth of a target concept, extra instruction is futile, and one-size-fits-all broadcasts can be up to proportionally slower than tailored teaching across diverse learner types.

Main Finding

The paper develops a formal, generative model of "minds" (learners) with prerequisite structure and uses it to show that limits on learning and adoption stem not only from information supply but from learner-side decoding capacity. Two distinct bottlenecks—(i) a structural barrier given by prerequisite reachability/depth and (ii) an epistemic/information-theoretic barrier given by uncertainty about the teacher’s target and the state-dependent usable information—jointly bound how fast teaching can complete. Key corollaries: (a) strict threshold effects in training (completion probability can jump from 0 to 1 when teaching horizon reaches the structural distance to a target), producing non‑concave returns to instructional effort; (b) broadcasting a common curriculum to heterogeneous learners can be strictly and parametrically (linear-in-number‑of‑types) slower than personalized instruction; and (c) acquiring prerequisites refines the learner-side statistical experiment (Blackwell dominance), so prerequisites do more than add concepts—they increase the effective information throughput of subsequent signals.

Key Points

• Formal notion of a mind: concept space C, axiom set A, and a finite family of expansion rules Em each of the form (S ⇒ c) where finite S are prerequisites that unlock concept c.
• Expansion rules induce a closure operator clm(·) (via repeated one-step expansions Φm), and the understanding horizon Um = clm(Am) is the set of all concepts reachable in principle.
• Under a finite-horizon assumption, the family of reachable acquired states above the axioms forms an antimatroid (equivalently a learning space). The paper proves the converse: every antimatroid admits representation by some mind (Theorem 2.27).
• Teaching modeled as sequential communication with a latent target known to the teacher but not the learner. Crucially, the learner parses instructional signals through a prerequisite‑gated parser: a signal about concept c is usable only when c is currently ordered/parseable for the learner; otherwise it collapses to a common null observation. This makes the effective channel output alphabet state‑dependent.
• Relativity of randomness: the same raw broadcast can be informative for one learner state and noise for another. As a learner acquires prerequisites, the parsed experiment Blackwell‑improves (i.e., becomes more informative).
• Two bottlenecks to fast teaching:
  • Structural barrier: the learner must traverse prerequisite‑respecting states until the target becomes parseable; the shortest such route (structural distance/depth) is a hard lower bound.
  • Epistemic/information barrier: even when parseable, identification requires sufficient usable information to discriminate the target; cumulative mutual information through the state‑dependent channel yields an information lower bound on expected teaching time.
• For deterministic targets in fixed-horizon settings, there are discontinuous structural thresholds: if the teaching horizon is below the structural distance, completion is impossible regardless of signals; once the horizon reaches that depth, completion may become feasible — giving non‑concave returns to instructional time and simple failures of uniform resource allocation.
• Heterogeneous learners: using a single broadcast curriculum can be slower than tailoring instruction; the paper gives a broadcast impossibility result (Theorem 5.6) where broadcast can be slower by a factor linear in the number of learner types.
• Simple examples (arithmetic learners, text‑editing learners) illustrate how different prerequisite wiring yields different learning paths and different responsiveness to identical instructional sequences.

Data & Methods

• This is a theoretical/mathematical paper; no empirical data are used.
• Main modeling ingredients:
  • Concept space C (finite or countable).
  • Mind m = (C, Am, Em) with finite prereq sets (finitary expansion rules).
  • One‑step expansion operator Φm and closure clm(K) = least fixed point containing K (constructed by iterating Φm).
• Key mathematical tools and results:
  • Order‑theoretic closure operators and fixed point theorems (Knaster–Tarski).
  • Combinatorial characterization: reachable family forms an antimatroid/learning space under finite‑horizon assumption; equivalence shown (Theorem 2.27).
  • Teaching modeled as a sequential signaling game with a latent target; learner’s parser maps raw signals into parsed observations that depend on current acquired state.
  • Information‑theoretic analysis: entropy, mutual information (log base 2, bits). The parsed experiments are compared via Blackwell ordering; larger acquired states induce Blackwell refinements.
  • Lower bound on expected completion time combining (a) structural shortest‑path length to target and (b) an information accumulation bound (mutual information throughput) — formal statements and proofs given in the text.
  • Constructive examples and impossibility results (e.g., broadcast penalty) with combinatorial proofs showing linear dependence on number of types.
• Assumptions and modeling choices:
  • Prerequisite sets are finite.
  • Axioms and expansion rules (the learner’s architecture) are fixed during the teaching interaction; changing architecture (development) is modeled as moving to a different mind, not as part of the teaching problem.
  • No modeling of semantic truth, inconsistency, forgetting, analogy, or detailed internal representations—concepts are abstract units and understanding is accessibility under rules.
  • Teacher knows the target; the learner does not.
  • Signals are raw inputs that either parse to informative observations if prerequisites are met, or to a common null otherwise.
  • Logs and information measures in bits; finite‑horizon often assumed for combinatorial results.

Implications for AI Economics

• As generative AI lowers the cost of producing explanations, proofs, worked examples, and analyses, the binding constraint shifts from supply of information to absorptive/decoding capacity of users (humans or downstream models). This paper gives a formal account of that absorber-side bottleneck.
• Practical implications for deployment of AI tutoring, documentation, and automated assistance:
  • Investment in prerequisite acquisition (foundational skills, background training) can be disproportionately valuable because it increases the information throughput of subsequent AI outputs (Blackwell improvement). In other words, "contextual readiness" multiplies the value of AI‑produced content.
  • Threshold effects imply non‑concave returns to training effort: partial, evenly spread training may be much less effective than concentrating effort to push learners past prerequisite depths that unlock large chunks of understandability. This favors targeted, intensive upskilling over shallow mass distribution when resources are scarce.
  • Personalization matters sharply. Broadcasting a one‑size‑fits‑all curriculum can be parametrically inefficient relative to individualized instruction; scaling AI assistance without matching prerequisite heterogeneity risks large slowdowns in adoption or capability acquisition.
  • Product and curriculum designers should detect prerequisite structures and sequence content to ensure signals are parseable (e.g., automatically scaffold explanations to current user state), rather than only optimizing content quality or quantity.
  • Policy: workforce retraining funded broadly may underperform policies that concentrate resources on cohorts where prerequisite thresholds can be crossed; similarly, digital public goods (explanations, guides) are more valuable when paired with investments that raise absorptive capacity.
• Limitations and directions for empirical work:
  • The model abstracts from semantic correctness, forgetting, stochastic internal update rules, and cognitive development of the architecture. Empirical work should map real educational tasks to concept spaces and estimate prerequisite graphs to test the predicted threshold, non‑concavity, and personalization gains.
  • Quantifying the information throughput of real instructional signals (e.g., LLM explanations) as a function of learner state is an empirical challenge but central to operationalizing the results.
  • Extensions could incorporate costs of producing signals, noisy or probabilistic parsing, adaptive teachers with partial knowledge of learner architecture, or dynamic changes in learner architecture (development, meta‑learning).
• Overall, the paper supplies a clear formal mechanism explaining why cheap generation of explanations (via AI) does not automatically translate into equal increases in usable knowledge: understanding requires matching content to learner prerequisites, and failure to do so can make abundant AI outputs effectively noise for many users.