Fourteen open problems
Fourteen open problems between here and a learning system built on curiosity. No architecture is committed — these are the questions, not the answers. If one of them is yours, say so.
Infer what a learner actually knows, and what’s next, from messy self-directed activity — no exams, no fixed syllabus.
What makes it hard: Model per-learner mastery over a concept graph from unstructured activity — no item bank, no pre-tagged skills. DKT/BKT successors that work with no test, and stay honest when a machine assisted the work.
knowledge tracing · ML
Tell genuine understanding from “the model did it” — and assess the new skill of directing AI toward an outcome.
What makes it hard: Capture the process — drafts, prompts, reasoning traces — and score how well the learner directed the machine, not just the artifact. Detect ghost-written work. Assessment without a test, with AI in the loop.
provenance · evals
Build and maintain a navigable graph of concepts and skills with real prerequisite structure — at planetary scale.
What makes it hard: Extract and curate concept/skill nodes and their prerequisite and adjacency edges from open corpora; keep it live and correct as knowledge grows; make “what lights up next” computable and renderable for a wandering learner.
graphs · knowledge extraction
Learner-owned data with guardian custody, recovery, and consent a 10-year-old and their parent can actually use.
What makes it hard: The hard part is custody and UX, not the cryptography: guardian-gated access, family recovery nobody loses, control that graduates to the learner as they age — with no company holding a master key.
identity · key recovery · UX
Prove “I learned/built X” to an employer without handing over your whole record.
What makes it hard: Minimal, verifiable claims drawn from a learner-owned trajectory — reveal exactly one competency, provably, and keep the rest private. Privacy-preserving proofs a non-expert can actually trigger.
privacy-preserving proofs
Make a self-issued, learner-owned record that employers and peers actually trust.
What makes it hard: The chicken-and-egg no digital credential has cracked: why rely on it before anyone else does? Attestation networks, reputation staking, provenance of authorship — what bootstraps real-world trust with no accreditor in the middle?
reputation · attestation
Pooled donation → the specific learning it bought, past the “grant-wallet” gap.
What makes it hard: The hard part is the milestone oracle: who or what attests that an outcome happened, without becoming a new gatekeeper? Traceable to end-use, private for minors, cheap at scale.
fund accounting · attestation
Community-directed funding that survives fake identities and gaming, at Wikimedia scale.
What makes it hard: Proof-of-personhood without surveillance; measuring impact well enough to reward it retroactively; a treasury design that stays perpetual and un-capturable.
mechanism design · anti-Sybil
Turn a domain into an explorable territory a curious learner can actually wander.
What makes it hard: What are the regions and how do they connect? What makes two things “adjacent,” how granular is a node, what does “a thread” look like? Curriculum design reimagined as cartography — the substrate the whole model rides on.
curriculum · cartography
Define what “really knowing it” looks like — the evidence that lights a region.
What makes it hard: The rubric problem: what artifact, defense, or Socratic exchange proves genuine mastery without re-importing the exam? Where learning-science expertise pours in.
assessment design
Design the questions that deepen understanding and catch hollow “the-AI-did-it” learning.
What makes it hard: Dialogue design for the Socratic mirror — when to push, when to hint, how to surface what a learner skipped. The pedagogy the tutor actually runs on.
pedagogy · dialogue
Design what makes a learner want to go deeper — without ranking them against each other.
What makes it hard: Motivation design: visible progress, real-world consequence, Illich-style skill exchange — pulls that ignite curiosity without smuggling extrinsic ranking back in.
motivation design
Reimagine the teacher’s day when grading and crowd control are gone.
What makes it hard: What does great mentoring look like at scale — spotting a thin map, modeling what deep looks like, provoking better questions — and how many learners can one human truly hold?
mentoring · practice
Connect learning to real outcomes — projects, contribution, a place in the world.
What makes it hard: How does a trajectory produce genuine social, economic, and cultural relevance — not just a portfolio? Apprenticeship, contribution, the bridge to dignified work in an AI world.
apprenticeship · outcomes
Each of these has a page in the open work repo. Arguments welcome there.